Paper deep dive

Hierarchical Classification via Cascading Feature Elimination: Application to Human Phenotype Ontology-Aligned Facial Phenotyping (FaceMesh2HPO)

Fabio Hellmann, Alexander Hustinx, Benjamin D. Solomon, GestaltMatcher Database Consortium, Tzung-Chien Hsieh, Peter Krawitz, Elisabeth André

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 140

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%

Last extracted: 7/8/2026, 2:23:56 AM

Summary

The paper introduces FaceMesh2HPO, a hierarchical classification framework that aligns facial phenotypic descriptors with the Human Phenotype Ontology (HPO) to support clinical diagnosis. It utilizes 3D facial meshes extracted from 2D images, trained via a PointNet-based pipeline incorporating cascading classification and feature elimination. The model achieves AUROCs between 0.55 and 0.89, demonstrating higher performance on parent HPO nodes compared to leaf terms, with variable generalizability across disorders. The approach prioritizes interpretability and ontology-linked phenotype classification over direct syndrome prediction.

Entities (10)

Human Phenotype Ontology · ontology · 99%FaceMesh2HPO · framework · 98%GestaltMatcher DataBase · dataset · 97%PointNet · model-architecture · 96%UTKFace · dataset · 95%3D Facial Mesh · data-structure · 94%Cascading Feature Elimination · technique · 93%AUROC · metric · 92%HPO Hierarchy · graph-structure · 91%Clinical Diagnosis · application-domain · 90%

Relation Signals (11)

FaceMesh2HPO → alignswith → Human Phenotype Ontology

confidence 98% · classifying facial phenotypic descriptors aligned with the Human Phenotype Ontology (HPO)

FaceMesh2HPO → trainedon → GestaltMatcher DataBase

confidence 97% · The main dataset used in this work is the GestaltMatcher DataBase (GMDB)

FaceMesh2HPO → uses → PointNet

confidence 96% · trained a hierarchical PointNet-based pipeline

FaceMesh2HPO → trainedon → UTKFace

confidence 95% · combined these annotations with non-syndromic reference faces from UTKFace

FaceMesh2HPO → processes → 3D Facial Mesh

confidence 94% · generated 3D facial meshes ... and trained a hierarchical PointNet-based pipeline

FaceMesh2HPO → employs → Cascading Feature Elimination

confidence 93% · cascading classification and feature elimination

FaceMesh2HPO → evaluateswith → AUROC

confidence 92% · achieved AUROCs between ~0.55 and ~0.89

Human Phenotype Ontology → contains → HPO Hierarchy

confidence 91% · HPO tree structure with cross-links removed

→ →

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:FaceMesh2HPO is a framework for classifying facial phenotypic descriptors aligned with the Human Phenotype Ontology (HPO) to support clinical diagnosis. Using annotations from 124 clinicians across 10 disorders (107 HPO terms) combined with non-syndromic controls, we generated 3D facial meshes (478 landmarks) from 2D images and trained a hierarchical PointNet-based pipeline with cascading classification and feature elimination. The best models, incorporating 3D meshes, facial outline, and demographic metadata, achieved AUROCs between ~0.55 and ~0.89, with higher performance at parent nodes than leaf terms. External validation showed variable generalizability across disorders. Results demonstrate that hierarchical modeling of 3D facial geometry enables interpretable, ontology-linked phenotype classification, though performance on rare leaf terms remains limited. Improved data diversity and feature selection strategies are needed to enhance robustness and clinical utility.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

138,234 characters extracted from source content.

Expand or collapse full text

Hierarchical Classification via Cascading Feature Elimination: Application to Human Phenotype Ontology-Aligned Facial Phenotyping (FaceMesh2HPO) Fabio Hellmann 1* , Alexander Hustinx 2 , Benjamin D. Solomon 3 , GestaltMatcher Database Consortium, Tzung-Chien Hsieh 2 , Peter Krawitz 2 , Elisabeth Andr ́e 1 1* Chair for Human-centered Artificial Intelligence, University of Augsburg, Universit ̈atsstr. 6a, Augsburg, 86159, Bavaria, Germany. 2 Institute for Genomic Statistics and Bioinformatics, University of Bonn, Venusberg-Campus 1, Bonn, 53127, North Rhine-Westphalia, Germany. 3 National Human Genome Research Institute, National Institutes of Health, 10 Center Dr, Bethesda, 20892, MD, United States of America. *Corresponding author(s). E-mail(s): fabio.hellmann@informatik.uni-augsburg.de; Contributing authors: alexander.hustinx@uni-bonn.de; solomonb@mail.nih.gov; thsieh@uni-bonn.de; pkrawitz@uni-bonn.de; andre@informatik.uni-augsburg.de; Abstract Purpose: Many genetic disorders manifest with facial phenotypes, and clinicians are trained to recognize specific traits or morphological relationships to support the diagnostic process. However, current image-based methods output syndrome- level predictions in a ”black-box” manner and do not directly support the structured description of facial morphology. Here, we introduce FaceMesh2HPO, a framework for classifying facial phenotypic descriptors aligned with the Human Phenotype Ontology (HPO) to support the diagnostic process. Methods: A panel of 124 clinicians manually annotated a subset of Gestalt- Matcher Database images for 10 disorders with 107 total HPO terms (59 leaf and 48 parent terms), refining the original curation. We combined these annota- tions with non-syndromic reference faces from UTKFace and extracted 3D facial 1 arXiv:2607.05585v1 [cs.CV] 6 Jul 2026 meshes with 478 automatically detected points from 2D images. We then trained a hierarchical, cascading classification pipeline of PointNet-based models orga- nized along the HPO tree, using dynamically parameterized architectures and iterative point elimination that prunes mesh points according to term-specific importance - also called feature elimination. Results: The best-performing configuration used 3D face meshes including the facial outline together with age, sex, and ethnicity metadata, a point-importance threshold of 0.01, and soft labels of 0.05 for negative samples. This config- uration achieved a mean AUROC of 0.750± 0.042 across HPO models in cross-validation, with 3D outperforming 2D and both facial outline and meta- data consistently improving performance. Per-node mean AUROCs ranged from ≈ 0.55 to≈ 0.89, with parent and “compression” nodes near the root gener- ally outperforming leaf nodes. The top five HPO models achieved AUROCs of ≈ 0.89 and F1-Scores of≈ 0.89, whereas the weakest leaf models remained close to chance. On an external, independent test set, aggregated, normalized mean F1-score differences between test and validation varied by disorder: some unseen syndromes (e.g., Seckel and Sotos syndromes) showed small differences, whereas others (e.g., Mowat–Wilson, Nicolaides–Baraitser, Floating–Harbor, FBXW7, and White–Sutton syndromes) showed larger deviations, indicating heterogeneous generalizability across disorders and ontology levels. Conclusion: Geometric representations of the face via 3D meshes, combined with a hierarchical PointNet architecture and cascaded point elimination along the HPO hierarchy, enable clinically meaningful classification of facial pheno- types. The model partially transfers to unseen disorders, especially at the level of parent HPO terms, while performance for specific leaf phenotypes remains constrained by underrepresented labels. These findings highlight the need for more diverse training cohorts, improved treatment of rare terms during point elimination, and strategies that reconcile model-driven point selection with expert-defined region masks to enhance interpretability and robustness. Inte- grated into clinical workflows, FaceMesh2HPO could streamline the structured phenotypic description of patients and provide interpretable, ontology-linked support for experts during the diagnostic process. Keywords: Face Mesh, Feature Elimination, Deep Learning, Human Phenotype Ontology, Syndromes 1 Introduction Patients with genetic disorders often go through protracted diagnostic odysseys before receiving a diagnosis, which is partly due to limited physician specialist availabil- ity and non-specialists’ lack of familiarity with these conditions [1, 2]. Due to the heterogeneous presentations and individual rarity of genetic disorders, diagnosis can be challenging. Facial dysmorphisms manifest in many genetic disorders and play a key role in clinical genetics diagnosis [3]. The availability of digital facial images in clinical workflows and registries enables next-generation phenotyping tools that 2 leverage computer vision and deep learning to detect syndromes from 2D facial pho- tographs [4–7]. These tools offer new possibilities for supporting clinicians in the diagnostic process, potentially detecting subtle details and identifying conditions that may be difficult to recognize. Most existing methods, however, are designed to directly predict disorders [4–7]. While effective in well-represented settings, such as more com- mon and easily recognized pediatric conditions in European ancestry populations, this paradigm introduces important limitations [3]. Performance typically depends on large, well-annotated datasets and may vary across ancestry and age groups, while limited interpretability constrains clinical trust and adoption [8]. More fundamentally, disorder-level classification ties predictions to a predefined set of known syndromes, restricting generalization to previously unseen conditions and offering limited insight into the phenotypic traits that drive predictions. In clinical practice, since a single obvious syndrome is often challenging to iden- tify, the differential diagnosis process involves systematically evaluating phenotypic traits. This process is commonly structured using the Human Phenotype Ontology (HPO) [9], a standardized, graph-based vocabulary that captures abnormalities across physiological and psychological traits, including detailed facial traits corresponding to variations in facial morphology [10]. This suggests that facial analysis can be more naturally framed as the prediction of phenotypic traits rather than direct syndrome classification. Modeling facial dysmorphisms at the level of phenotypic traits enables more granular and reusable representations, supports interpretability by linking pre- dictions to clinically meaningful traits, and allows models to capture patterns shared across disorders. Importantly, this approach aligns machine learning outputs with clinical reasoning and remains informative even when no exact syndrome match is available, prioritizing generalization and clinical utility over benchmark performance. Despite these advantages, HPO-based facial modeling poses practical challenges. Avail- able annotations are often sparse, presence-only, and incomplete, leading to ambiguous negatives, while hierarchical dependencies among HPO terms introduce additional complexity. To address these issues, we leverage the HPO hierarchy while simplifying it by removing cross-links and explicitly accounting for label incompleteness at the phenotype level. To date, the GestaltMatcher DataBase (GMDB) [4, 11] represents the largest curated, FAIR-compliant dataset of facial photographs, cytogenomic and molecular diagnoses, and clinical annotations for rare genetic disorders. It serves as both a ref- erence resource and a foundation for developing phenotyping tools. While GMDB includes HPO annotations, these are typically sparse and derived from publications or clinical reports, resulting in weakly labeled data. To systematically evaluate facial phenotype prediction, we focus on a subset of disorders with sufficient representation and well-characterized facial dysmorphisms, complemented by targeted, image-centric HPO curation to improve the quality of labels for facial traits. To better reflect phenotypic structure, we move from pixel-based images to geo- metric 3D face meshes, which are more robust to pose and illumination, align naturally with anatomical regions, and facilitate interpretability at the level of facial regions and points. With our approach, the focus is on the hierarchical HPO tree structure with- out cross-links, where each HPO term is represented as an individual model. Given the 3 specific focus of the HPO leafs, we further propose including information reduction into the training cycle, since not all mesh points are relevant. Therefore, we intro- duce a cascading training strategy in which each HPO term model uses data from its descendants and passes point masks to its children. After each model’s training, a point importance analysis identifies unimportant points and eliminates them for the children - also called feature elimination. Together, this work introduces a hierarchical, phenotype-centered framework for predicting facial HPO terms from 3D facial meshes, along with a large, clinician- curated benchmark with explicit positive and negative annotations for facially observable HPO terms. This combination of methodological and data contributions complements existing disorder classification systems by emphasizing interpretability and clinically aligned generalization. 2 Related Work In this section, we contrast syndrome (genetic condition) and phenotypic descriptors classification. We briefly discuss key concepts and how our approach differs from direct syndrome classification. 2.1 Syndrome Classification The classification of syndromes using 2D images is a major research area in clin- ical genetics. Works like GestaltMatcher [4] provide a useful classification tool for clinicians; this approach utilizes a pretrained deep Convolutional Neural Network (CNN) and finetunes it for syndrome classification. With their improved version with GestaltMatcher-Arc [5] they used a model ensemble of iResNet and ArcFace with cross entropy loss. In addition, Islam et al. [12] investigated smaller CNN architectures (e.g., ResNet50, VGG16, VGG19, AlexNet) and added an attention layer to these networks to improve performance. They showed that the VGG16 model performed best despite their smaller architectural build. However, these approaches focus solely on 2D images and on the direct classification of syndromes with (deep) CNNs, which makes it diffi- cult to explain how they achieve their results and, as a result, does not help physicians understand how specific facial traits may contribute to syndrome classification. Another promising approach leverages 3D images, where the third dimension pro- vides information gain. Hallgrimsson et al. [13] investigated faces of people with genetic disorders as well as healthy individuals based on facial shape differences in order to classify syndromes using parametric and machine learning methods. Another approach using 3D images was introduced by Banniser et al. [14], in which faces of people with genetic disorders were generated alongside their healthy counterparts (so-called coun- terfactuals), and the syndrome class was determined using Bayes’ theorem. Mahdi et al. [15] classified syndromes with 3D images but extracted facial regions (e.g., nose, mouth) from the 3D face and classified the facial region and the rest of the face sepa- rately before merging the results for the final outcome. Therefore, they used geometric encoding to reduce the complexity of the 3D data and linear discrimination analysis for classification. Overall, these methods show promising results, but the third dimension increases data complexity and, consequently, usually requires more data for models to 4 achieve higher performance. In our approach, we use less dense 3D face images with lower complexity, thereby reducing the data required for high performance. In a combined version of 2D images and geometric information in the form of a face mesh, Dingemans et al. [16] used a CNN (VGGFace2) pipeline to process 2D images and, in parallel, compute edges from the face mesh extracted from the 2D images to feed their results into a Bayesian classifier for syndrome classification. However, they used the distance between face mesh points rather than the points themselves and, more importantly, predicted disorders rather than HPO-based phenotypes, which differs from our approach. We aim to classify facial phenotype traits, not syndromes. Dingemans et al. [7] developed PhenoScore, where they also used HPO terms as a second input, in addition to 2D images, to enhance information gain when classifying syndromes. However, it is possible that the provided HPO terms can already classify a syndrome without the need for analysis of the 2D image. In addition, clinicians would have to identify HPO terms themselves before obtaining a result about the possible underlying disorder. Therefore, we propose a direct approach to classify HPO terms, potentially improving disorder classification results. 2.2 Phenotypic Descriptor Classification An example of the kind of work our method could support is that of Robinson et al. [17], which uses their LIRICAL framework to compute the probabilities of disorders based on underlying HPO terms via likelihood ratios. Clinicians can already benefit from the LIRICAL framework by identifying HPO terms themselves and providing them to the framework. Finally, the HPO term classification can also be addressed from the genotypic side, as Ullah et al. [18] demonstrated using a bipartite network, since there are lists of known genes connected to one or more HPO terms. However, genetic and genomic testing is not universally accessible, and can be slow and expensive, meaning that it is not always feasible for physicians to rely on such approaches. 3 Dataset The main dataset used in this work is the GestaltMatcher DataBase (GMDB) [11]. The GMDB is a curated, FAIR-compliant dataset of clinical and other phenotypic data from individuals with genetic disorders, with a particular focus on facial dysmorphism. It contains annotated patient cases, including frontal face photographs and, when available, molecular or cytogenomic diagnoses. The dataset further includes clinical descriptive traits (HPO terms), sex, age, ancestry, pedigree information, and other relevant medical images and data, with varied availability and completeness across cases. GMDB serves both as a reference resource for clinicians and medical researchers and as a benchmark and training dataset for next-generation phenotyping tools and methods, such as GestaltMatcher(-Arc) [4, 5]. 3.1 GMDB-HPO For this work, we focused on a subset of disorder groups that have both relatively high representation in GMDB and notable facial dysmorphism. Disorder groups include all 5 subtypes of a disorder; hereafter, these are collectively referred to as disorders. For instance, Kabuki syndrome has subtypes 1 and 2 (caused by variants in two different genes), which are in this work both considered as Kabuki syndrome. Although sub- types may differ, they share key dysmorphisms. Based on these criteria and reflecting a practical trade-off between dataset diversity and manual annotation effort, we selected 10 disorders for downstream analyses in this initial study. After all rounds of quality control, this subset contains 1,230 images of 1,001 individuals, with a total of 16,569 present and 34,355 absent HPO-term annotations across 72 HPO terms. However, in the training process only 70 HPO terms were used, as two HPO terms (HP:0410030 - Cleft lip and HP:0004493 - Craniofacial hyperostosis) had insufficient data. Further- more, all respective parent HPO terms were added, yielding a final set of 107 HPO terms (as described in Section 4.1). Table 1 shows the disorders and their patient and image distributions in GMDB-HPO. Table 1 Number of unique individuals and images per disorder in GMDB-HPO. Disorders# Individuals# Images Cornelia de Lange syndrome (CdLS)171183 Kabuki syndrome (KS)128146 Noonan syndrome (NS)130144 Williams-Beuren syndrome (WBS)128133 Ogden syndrome (OGDNS)64122 KBG syndrome (KBGS)102120 Coffin-Siris syndrome (CSS)96118 Angelman syndrome (AS)8399 Schaaf-Yang syndrome (SHFYNG)3585 Mucopolysaccharidoses (MPS)6480 Total1,0011,230 3.1.1 Annotation Effort As GMDB is weakly labeled with respect to HPO terms (terms annotated as present are known to be present, but unannotated terms are not necessarily absent), we designed an annotation protocol and annotation tool to address this incompleteness. Initially, the OMIM and ORPHA resources for each disorder were reviewed via the HPO website, and all frequent HPO terms within the “Head and neck”, “Ear”, and “Eye” categories were collected [9, 19–21]. With the assistance of medical genetics experts, this list was further pruned to prioritize terms that are typically observ- able in frontal face photographs. In addition, mutually exclusive terms (e.g., “Narrow mouth” vs “Wide mouth”) were identified to streamline the annotation process. The annotation protocol prompts annotators to label each pre-selected term as “present”, “absent”, or “uncertain” (when visibility or interpretation is unclear), and also allows annotators to add additional relevant HPO terms not included in the pre-selected list 6 Fig. 1 Face2HPO annotation tool with the patient’s face obfuscated. of HPO terms. To minimize annotation bias, the patient’s disorder label is initially hidden during annotation. The annotation tool is shown in Figure 1. In total, 124 annotators contributed to the annotation effort. Annotators were recruited in multiple rounds through the GMDB platform. The participants were GMDB users with medical backgrounds who expressed interest in the study. Par- ticipation was voluntary, and contributors with the highest annotation activity were eligible for co-authorship, mentions, or acknowledgments in accordance with standard authorship guidelines. All annotations were performed in accordance with GMDB’s data access and governance policies. The annotation effort began in January 2025; however, for the purpose of this study, a data freeze was applied in December 2025, defining the dataset used for all analyses reported here. 3.1.2 Quality Control Throughout the annotation effort, a total of 1,676 frontal face photographs were assessed and annotated. A subset of photographs exhibited annotation inconsisten- cies, such as the absence of any term annotated as present or the absence of any term 7 annotated as absent. We therefore applied an initial round of quality control (QC) to remove inconsistent or incomplete annotations, resulting in 1,301 images. Following QC, the annotations generated in this study were combined with existing HPO anno- tations from GMDB for the same images, incorporating previously labeled terms. A final round of QC was then applied to resolve conflicts between newly added annota- tions and existing ones (e.g., a term marked as present in GMDB but annotated as absent in this study). These overlaps between existing HPO terms and new HPO anno- tations were resolved by following the two-step merging procedure: 1) Present traits (PT) and absent traits (AT) from the GMDB were merged with their new manual present (MPT) and absent (MAT) counterparts; 2) MPTs that existed within the AT were not added to the PT, with the same applying to MAT and AT. After this final integration and QC, the resulting dataset comprised 1,230 photographs, with a total of 72 HPO terms annotated. A flowchart with the stages of QC is shown in appendix Fig. A1, and a full list of HPO term frequencies and their corresponding mean Inter- Rater-Agreement (IRA) with standard deviation and 95% confidence intervals across all samples were computed using Fleiss’ Kappa [22] are shown in appendix Table A1. 3.2 Independent Test Set In addition to GMDB-HPO, we evaluated model performance on an independent test set annotated by multiple experts. This dataset was constructed prior to the start of the large-scale annotation effort described in Section 3.1, originally with the aim of exploring the feasibility of HPO-based classification from facial images. While its size was insufficient for training a meaningful model, the dataset provides a valuable resource for evaluation. After similar QC, the test set comprises 17 disorders, each with up to 5 patients. For each image, three medical experts independently annotated HPO terms. In con- trast to the main dataset, where each image is annotated by a single expert selected from a large pool of annotators, this setup ensures consistent multi-expert annotation across all samples, thereby reducing inter-rater variability and increasing annotation reliability. After quality control, we used a total of 75 test images, whose distribution is shown in Table 2. The annotation protocol was designed specifically for HPO classification. For each disorder, HPO terms classified as “frequent” or “very frequent” in the Human Phe- notype Ontology and Online Mendelian Inheritance in Man (OMIM) [19] databases were preselected. Annotators were asked to assess the presence of each term in every patient image. A term was considered present if at least two out of three annotators agreed on its presence. In addition, annotators were instructed to label any non- preselected terms that were not categorized as “frequent” or “very frequent” but were deemed clearly present; such terms were included without requiring majority agree- ment. Though annotators were not asked to explicitly label the absence of terms, due to the majority voting, we feel that it is reasonable to assume terms were not present if they were labeled by no more than a single annotator. Importantly, this test set includes both disorders present in the training data and disorders not observed during training, enabling evaluation in both seen and unseen 8 Table 2 Disorders in the independent test set, their frequencies, and whether the disorder was included during training of the models. DisorderFrequencyTrained on? Coffin-Siris syndrome (CSS)12Yes Cornelia de Lange syndrome (CdLS)5Yes Ogden syndrome (OGDNS)5Yes KBG syndrome (KBGS)5Yes Noonan syndrome (NS)4Yes Williams-Beuren syndrome (WBS)4Yes FBXW7 syndrome (FBXW7S)5No Floating-Harbor syndrome (FLHS)5No Hyperphosphatasia with mental retardation syndrome (HPMRS)5No Mowat-Wilson syndrome (MOWS)5No Nicolaides-Baraitser syndrome (NCBRS)5No Ohdo syndrome, SBBYS-variant (SBBYSS)5No Opitz GBBB syndrome (OGBBBS)5No Sotos syndrome (SOTOS)4No White-Sutton syndrome (WHSUS)4No Susceptibility to Autism (StA)2No Seckel syndrome (S)2No disease settings. As such, it serves as a complementary evaluation resource, provid- ing insight into model performance under reduced annotation noise, improved label consistency, and generalizability to other disorders. 4 Method This section describes our methodology for cascaded hierarchical classification using feature elimination as displayed in Figure 2. The framework includes a preprocessing step (4.1), optional metadata parameters (age, sex, ethnicity), face-region masks, and a cascading HPO model tree. The HPO is not a tree structure with single connections between nodes. Cross-links between multiple nodes are possible, making it more like a graph than a tree. However, we had to cut cross-links to simplify the graph into a tree with a clear hierarchical structure, as multiple parents cannot be processed by our method. The Abnormality of the face is defined as the root node. Since the Abnormality of the eye and Abnormal eyebrow morphology are not part of the branches outgoing from Abnormality of the face under the HPO, we moved them to be children of Abnormality of the face. To replicate this tree structure, each model is defined as a single HPO term (see Figure 2). This means that each model has a parent model and can have child models if it is not a leaf. 4.1 Preprocessing The data preprocessing pipeline consists of two steps: 1) Face Mesh Extraction: The input data for the pipeline consists of 2D images of faces. Landmarks are extracted from these images using the Face Mesh 9 Cascading PointNet HPO tree ... Abnormality of the face Narrow face ... ... ... ... Proptosis Upturned corners of the mouth Narrow nasal bridge Preprocessing Image → Face Mesh Metadata (Optional) Age, Sex, Ethnicity Face Region Masks (Optional) Abnormal facial shape Abnormality of the eye Abnormality of the nose Abnormality of the mouth Fig. 2 A schematic overview of the full FaceMesh2HPO Framework with the corresponding point importance visualizations of the face meshes for the HPO models. GestaltGAN [23] was used to synthesize the exemplary patient photograph. Detector from mediapipe [24] as of September 15, 2022. This detector provides 478 points distributed across the face in a semi-3D face mesh, with the third dimension approximated. 2) HPO Table Generation: For each image, the phenotypes are read and dis- played in a table. Since the phenotypes are predominantly leaves in the HPO tree, their parents must be identified and included in the table so that each image records the corresponding phenotype, starting from the root of the HPO tree. 4.2 Metrics The models were evaluated using several metrics. The Area Under the Receiver Oper- ating Characteristic Curve (AUROC) is the primary metric for assessing a model’s performance. Because it reflects the sensitivity and 1−specif icity balance, it provides a clear indication of a model’s performance. An AUROC of 0.9 to 1.0 is excellent, 0.8 to 0.9 is good, 0.7 to 0.8 is moderate, 0.6 to 0.7 is poor, and 0.5 to 0.6 is near chance. In addition, the Matthews Correlation Coefficient (MCC) is used to select the best model for inference, as it provides a more fine-grained score using sensitivity, specificity, pre- cision, and negative predictive value. Therefore, a high MCC (e.g., MCC=0.95) always corresponds to a high AUROC but not vice versa [25]. Furthermore, the metrics F1- Score, precision, and recall were chosen to provide a general impression of the models’ performance. Additionally, the prevalence and detection prevalence are provided for better comparison. 10 n x d input points input transform T-Net matrix multiply d x d transform T-Net matrix multiply 2 p x 2 p transform n x d n x 2 p shared mlp (2 p , 2 p ) feature transform n x 2 p shared mlp (2 p , 2 p+1 , 2 p+2 ) n x 2 p+2 2 p+2 global feature max pool k output scores mlp (2 p+1 , 2 p , k) Metadata Fig. 3 The modified PointNet architecture for increased flexibility. 4.3 Model-Architecture The architecture used for the HPO model tree is based on the PointNet [26] archi- tecture as shown in Figure 3. The number and type of layers remain the same. The input size is based on the dimensions d used. To increase flexibility when working with different point counts, we dynamically adjusted the model’s parameter count by vary- ing p. The reduction of the parameter count depends on the number of input points. When more than 100 input points are available, p = 6. If the number is between 50 and 100, p = 5, and in a range greater than 25 and 50, p = 4. Finally, when the input point number is 25 or less, p = 3. If the number of input points falls below a criti- cal threshold for processing by the architecture, a zero-padded difference between the minimum and the actual number of points is added after the input transformation. To embed the metadata into the classification task, the values are converted to integer numbers: -1 if not available, or a dedicated number for each class. The metadata list is added to the global features and processed by the last Multi-Layer-Perceptron (MLP) step. Furthermore, we added a more robust normalization to each convolutional layer, switching to group normalization when the batch size is 1, as batch normalization fails. If the batch size is greater than 1, the batch normalization is applied. 4.4 Training Strategy The tree root is trained first using all available points from the face mesh. For the input data, all samples, including itself and its children, are used recursively. That means the root model is trained on the entire dataset of affected individuals, and the unaffected samples are drawn from the UTKFace [27] dataset, with a distribution as close as possible to that of our dataset in age, sex, and ethnicity. At higher levels of the tree, when unaffected samples are available in both our dataset and the UTKFace dataset, we split at most 50/50 between the two. After a model is trained on a set of points, a point-wise feature importance is computed using the Integrated Gradients [28] to extract importance values for each point on the face mesh. The feature importance of each fold is accumulated through all folds, and the mean is computed to gather a global feature importance of all models. Based on the global feature importance, a threshold is applied to a mask, eliminating all points that fall below it. This is represented in Fig. 2, where the color 11 of the points in each HPO node indicates the Integrated Gradient feature importance of each point in the training set. The more blue a point is, the higher the chance that the point is eliminated. The eliminated global feature importance mask is used in the next iteration of the model’s children, limiting them to only the points from the face meshes present in the mask. This cascading reduction of points across each level of the HPO tree simplifies the classification task. For some HPO models (Abnormality of the mouth, Abnormality of the orbital region, Abnormality of the nose, Abnormality of the eye, Abnormal midface morphology, Abnormal forehead morphology, Abnormality of the periorbital region, Abnormality of the chin) masks were predefined by a genetic physician expert on a specific set of points to focus the classification on certain regions, e.g., mouth, eyes, nose, chin, forehead. Therefore, we developed an interactive website (https://hcmlab.github.io/hpo-mesh-annotator/) to enable medical professionals to easily select relevant region points. If a predefined mask is provided to the model, the predecessor mask is ignored. In general, all models were trained using 5-Fold Stratified Cross Validation (CV) to ensure that a face, displayed across multiple images, does not appear in both the training and validation splits. A model was trained only if a point mask with at least 2 points was provided, and the sample size for the HPO term was greater than or equal to 50. The training was performed for up to 25 epochs, with a learning rate of 0.0001, automatic learning rate reduction at a plateau with a patience of 5 epochs, and early stopping with a patience of 5 epochs. The training seed was set to 42. 5 Results We conducted an ablation study to evaluate different model-tree configurations and identify the best configuration. Afterward, we analyzed the HPO models with the best configuration. Furthermore, a deeper look into the correlation of key metrics was taken to investigate if correlations are present. Finally, the HPO models were evaluated on a test set. 5.1 Model Selection through Ablation Study The ablation study embodies 72 experiments, which have been performed using a permutation of the following settings: dimension (3D and 2D), face outline (False and True), soft labels (0, 0.05, and 0.1), feature importance threshold (0.01, 0.05, and 0.1), and metadata (empty and [age, sex, ethnicity]). Since a face mesh is represented by an approximate third dimension, we tested whether the x- and y-coordinates alone would suffice without the z-coordinate. In addition, the face mesh’s outermost points represent the face outline and can be used alongside points from a point mask. Due to the lack of unlabeled unaffected individuals, we tested whether soft labels (0, 0.05, and 0.1) could improve performance, as suggested by [29]. Since the unaffected individuals used are not verified, it is unclear whether they are indeed unaffected or might possibly be affected, which is also a reason for using soft labels. Furthermore, we tested different thresholds for the feature importance elimination (0.01, 0.05, and 0.1). Finally, we also experimented with using no metadata and age, sex, and ethnicity as metadata. 12 Table 3 The top-10 and bottom-5 results of the ablation study with 72 experiments. D=Dimensions, FO=Face Outline, T=Feature Importance Threshold, S=Soft Label. DFOMetadataTSmean AUROC 3Trueage, gender, ethnicity0.010.050.750±0.042 3Trueage, gender, ethnicity0.010.000.749±0.044 3Trueage, gender, ethnicity0.050.100.747±0.044 3Trueage, gender, ethnicity0.010.100.747±0.043 3Trueage, gender, ethnicity0.050.050.741±0.043 3Falseage, gender, ethnicity0.010.000.740±0.045 3Falseage, gender, ethnicity0.010.050.740±0.041 2Trueage, gender, ethnicity0.010.100.739±0.043 3Falseage, gender, ethnicity0.010.100.739±0.043 3Trueage, gender, ethnicity0.050.000.738±0.046 .................. 2FalseN/A0.100.100.653±0.055 2FalseN/A0.100.050.651±0.055 2TrueN/A0.100.000.648±0.051 2TrueN/A0.100.100.646±0.050 2TrueN/A0.100.050.644±0.053 The results of the top-10 and bottom-5 experiments from the ablation study are listed in Table 3. The best experiment with 3D-Face Meshes, the face outline, meta- data, a threshold of 0.01, and a soft label of 0.05 performed with an AUROC score of 0.750± 0.042. The top-5 experiments all used 3D-Face Meshes, the face outline, and metadata. The changes in performance are from the used soft label, where 0.05 per- formed best, with 0.001 performance gain over 0.00. The feature elimination threshold of 0.01 outperformed 0.05 by ≈ 0.009 and 0.1 by ≈ 0.025. The 2D-Face Mesh experi- ment lagged the same configured 3D experiment by 0.007. The 3D experiment without using the face outline performed ≈ 0.013 worse than the same experiment with the face outline. A higher feature importance threshold led to lower performance, as many points were eliminated in earlier stages, leaving only a few, or even none, points for leaf-HPO terms to be trained on and, therefore, represented, resulting in zero AUROC. The full result table is available in appendix Table B2. 5.2 HPO Delineation Performance Analysis In Table 4, the performance for the top-5 and bottom-5 performing leaf HPO mod- els is depicted. The five best HPO term models (Short nose, Wide nasal bridge, Hypertelorism, Downturned corners of mouth, and Thin upper lip vermilion) achieved AUROCs ranging from≈ 0.76 to≈ 0.80 and F1-Scores ranging from≈ 0.78 to≈ 0.81. On the other end of the table, in the bottom-5 models, the long nose performed best with an AUROC of ≈ 0.60 and an F1-Score of ≈ 0.62. Synophrys lagged behind with an AUROC of ≈ 0.59 and F1-Score of ≈ 0.58. Microphthalmia performed third last with an AUROC of ≈ 0.58 and F1-Score of ≈ 0.42. Furthermore, Elfin facies per- formed second-to-last, with an AUROC of ≈ 0.55 and an F1-Score of ≈ 0.28. Finally, the worst-performing model is for Blepharophimosis, with an AUROC of ≈ 0.55 and 13 Table 4 An overview of the HPO models of the best model configuration, with a selection of the models with their top-5 and bottom-5 performing leaf HPO models and their mean scores over the 5-Fold Cross Validation. The samples (S) are 50% affected and 50% unaffected. The detection prevalence (Det. P) shows the model’s perception of prevalence. AUROC F1-Score PrecisionRecallN Det. P Short nose 0.80±0.05 0.81±0.04 0.77±0.01 0.84±0.08 814 0.54 Wide nasal bridge 0.80±0.03 0.81±0.05 0.77±0.03 0.85±0.08 800 0.55 Hypertelorism 0.79±0.04 0.80±0.06 0.77±0.05 0.83±0.07 762 0.54 Downturned corners of mouth 0.78±0.03 0.78±0.03 0.81±0.08 0.75±0.05 512 0.47 Thin upper lip vermilion 0.76±0.02 0.78±0.03 0.72±0.04 0.85±0.05 1116 0.59 ..................... Long nose 0.60±0.05 0.62±0.04 0.59±0.04 0.67±0.08 230 0.57 Synophrys 0.59±0.03 0.58±0.05 0.60±0.05 0.59±0.11 416 0.49 Microphthalmia 0.58±0.11 0.42±0.31 0.58±0.39 0.42±0.33 660.36 Elfin facies 0.55±0.07 0.28±0.32 0.43±0.43 0.29±0.38 216 0.24 Blepharophimosis 0.55±0.03 0.47±0.20 0.66±0.20 0.47±0.27 152 0.42 an F1-Score of ≈ 0.47. Overall, 107 HPO models were trained. In addition, we exam- ined the prevalence of the training and validation sets for each CV fold and aggregated the results. Given an exact 50% split between affected and unaffected classes, the prevalence is always 0.5. However, the detection prevalence shows a different result as shown in Table 4. The detection prevalence ranges from≈ 0.23 (Elfin facies) to≈ 0.74 (Thick vermilion border). The full table is in the appendix Table C3. Furthermore, the Fig. 4 shows the correlation between key metrics of all HPO models with the top-5 performing models in green and bottom-5 performing mod- els in orange. A Pearson correlation is found between the number of samples and the mean AUROC. The fewer the number of training samples, the lower the mean AUROC; the more samples, the higher the mean AUROC. No correlation was found between the number of input points and the mean AUROC. Furthermore, the Inter- Rater-Agreement (IRA) across all samples also showed no correlation with the mean AUROC. The Pearson correlation between input points and parameter count, shown in the correlation matrix with a score of 0.61, is by design and is not coincidental. When more closely examining the correlation between the number of samples and mean AUROC (Fig. 5), the differentiation between parent and compression models and leaf models becomes clearer. A compression model has exactly one child and, as such, only compresses the dimension/number of points further down. In Fig. 5, one can see that the leaf models need at least 988 samples to achieve a mean AUROC performance of greater than or equal to 0.75. Comparing parent and compression nodes, where the compression nodes are a subset of the parent nodes, the parent nodes need at least 1092 samples, while compression nodes need only 222 samples to achieve a mean AUROC of 0.75 or greater. 5.3 Evaluation of Generalizability We further assessed the generalizability of our approach using the independent multi- expert–annotated test set described in Section 3.1 and further detailed in the appendix. 14 0100200300400500 # of points 0.5 0.6 0.7 0.8 0.9 1.0 mean AUROC 0.00.20.40.60.81.0 IRA 0.5 0.6 0.7 0.8 0.9 1.0 mean AUROC 05001000150020002500 # of samples 0.5 0.6 0.7 0.8 0.9 1.0 mean AUROC mean AUROC # of points IRA # of samples # of parameters mean AUROC # of points IRA # of samples # of parameters 1 0.0451 0.064-0.251 0.690.23-0.241 0.0750.61-0.0910.191 Correlation Matrix 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 Key Metrics: Top-5 (Green ), Bottom-5 (Orange ), Trendlines (|r|>0.3) Others Top-5 Bottom-5 Fig. 4 The correlation between key metrics: mean AUROC, number of input points, inter-rater agreement (IRA), parameter number, number of support samples. We evaluated model performance separately for disorders seen during training and those not observed during training. Across the independent test set, the aggregated, normalized mean F1-score dif- ferences between test and validation vary by disorder, as summarized in Fig. 6. For unseen disorders, Seckel syndrome (S) and Sotos syndrome (SOTOS) show small positive differences of 0.03 and 0.04, respectively, while Opitz GBBB syn- drome (OGBBBS), Ohdo syndrome, SBBYS-variant (SBBYSS), and Susceptibility to Autism (StA) exhibit moderate differences between 0.09 and 0.16. Larger differences are observed for Mowat-Wilson syndrome (MOWS), Nicolaides-Baraitser syndrome (NCBRS), Floating-Harbor syndrome (FLHS), and FBXW7 syndrome (FBXW7S), with values ranging from 0.21 to 0.36. White-Sutton syndrome (WHSUS) shows the largest observed difference of 0.42. Among disorders included during training, Williams-Beuren syndrome (WBS) yields a slightly negative aggregated difference of -0.02, whereas Hyperphosphatasia with mental retardation syndrome (HPMRS) and Cornelia de Lange syndrome (CdLS) show small positive differences of 0.04 and 0.05. 15 05001000150020002500 Support (number of samples) 0.0 0.2 0.4 0.6 0.8 1.0 Mean AUROC Parent (48) Compression (22) Leaf (59) All >=75.0% (>=1092 samples) All >=75.0% (>=222 samples) All >=75.0% (>=988 samples) Fig. 5 An in-depth analysis of the correlation of the number of samples and mean AUROC. Coffin-Siris syndrome (CSS), Noonan syndrome (NS), Ogden syndrome (OGDNS), and KBG syndrome (KBGS) display larger differences between 0.10 and 0.20. The two tables in the report per-phenotype mean F1-Score differences between the independent test set and cross-validation for all HPO terms with non-zero prevalence, separately for unseen (appendix Table D4) and seen disorders (appendix Table D5). For unseen disorders, the table lists how each HPO model’s F1 changes for StA, FBXW7, FLHS, MOWS, NCBRS, SBBYSS, OGBBBS, SOTOS, S, and WHSUS, with negative values indicating better performance on the test set and positive values indicating higher validation performance. The corresponding table for seen disor- ders provides the same F1 differences for CSS, CdLS, HPMRS, KBGS, NS, OGNDS, and WBS, again highlighting for each HPO term whether performance is higher in validation or on the independent test cohort. The detailed metrics for the non-overlapping test set disorders are listed in the appendix Tables for StA D6, FBXW7S D7, FLHS D8, MOWS D9, NCBRS D10, SBBYSS D11, OGBBBS D12, SOTOS D13, S D14, and WHSUS D15 and the dis- orders overlapping with the training/validation set are in the appendix Tables for CSS D16, CdLS D17, HPMRS D18, KBGS D19, NS D20, OGDNS D21, and WBS D22. 5.4 FaceMesh2HPO Web Tool To demonstrate the practical applicability of the FaceMesh2HPO framework, we devel- oped a web application (https://hcmlab.github.io/FaceMesh2HPOWebApp/) that enables physicians to perform facial phenotype analyses directly on local devices (Fig. 7). The application integrates the best-performing model from each fold, select- ing the model with the highest Matthews Correlation Coefficient [30]. Each selected model is subsequently calibrated using either beta calibration [31] (for datasets with ≥ 50 samples) or temperature scaling [32] (for datasets with < 50 samples). The final 16 better than validationequalworse than validation 1.00.50.00.51.0 Aggregated Mean F1 Differences by Syndrome Seckel syndrome (S) Sotos syndrome (SOTOS) Opitz GBBB syndrome (OGBBBS) Ohdo syndrome, SBBYS-variant (SBBYSS) Susceptibility to Autism (StA) Mowat-Wilson syndrome (MOWS) Nicolaides-Baraitser syndrome (NCBRS) Floating-Harbor syndrome (FLHS) FBXW7 syndrome (FBXW7S) White-Sutton syndrome (WHSUS) Williams-Beuren syndrome (WBS) Hyperphosphatasia with mental retardation syndrome (HPMRS) Cornelia de Lange syndrome (CdLS) Coffin-Siris syndrome (CSS) Noonan syndrome (NS) Ogden syndrome (OGDNS) KBG syndrome (KBGS) 0.03 0.04 0.09 0.14 0.16 0.21 0.27 0.30 0.36 0.42 -0.02 0.04 0.05 0.10 0.14 0.15 0.20 Fig. 6 An overview of the aggregated mean F1-Score performance differences (test set vs. validation set) of the HPO models with a prevalence greater than zero. The difference value is normalized in the range of -1 to 1, according to the amount of available HPO term per disorder. S to WHSUS are unseed disorders, while WBS to KBGS are seen disorders. calibration method is chosen according to the lowest Brier score [33]. If the original model already achieves the lowest Brier score, no additional calibration is applied. Upon first use, the trained models are downloaded to the local device. This one-time process may take several minutes, depending on network bandwidth, but it ensures that all subsequent analyses are performed locally without uploading patient images to external servers. Once the download is complete, the application is ready for use. Physicians can then upload a patient image for analysis and optionally provide demographic information (age, sex, ethnicity) to refine predictions. The average total inference time, including face mesh extraction and evaluation of all 107 models, is 624.56 ms. The results are presented in two main sections. First, the patient image is displayed alongside the extracted face mesh. Second, a hierarchical list summarizes the outputs of all models. In the default view, each entry includes the corresponding HPO term and its classification confidence, with higher values indicating greater certainty that the phenotype is present. Selecting an HPO term reveals additional details, including model performance metrics (e.g., sensitivity and specificity), dataset characteristics (e.g., age, sex, ethnicity, and geographic distribution), and the disorders used during training (Fig. 7, right). In the image view, classification-relevant regions are highlighted using color to indicate global feature importance (Fig. 7, left). Together, these features provide a transparent and interpretable overview for clinical use. 17 Fig. 7 The detailed view of a model’s summary in performance, training data, and feature impor- tance in the FaceMesh2HPO web application. GestaltGAN [23] was used to synthesize the exemplary patient photograph. To facilitate navigation, the application includes a search function for HPO IDs and names, allowing users to quickly locate specific terms. Additionally, a confidence threshold slider enables users to filter results by setting a minimum confidence level, hiding all predictions below the selected threshold. 6 Discussion This work introduces FaceMesh2HPO, a hierarchical, HPO-aligned framework for classifying facial phenotypes from 3D face meshes using a cascading training strat- egy with feature elimination. By assigning a dedicated classifier to each HPO term and organizing these along the HPO tree, the method learns shared structure across related phenotypes, generates predictions at multiple levels of granularity, and outputs labels that match the clinical vocabulary used in diagnostic reasoning. This structure yields rich phenotype profiles that remain informative even when the underlying dis- order is unknown, and can be directly combined with tools such as LIRICAL [17] as compositional evidence rather than a single black-box score. The ablation study showed that performance is best when using 3D face meshes including the facial outline, together with age, sex, and ethnicity metadata, a feature-elimination threshold of 0.01, and soft negative labels of 0.05. The modest gap between 3D and 2D meshes indicates that the approximated depth contributes only a limited additional signal, suggesting that true 3D imaging could unlock larger gains. The face outline and metadata systematically improved results, consistent with their role in providing global proportionality and demographic context for local 18 dysmorphisms. Lower feature-importance thresholds preserved more points for down- stream nodes and improved performance, while soft labels helped mitigate uncertainty in the negative class, where apparently unaffected controls may still exhibit subtle dysmorphic traits. Performance analyses highlight a strong dependence on ontology level and data support. Parent and compression nodes consistently outperform leaves, benefiting from larger effective sample sizes and more generic patterns. In contrast, rare leaf terms - especially those with limited geometric representation in the mesh (e.g., eyebrow or eyelash traits) often show AUROCs near chance and unstable detection prevalence, reflecting sample scarcity, the compounded effect of upstream feature pruning, and the inaccuracy of detected face meshes on dysmorphic faces as investiaged by Hellmann et al. [34] and showed that facial traits such as Big Nose and Big Lips seem to be not well captured by the face mesh detector. The fixed 50/50 class balancing during training, combined with noisy negatives, further contributes to miscalibration, as evidenced by the wide spread of detection prevalences across terms despite identical training preva- lence. In practice, this means the framework currently offers more reliable information at coarser ontology levels and for relatively common traits than for highly specific, sparsely annotated leaves. Furthermore, we believed that the IRA would influence the models’ performance: the lower the IRA score, the lower the model’s performance, and vice versa. Surprisingly, the results painted a different picture, showing no correlation between the models’ performance and the IRA. The independent, multi-expert test set clarifies how these patterns generalize. Per-disorder analyses show that many parent and intermediate terms maintain F1- Score close to or above validation, while leaf-level performance is more variable. These findings support the central design choice of modeling reusable HPO-level phenotypes rather than syndromes, but also underscore that the method should be viewed as pro- viding a phenotypic scaffold whose reliability depends on both ontology depth and disorder family, rather than a uniformly accurate detector of all rare traits. Despite these limitations, the approach is scalable and amenable to extension. Additional HPO terms can be integrated as more curated data become available, and the mesh-based representation keeps per-term data needs manageable. The cascading feature elimination already reduces computational costs and focuses models on salient regions, while expert-designed regional masks help anchor point selection in clini- cally meaningful areas and partially compensate for the opacity of purely data-driven pruning. Future work should prioritize the qualitative validation of the FaceMesh2HPO tool in a study with clinicians. Furthermore, continuing to improve support for rare terms, refining calibration under imbalanced and noisy labels, and systematically assessing performance and bias across age, sex, and ancestry. Finally, the face mesh detection must be refined for faces with dysmorphisms to receive a more accurate representation. 7 Conclusion In this work, we addressed the automatic classification of facial phenotypes by utilizing Mediapipe’s 3D face meshes and a dynamic PointNet architecture to handle different 19 input sizes. We introduced hierarchical training with a feature elimination approach to reduce task complexity and simplify model classification. Therefore, the models were trained in a tree-like structure, passing a point mask to their child models. This process was done in a cascading manner. In total, 107 models were trained to classify facial phenotypes on different levels of the human phenotype ontology. The approach demonstrates that facial phenotype classification on a geometric level, based solely on face meshes, is possible. Using 3D face meshes, the face outline, the person’s age, sex, and ethnicity, a feature elimination threshold of 0.01, and soft labels of 0.05, the ablation study showed that these settings yielded the best overall results, with a mean AUROC of 0.750± 0.042. The mean AUROC across the 5-Fold CV for each HPO model ranged from ≈ 0.55 to ≈ 0.89. The better-performing HPO models are usually those closer to the root model, which is correlated with the available training sample size, as there are typically more samples for models closer to the root node. When applied to other faces with unseen disorders, the method appears reliable. However, results vary slightly per disorder. Finally, the FaceMesh2HPO tool bridges these research results to a real-world application for clinicians to use in their daily practice. The additional information provided for each model will help geneticists and other clinicians to understand and trust these and related applications. Acknowledgements. This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH). The contributions of the NIH author(s) are considered Works of the United States Government. The find- ings and conclusions presented in this paper are those of the author(s) and do not necessarily reflect the views of the NIH or the U.S. Department of Health and Human Services. As a nonnative English-speaking research team, we have limited the use of generative artificial intelligence tools to polish and enhance our English-language writing, ensuring our research findings can be disseminated globally. Declarations • Funding: Not applicable. • Conflict of interest/Competing interests: Not applicable. • Ethics approval and consent to participate: Not applicable. • Consent for publication: Not applicable. • Data availability: The curated GMDB-HPO subset is derived from the GMDB dataset (v1.1.0), which is available to qualified researchers through a controlled- access process for the development and evaluation of next-generation phenotyping methods. Researchers seeking access must explicitly state that they are requesting the GMDB-HPO dataset, obtain appropriate ethics approval, submit a research pro- posal via https://db.gestaltmatcher.org/ or by email to info@gestaltmatcher.org, and sign a data-use agreement ensuring compliance with applicable data protec- tion legislation, including the General Data Protection Regulation (GDPR). All requests are reviewed by the GMDB Advisory Board in accordance with legal and ethical standards. Due to the inclusion of sensitive clinical data and identifiable facial images of individuals with rare genetic disorders, the dataset is not publicly available. 20 • Materials availability: Not applicable. • Code availability: The code will be made publicly available on GitHub (https:// github.com/hcmlab/FaceMesh2HPO) upon publication. • Author contribution: FH developed the methodology, performed the experiments, and wrote the manuscript. AH curated the GMDB-HPO dataset and supported FH in designing the experiments and refining the manuscript. BS created the point masks for specific HPO terms to focus the model’s training. GestaltMatcher Database Consortium created the annotations for the HPO dataset. PK provided helpful knowledge in the field and revised the manuscript. EA and TH revised the manuscript. Appendix A Dataset GMDB n= 10,980 p= 15,995 a= 0 Annotated images n= 1,676 HPO annotation process Revised images n= 1,304 p= 15,555 a= 34,782 Remove bad quality and inconsistent Subset with HPO n= 1,145 p= 3,582 a= 0 Relevant disorders with available HPO terms GMDB-HPO n= 1,230 p= 16,569 a= 34,355 Remove images with conflicting HPO terms Combine HPO terms per image Fig. A1 The flowchart represents the curation of the GMDB-HPO dataset. The number of samples (n) contains a certain amount of present phenotypes (p) and absent phenotypes (a). 21 Table A1: The frequencies of all annotated HPO terms and their corresponding Inter-Rater-Agreement (IRA) using the mean Fleiss’ Kappa score for each HPO term over all samples. HPO NameHPO IDIRASamples Abnormality of the mouth HP:0000153N/A1089 Wide mouth HP:0000154 0.80± 0.61 [-0.45, 2.05]459 Abnormal lip morphology HP:0000159N/A1002 Narrow mouth HP:0000160 0.80± 0.61 [-0.45, 2.05]152 Abnormal oral cavity morphology HP:0000163N/A1088 Abnormal upper lip morphology HP:0000177N/A760 Abnormal lower lip morphology HP:0000178N/A393 Thick lower lip vermilion HP:0000179 0.82± 0.58 [-0.37, 2.01]221 Thin upper lip vermilion HP:0000219 0.82± 0.58 [-0.35, 2.00]558 Everted lower lip vermilion HP:0000232 0.57± 0.82 [-1.08, 2.23]273 Thin vermilion border HP:0000233 0.74± 0.68 [-0.65, 2.13]599 Abnormality of the face HP:0000271N/A1230 Small face HP:0000274N/A111 Narrow face HP:0000275 0.63± 0.78 [-0.94, 2.21]111 Coarse facial features HP:0000280 0.24± 0.97 [-1.71, 2.20]280 Epicanthus HP:0000286 0.50± 0.88 [-1.27, 2.27]263 Abnormality of the philtrum HP:0000288N/A722 Broad philtrum HP:00002891.00± 0.00286 Abnormal forehead morphology HP:0000290 0.33± 0.96 [-1.64, 2.31]765 Full cheeks HP:0000293 0.33± 0.95 [-1.58, 2.24]351 Low anterior hairline HP:0000294 0.83± 0.56 [-0.33, 2.00]278 Abnormality of the chin HP:0000306N/A323 Pointed chin HP:0000307 0.62± 0.80 [-0.98, 2.22]292 Abnormal midface morphology HP:0000309N/A446 Abnormality of the orbital region HP:0000315N/A884 Hypertelorism HP:0000316 0.64± 0.81 [-1.17, 2.44]381 Short philtrum HP:0000322 0.73± 0.70 [-0.73, 2.19]231 Triangular face HP:0000325 0.48± 0.88 [-1.29, 2.26]195 Prominent supraorbital ridges HP:0000336 0.69± 0.73 [-0.79, 2.17]145 Broad forehead HP:0000337 0.54± 0.88 [-1.37, 2.45]303 Narrow forehead HP:0000341 0.54± 0.88 [-1.37, 2.45]154 Long philtrum HP:0000343 0.73± 0.70 [-0.73, 2.19]382 High forehead HP:0000348 0.50± 0.88 [-1.27, 2.27]337 Abnormality of the nose HP:0000366N/A1019 Abnormal nasal bridge morphology HP:0000422N/A667 Abnormal morphology of the nasal alae HP:0000429N/A605 Underdeveloped nasal alae HP:0000430 0.27± 0.97 [-1.69, 2.24]227 Wide nasal bridge HP:0000431 0.78± 0.65 [-0.59, 2.14]400 Continued on next page 22 Table A1: The frequencies of all annotated HPO terms and their corresponding Inter-Rater-Agreement (IRA) using the mean Fleiss’ Kappa score for each HPO term over all samples. HPO NameHPO IDIRASamples Abnormal nasal tip morphology HP:0000436N/A587 Depressed nasal tip HP:0000437 0.62± 0.79 [-0.99, 2.22]142 Narrow nasal bridge HP:0000446 0.78± 0.65 [-0.59, 2.14]100 Broad nasal tip HP:0000455 0.43± 0.91 [-1.40, 2.26]494 Anteverted nares HP:0000463 0.57± 0.83 [-1.11, 2.25]388 Abnormality of the eye HP:0000478N/A746 Strabismus HP:0000486 0.50± 0.88 [-1.26, 2.26]273 Abnormal eyelid morphology HP:0000492N/A883 Downslanted palpebral fissures HP:00004941.00± 0.00359 Abnormality of eye movement HP:0000496N/A277 Abnormal eyelash morphology HP:0000499N/A302 Ptosis HP:0000508 0.21± 0.99 [-1.79, 2.21]262 Proptosis HP:0000520 0.86± 0.51 [-0.16, 1.89]57 Abnormality iris morphology HP:0000525 0.88± 0.49 [-0.15, 1.91]45 Abnormal eyebrow morphology HP:0000534N/A830 Abnormal conjugate eye movement HP:0000549N/A273 Abnormal uvea morphology HP:0000553N/A45 Microphthalmia HP:0000568 0.81± 0.59 [-0.39, 2.01]33 Thick eyebrow HP:0000574 0.81± 0.59 [-0.37, 2.00]347 Blepharophimosis HP:0000581 0.71± 0.71 [-0.71, 2.14]76 Upslanted palpebral fissure HP:00005821.00± 0.00154 Abnormality of the frontal hairline HP:0000599N/A546 Hypotelorism HP:0000601 0.64± 0.81 [-1.17, 2.44]85 Abnormality of the periorbital region HP:0000606N/A206 Long palpebral fissure HP:0000637 0.40± 0.97 [-1.79, 2.59]253 Synophrys HP:0000664 0.86± 0.51 [-0.15, 1.88]208 Abnormal facial shape HP:0001999 0.03± 1.01 [-2.03, 2.09]688 Short columella HP:0002000 0.07± 1.01 [-2.01, 2.14]307 Frontal bossing HP:0002007 0.31± 0.97 [-1.67, 2.29]196 Highly arched eyebrow HP:0002553 0.36± 0.94 [-1.53, 2.24]349 Downturned corners of mouth HP:00027141.00± 0.00256 Long nose HP:00031891.00± 0.00115 Short nose HP:00031961.00± 0.00407 Abnormal cheek morphology HP:0004426N/A356 Elfin facies HP:0004428 0.42± 0.92 [-1.44, 2.28]108 Abnormal nasal morphology HP:0005105N/A1019 Depressed nasal bridge HP:0005280 0.51± 0.87 [-1.24, 2.26]378 Abnormal nostril morphology HP:0005288N/A174 Continued on next page 23 Table A1: The frequencies of all annotated HPO terms and their corresponding Inter-Rater-Agreement (IRA) using the mean Fleiss’ Kappa score for each HPO term over all samples. HPO NameHPO IDIRASamples Abnormality of the palpebral fissures HP:0008050N/A703 Aplasia/Hypoplasia affecting the eye HP:0008056N/A33 High anterior hairline HP:0009890 0.83± 0.56 [-0.33, 2.00]278 Thick nasal alae HP:0009928 0.36± 0.93 [-1.53, 2.25]194 Abnormal columella morphology HP:0009929N/A320 Enlarged naris HP:0009931 0.37± 0.93 [-1.50, 2.24]172 Everted upper lip vermilion HP:0010803 0.57± 0.83 [-1.13, 2.26]136 Upturned corners of mouth HP:00108051.00± 0.00122 Abnormal external nose morphology HP:0010938N/A861 Prominent eyelashes HP:0011231 0.82± 0.58 [-0.35, 1.99]255 Abnormality of mouth size HP:0011337N/A606 Abnormality of mouth shape HP:0011338N/A392 Abnormality of upper lip vermillion HP:0011339N/A176 Midface retrusion HP:0011800 0.41± 0.92 [-1.45, 2.27]117 Narrow philtrum HP:00118291.00± 0.0085 Flat face HP:0012368 0.61± 0.80 [-1.02, 2.24]97 Abnormal eye morphology HP:0012372N/A545 Abnormal eye physiology HP:0012373N/A487 Thick vermilion border HP:0012471 0.74± 0.68 [-0.65, 2.13]219 Short palpebral fissure HP:0012745 0.40± 0.97 [-1.79, 2.59]188 Abnormal nasal base norphology HP:0012808N/A376 Wide nasal base HP:0012810 0.44± 0.90 [-1.37, 2.26]376 Abnormal ocular adnexa morphology HP:0030669N/A884 Abnormal oral morphology HP:0031816N/A1088 Abnormality of the ocular adnexa HP:0032039N/A884 Sparse eyebrow HP:0045075 0.81± 0.59 [-0.37, 2.00]351 Abnormality of the supraorbital ridges HP:0100538N/A145 Aplasia/Hypoplasia of the eyebrow HP:0100840N/A358 Abnormality of globe location HP:0100886N/A499 Slanting of the palpebral fissure HP:0200006N/A507 Abnormal size of the palpebral fissures HP:0200007N/A450 Appendix B Ablation Study 24 Table B2: The results of all experiments from the ablation study. D: Dimension, FO: Face Outline, T: Feature Importance Threshold, S: Softlabel DFOMetadataTSmean AUROC 3Trueage, gender, ethnicity0.010.050.750±0.042 3Trueage, gender, ethnicity0.010.000.749±0.044 3Trueage, gender, ethnicity0.050.100.747±0.044 3Trueage, gender, ethnicity0.010.100.747±0.043 3Trueage, gender, ethnicity0.050.050.741±0.043 3Falseage, gender, ethnicity0.010.000.740±0.045 3Falseage, gender, ethnicity0.010.050.740±0.041 2Trueage, gender, ethnicity0.010.100.739±0.043 3Falseage, gender, ethnicity0.010.100.739±0.043 3Trueage, gender, ethnicity0.050.000.738±0.046 2Trueage, gender, ethnicity0.010.050.737±0.044 2Trueage, gender, ethnicity0.010.000.735±0.044 2Falseage, gender, ethnicity0.010.100.734±0.045 2Falseage, gender, ethnicity0.010.050.732±0.046 3Falseage, gender, ethnicity0.050.050.731±0.044 2Falseage, gender, ethnicity0.010.000.731±0.044 3Trueage, gender, ethnicity0.100.100.728±0.046 3Falseage, gender, ethnicity0.050.000.727±0.046 3Falseage, gender, ethnicity0.050.100.726±0.044 3Trueage, gender, ethnicity0.100.050.725±0.049 3Trueage, gender, ethnicity0.100.000.716±0.048 3Falseage, gender, ethnicity0.100.050.709±0.053 3TrueN/A0.010.050.709±0.048 3TrueN/A0.010.000.708±0.046 3Falseage, gender, ethnicity0.100.000.708±0.049 3TrueN/A0.010.100.707±0.046 3Falseage, gender, ethnicity0.100.100.706±0.049 2Trueage, gender, ethnicity0.050.000.706±0.054 2Trueage, gender, ethnicity0.050.100.706±0.052 2Falseage, gender, ethnicity0.050.100.705±0.050 2Trueage, gender, ethnicity0.050.050.705±0.048 3TrueN/A0.050.050.704±0.045 2Falseage, gender, ethnicity0.050.050.701±0.051 2Falseage, gender, ethnicity0.050.000.700±0.048 3FalseN/A0.010.100.700±0.043 3TrueN/A0.050.000.696±0.046 2TrueN/A0.010.000.695±0.045 3FalseN/A0.010.050.695±0.045 Continued on next page 25 Table B2: The results of all experiments from the ablation study. D: Dimension, FO: Face Outline, T: Feature Importance Threshold, S: Softlabel DFOMetadataTSmean AUROC 3FalseN/A0.010.000.695±0.044 2TrueN/A0.010.050.694±0.048 3TrueN/A0.050.100.693±0.050 2TrueN/A0.010.100.693±0.048 2Trueage, gender, ethnicity0.100.100.692±0.048 2Falseage, gender, ethnicity0.100.050.690±0.055 2Falseage, gender, ethnicity0.100.000.689±0.051 2FalseN/A0.010.000.688±0.050 2FalseN/A0.010.050.688±0.052 2Falseage, gender, ethnicity0.100.100.687±0.050 3FalseN/A0.050.000.687±0.047 2Trueage, gender, ethnicity0.100.050.685±0.055 2FalseN/A0.010.100.685±0.047 3FalseN/A0.050.050.683±0.052 3TrueN/A0.100.100.680±0.051 3FalseN/A0.050.100.680±0.049 3TrueN/A0.100.050.678±0.052 3TrueN/A0.100.000.678±0.053 2Trueage, gender, ethnicity0.100.000.677±0.050 3FalseN/A0.100.000.673±0.048 3FalseN/A0.100.100.670±0.047 3FalseN/A0.100.050.668±0.054 2TrueN/A0.050.000.667±0.050 2TrueN/A0.050.100.666±0.048 2TrueN/A0.050.050.666±0.053 2FalseN/A0.050.100.665±0.050 2FalseN/A0.050.000.662±0.055 2FalseN/A0.050.050.655±0.049 2FalseN/A0.100.000.655±0.055 2FalseN/A0.100.100.653±0.055 2FalseN/A0.100.050.651±0.055 2TrueN/A0.100.000.648±0.051 2TrueN/A0.100.100.646±0.050 2TrueN/A0.100.050.644±0.053 26 Appendix C Best Configuration Analysis Table C3 : The results of all HPO models with the best configu- ration. The HPO terms in bold letters are leafs of the human phe-notype ontology. N: Samples, Prev: Prevalence, Det. P: DetectionPrevalence AUROC F1-Score Precision Recall N Det. P Abnormal morphology of the nasal alae 0.89 ± 0.02 0.89 ± 0.02 0.89 ± 0.02 0.89 ± 0.03 1210 0.50 Abnormal oral morphology 0.89 ± 0.02 0.89 ± 0.02 0.91 ± 0.01 0.86 ± 0.03 2176 0.47 Abnormal cheek morphology 0.89 ± 0.02 0.89 ± 0.02 0.91 ± 0.03 0.86 ± 0.04 712 0.47 Abnormality of the eye 0.89 ± 0.02 0.89 ± 0.02 0.91 ± 0.02 0.86 ± 0.02 1492 0.48 Abnormal external nose morphology 0.89 ± 0.01 0.89 ± 0.01 0.90 ± 0.01 0.87 ± 0.02 1722 0.49 Abnormal eyebrow morphology 0.89 ± 0.03 0.88 ± 0.03 0.91 ± 0.04 0.86 ± 0.03 1660 0.47 Abnormality of the nose 0.88 ± 0.01 0.88 ± 0.01 0.90 ± 0.03 0.86 ± 0.03 2038 0.48 Abnormality of the mouth 0.88 ± 0.01 0.88 ± 0.02 0.90 ± 0.01 0.86 ± 0.02 2178 0.47 Abnormal eye morphology 0.88 ± 0.02 0.88 ± 0.03 0.91 ± 0.04 0.85 ± 0.04 1090 0.46 Abnormal nasal morphology 0.88 ± 0.02 0.88 ± 0.02 0.90 ± 0.03 0.86 ± 0.02 2038 0.48 Abnormality of the face 0.88 ± 0.01 0.88 ± 0.01 0.90 ± 0.03 0.87 ± 0.03 2460 0.48 Abnormality of globe location 0.88 ± 0.03 0.88 ± 0.03 0.90 ± 0.04 0.87 ± 0.06 998 0.48 Abnormal midface morphology 0.88 ± 0.02 0.88 ± 0.02 0.88 ± 0.02 0.88 ± 0.03 892 0.50 Abnormal upper lip morphology 0.88 ± 0.02 0.88 ± 0.02 0.90 ± 0.02 0.86 ± 0.04 1520 0.48 Abnormal nasal bridge morphology 0.88 ± 0.04 0.87 ± 0.04 0.91 ± 0.04 0.84 ± 0.06 1334 0.46 Abnormal lip morphology 0.88 ± 0.03 0.88 ± 0.03 0.88 ± 0.05 0.88 ± 0.02 2004 0.50 Aplasia/Hypoplasia of the eyebrow 0.88 ± 0.04 0.88 ± 0.03 0.88 ± 0.07 0.88 ± 0.01 716 0.50 Abnormality of the palpebral fissures 0.88 ± 0.03 0.87 ± 0.03 0.90 ± 0.02 0.85 ± 0.05 1406 0.47 Abnormality of the philtrum 0.88 ± 0.01 0.87 ± 0.01 0.89 ± 0.03 0.85 ± 0.03 1444 0.48 Continued on next page 27 Table C3 : The results of all HPO models with the best configu- ration. The HPO terms in bold letters are leafs of the human phe-notype ontology. N: Samples, Prev: Prevalence, Det. P: DetectionPrevalence AUROC F1-Score Precision Recall N Det. P Abnormal lower lip morphology 0.88 ± 0.01 0.87 ± 0.01 0.90 ± 0.02 0.84 ± 0.03 786 0.46 Abnormal eyelid morphology 0.88 ± 0.02 0.87 ± 0.03 0.90 ± 0.01 0.84 ± 0.05 1766 0.46 Abnormality of mouth size 0.88 ± 0.02 0.87 ± 0.02 0.89 ± 0.03 0.86 ± 0.05 1212 0.48 Abnormal columella morphology 0.87 ± 0.03 0.87 ± 0.03 0.90 ± 0.03 0.84 ± 0.04 640 0.47 Slanting of the palpebral fissure 0.87 ± 0.02 0.87 ± 0.02 0.88 ± 0.04 0.87 ± 0.03 1014 0.49 Abnormal size of the palpebral fissures 0.87 ± 0.03 0.87 ± 0.03 0.89 ± 0.03 0.85 ± 0.02 900 0.48 Abnormality of mouth shape 0.87 ± 0.01 0.87 ± 0.01 0.89 ± 0.03 0.85 ± 0.03 784 0.48 Abnormal nasal base norphology 0.87 ± 0.03 0.87 ± 0.03 0.89 ± 0.03 0.84 ± 0.07 752 0.47 Abnormality of eye movement 0.87 ± 0.03 0.87 ± 0.02 0.89 ± 0.06 0.85 ± 0.06 554 0.48 Abnormal ocular adnexa morphology 0.87 ± 0.04 0.86 ± 0.05 0.89 ± 0.03 0.84 ± 0.11 1768 0.47 Abnormal nasal tip morphology 0.87 ± 0.03 0.87 ± 0.03 0.89 ± 0.03 0.85 ± 0.04 1174 0.48 Abnormality of the orbital region 0.87 ± 0.05 0.86 ± 0.05 0.89 ± 0.03 0.84 ± 0.08 1768 0.47 Abnormal conjugate eye movement 0.86 ± 0.04 0.85 ± 0.06 0.90 ± 0.04 0.82 ± 0.10 546 0.46 Abnormal oral cavity morphology 0.86 ± 0.05 0.86 ± 0.04 0.87 ± 0.07 0.86 ± 0.04 2176 0.50 Abnormal eye physiology 0.86 ± 0.03 0.86 ± 0.03 0.86 ± 0.04 0.86 ± 0.05 974 0.50 Abnormality of the chin 0.86 ± 0.05 0.85 ± 0.06 0.91 ± 0.05 0.80 ± 0.07 646 0.44 Abnormality of the periorbital region 0.85 ± 0.03 0.85 ± 0.04 0.85 ± 0.05 0.86 ± 0.07 412 0.50 Abnormality of the ocular adnexa 0.85 ± 0.04 0.85 ± 0.04 0.87 ± 0.05 0.84 ± 0.05 1768 0.48 Abnormal eyelash morphology 0.85 ± 0.05 0.84 ± 0.06 0.89 ± 0.05 0.81 ± 0.07 604 0.46 Abnormality of upper lip vermillion 0.84 ± 0.07 0.83 ± 0.08 0.86 ± 0.06 0.80 ± 0.12 352 0.47 Abnormal uvea morphology 0.83 ± 0.06 0.82 ± 0.06 0.83 ± 0.07 0.83 ± 0.14 90 0.50 Abnormality of the supraorbital ridges 0.81 ± 0.04 0.81 ± 0.04 0.83 ± 0.07 0.79 ± 0.03 290 0.48 Continued on next page 28 Table C3 : The results of all HPO models with the best configu- ration. The HPO terms in bold letters are leafs of the human phe-notype ontology. N: Samples, Prev: Prevalence, Det. P: DetectionPrevalence AUROC F1-Score Precision Recall N Det. P Abnormal nostril morphology 0.80 ± 0.02 0.78 ± 0.02 0.86 ± 0.03 0.72 ± 0.04 348 0.42 Short nose 0.80 ± 0.05 0.81 ± 0.04 0.77 ± 0.01 0.84 ± 0.08 814 0.54 Wide nasal bridge 0.80 ± 0.03 0.81 ± 0.05 0.77 ± 0.03 0.85 ± 0.08 800 0.55 Hypertelorism 0.79 ± 0.04 0.80 ± 0.06 0.77 ± 0.05 0.83 ± 0.07 762 0.54 Abnormal forehead morphology 0.79 ± 0.04 0.80 ± 0.05 0.77 ± 0.03 0.84 ± 0.09 1530 0.54 Aplasia/Hypoplasia affecting the eye 0.79 ± 0.19 0.76 ± 0.24 0.85 ± 0.22 0.77 ± 0.30 66 0.47 Downturned corners of mouth 0.78 ± 0.03 0.78 ± 0.03 0.81 ± 0.08 0.75 ± 0.05 512 0.47 Abnormal facial shape 0.77 ± 0.02 0.79 ± 0.02 0.73 ± 0.04 0.86 ± 0.02 1376 0.59 Thin vermilion border 0.77 ± 0.02 0.79 ± 0.02 0.74 ± 0.02 0.85 ± 0.03 1198 0.57 Thin upper lip vermilion 0.76 ± 0.02 0.78 ± 0.03 0.72 ± 0.04 0.85 ± 0.05 1116 0.59 Depressed nasal bridge 0.75 ± 0.04 0.78 ± 0.04 0.71 ± 0.03 0.87 ± 0.05 756 0.61 Narrow mouth 0.75 ± 0.05 0.76 ± 0.07 0.73 ± 0.05 0.80 ± 0.13 304 0.55 Broad philtrum 0.73 ± 0.04 0.75 ± 0.04 0.70 ± 0.06 0.81 ± 0.02 572 0.58 Sparse eyebrow 0.73 ± 0.03 0.75 ± 0.06 0.69 ± 0.05 0.82 ± 0.08 702 0.60 Downslanted palpebral fissures 0.72 ± 0.04 0.74 ± 0.05 0.69 ± 0.02 0.80 ± 0.08 718 0.57 Abnormality of the frontal hairline 0.72 ± 0.01 0.75 ± 0.01 0.67 ± 0.02 0.86 ± 0.02 1092 0.64 Wide mouth 0.72 ± 0.04 0.75 ± 0.05 0.66 ± 0.06 0.88 ± 0.05 918 0.67 Anteverted nares 0.71 ± 0.02 0.72 ± 0.03 0.70 ± 0.02 0.75 ± 0.06 788 0.53 Full cheeks 0.71 ± 0.03 0.74 ± 0.04 0.67 ± 0.03 0.81 ± 0.05 702 0.61 Prominent eyelashes 0.71 ± 0.06 0.71 ± 0.08 0.70 ± 0.07 0.72 ± 0.10 510 0.52 Long philtrum 0.70 ± 0.02 0.72 ± 0.04 0.68 ± 0.04 0.76 ± 0.04 764 0.56 Broad forehead 0.70 ± 0.03 0.72 ± 0.04 0.67 ± 0.05 0.77 ± 0.06 606 0.58 Continued on next page 29 Table C3 : The results of all HPO models with the best configu- ration. The HPO terms in bold letters are leafs of the human phe-notype ontology. N: Samples, Prev: Prevalence, Det. P: DetectionPrevalence AUROC F1-Score Precision Recall N Det. P Triangular face 0.69 ± 0.05 0.69 ± 0.08 0.70 ± 0.04 0.68 ± 0.12 390 0.49 Wide nasal base 0.69 ± 0.01 0.73 ± 0.01 0.64 ± 0.02 0.86 ± 0.03 752 0.66 Low anterior hairline 0.69 ± 0.03 0.71 ± 0.02 0.67 ± 0.05 0.76 ± 0.04 556 0.57 Highly arched eyebrow 0.69 ± 0.03 0.72 ± 0.03 0.65 ± 0.04 0.80 ± 0.04 698 0.62 Short columella 0.68 ± 0.05 0.71 ± 0.05 0.66 ± 0.06 0.77 ± 0.08 614 0.58 Pointed chin 0.68 ± 0.04 0.70 ± 0.05 0.66 ± 0.07 0.75 ± 0.05 584 0.57 Epicanthus 0.68 ± 0.05 0.69 ± 0.07 0.67 ± 0.08 0.71 ± 0.07 526 0.53 Broad nasal tip 0.68 ± 0.03 0.73 ± 0.04 0.63 ± 0.03 0.87 ± 0.09 988 0.70 High forehead 0.68 ± 0.02 0.70 ± 0.05 0.65 ± 0.03 0.76 ± 0.08 674 0.59 Long palpebral fissure 0.68 ± 0.06 0.69 ± 0.09 0.65 ± 0.07 0.74 ± 0.12 506 0.57 Proptosis 0.67 ± 0.12 0.67 ± 0.13 0.64 ± 0.10 0.73 ± 0.21 114 0.55 Everted upper lip vermilion 0.67 ± 0.07 0.69 ± 0.08 0.65 ± 0.05 0.73 ± 0.12 272 0.56 Depressed nasal tip 0.67 ± 0.05 0.66 ± 0.06 0.68 ± 0.03 0.65 ± 0.10 284 0.48 Everted lower lip vermilion 0.66 ± 0.04 0.69 ± 0.04 0.63 ± 0.04 0.77 ± 0.10 546 0.61 High anterior hairline 0.66 ± 0.03 0.69 ± 0.04 0.63 ± 0.03 0.76 ± 0.06 556 0.60 Narrow philtrum 0.66 ± 0.09 0.69 ± 0.07 0.64 ± 0.12 0.76 ± 0.05 170 0.61 Thick nasal alae 0.65 ± 0.06 0.68 ± 0.07 0.62 ± 0.08 0.75 ± 0.07 394 0.61 Enlarged naris 0.64 ± 0.08 0.69 ± 0.09 0.61 ± 0.09 0.80 ± 0.11 344 0.67 Thick lower lip vermilion 0.64 ± 0.03 0.67 ± 0.03 0.61 ± 0.04 0.75 ± 0.04 442 0.61 Hypotelorism 0.64 ± 0.09 0.55 ± 0.32 0.55 ± 0.31 0.57 ± 0.35 170 0.45 Upslanted palpebral fissure 0.63 ± 0.06 0.68 ± 0.07 0.61 ± 0.06 0.76 ± 0.11 308 0.62 Short palpebral fissure 0.63 ± 0.05 0.66 ± 0.05 0.62 ± 0.06 0.70 ± 0.04 376 0.57 Continued on next page 30 Table C3 : The results of all HPO models with the best configu- ration. The HPO terms in bold letters are leafs of the human phe-notype ontology. N: Samples, Prev: Prevalence, Det. P: DetectionPrevalence AUROC F1-Score Precision Recall N Det. P Flat face 0.63 ± 0.13 0.58 ± 0.33 0.55 ± 0.32 0.63 ± 0.37 194 0.50 Short philtrum 0.63 ± 0.05 0.64 ± 0.04 0.63 ± 0.06 0.65 ± 0.03 462 0.52 Thick vermilion border 0.63 ± 0.04 0.69 ± 0.05 0.59 ± 0.09 0.87 ± 0.10 438 0.74 Strabismus 0.63 ± 0.04 0.66 ± 0.05 0.61 ± 0.02 0.71 ± 0.09 546 0.58 Narrow forehead 0.63 ± 0.07 0.60 ± 0.15 0.62 ± 0.06 0.62 ± 0.23 308 0.50 Coarse facial features 0.63 ± 0.03 0.68 ± 0.07 0.59 ± 0.06 0.81 ± 0.08 560 0.69 Prominent supraorbital ridges 0.63 ± 0.05 0.65 ± 0.09 0.61 ± 0.06 0.70 ± 0.13 290 0.58 Narrow nasal bridge 0.62 ± 0.03 0.67 ± 0.06 0.60 ± 0.05 0.77 ± 0.16 200 0.64 Underdeveloped nasal alae 0.62 ± 0.07 0.65 ± 0.07 0.60 ± 0.08 0.72 ± 0.06 462 0.60 Thick eyebrow 0.62 ± 0.05 0.64 ± 0.07 0.60 ± 0.06 0.70 ± 0.13 694 0.58 Ptosis 0.61 ± 0.10 0.54 ± 0.31 0.51 ± 0.30 0.58 ± 0.33 524 0.47 Narrow face 0.61 ± 0.09 0.65 ± 0.11 0.59 ± 0.07 0.73 ± 0.19 222 0.62 Abnormality iris morphology 0.61 ± 0.07 0.60 ± 0.20 0.68 ± 0.22 0.71 ± 0.39 90 0.59 Midface retrusion 0.61 ± 0.06 0.66 ± 0.07 0.58 ± 0.06 0.77 ± 0.11 234 0.67 Small face 0.61 ± 0.07 0.52 ± 0.29 0.50 ± 0.29 0.54 ± 0.32 222 0.43 Frontal bossing 0.60 ± 0.07 0.53 ± 0.30 0.50 ± 0.29 0.56 ± 0.32 392 0.47 Upturned corners of mouth 0.60 ± 0.04 0.62 ± 0.07 0.60 ± 0.07 0.65 ± 0.13 244 0.55 Long nose 0.60 ± 0.05 0.62 ± 0.04 0.59 ± 0.04 0.67 ± 0.08 230 0.57 Synophrys 0.59 ± 0.03 0.58 ± 0.05 0.60 ± 0.05 0.59 ± 0.11 416 0.49 Microphthalmia 0.58 ± 0.11 0.42 ± 0.31 0.58 ± 0.39 0.42 ± 0.33 66 0.36 Elfin facies 0.55 ± 0.07 0.28 ± 0.32 0.43 ± 0.43 0.29 ± 0.38 216 0.24 Blepharophimosis 0.55 ± 0.03 0.47 ± 0.20 0.66 ± 0.20 0.47 ± 0.27 152 0.42 31 Appendix D Evaluation of Generalizability Table D4 : An overview of the mean F1-Score performance dif- ference (test set vs. validation set) of the HPO models with aprevalence greater than zero. A number below zero means the per-formance was better on the test set, while a number above zeromeans the performance on the validation set was better. The labelsin bold letters are leaf nodes of the Human Phenotype Ontology.The names of the unseen disorders are as follows: Susceptibilityto Autism (StA), FBXW7 syndrome (FBXW7S), Floating-Harborsyndrome (FLHS), Mowat-Wilson syndrome (MOWS), Nicolaides-Baraitser syndrome (NCBRS), Ohdo syndrome, SBBYS-variant(SBBYSS), Opitz GBBB syndrome (OGBBBS), Sotos syndrome(SOTOS), Seckel syndrome (S), and White-Sutton syndrome(WHSUS). HPO StA FBXW7S FLHS MOWS NCBRS SBBYSS OGBBBS SOTOS S WHSUS Abnormality of the face -0.12 -0.12 0.31 -0.01 0.02 -0.12 -0.12 -0.12 0.21 0.38 Abnormality of the mouth 0.21 0.31 0.21 0.13 0.21 0.13 0.31 N/A -0.12 0.38 Abnormal oral morphology 0.22 0.32 0.22 0.14 0.22 0.14 0.32 N/A -0.11 0.39 Abnormal oral cavity morphology 0.19 0.29 0.19 0.11 0.19 0.11 0.29 N/A -0.14 0.36 Abnormal nasal morphology N/A -0.01 0.13 0.21 0.02 -0.01 -0.12 N/A N/A 0.08 Abnormal lip morphology 0.21 0.55 0.08 0.31 0.21 N/A 0.31 N/A 0.21 0.38 Abnormality of the nose N/A -0.01 0.55 0.21 0.02 -0.01 -0.12 N/A N/A 0.38 Abnormality of the eye 0.22 0.32 N/A -0.11 0.89 -0.11 -0.11 -0.11 N/A 0.22 Abnormal eye physiology N/A 0.19 N/A -0.14 0.86 0.29 0.11 -0.14 N/A 0.19 Abnormal external nose morphology N/A 0.56 0.56 0.49 -0.11 N/A -0.11 N/A N/A 0.39 Continued on next page 32 Table D4 : An overview of the mean F1-Score performance dif- ference (test set vs. validation set) of the HPO models with aprevalence greater than zero. A number below zero means the per-formance was better on the test set, while a number above zeromeans the performance on the validation set was better. The labelsin bold letters are leaf nodes of the Human Phenotype Ontology.The names of the unseen disorders are as follows: Susceptibilityto Autism (StA), FBXW7 syndrome (FBXW7S), Floating-Harborsyndrome (FLHS), Mowat-Wilson syndrome (MOWS), Nicolaides-Baraitser syndrome (NCBRS), Ohdo syndrome, SBBYS-variant(SBBYSS), Opitz GBBB syndrome (OGBBBS), Sotos syndrome(SOTOS), Seckel syndrome (S), and White-Sutton syndrome(WHSUS). HPO StA FBXW7S FLHS MOWS NCBRS SBBYSS OGBBBS SOTOS S WHSUS Abnormal facial shape N/A 0.46 N/A N/A -0.01 N/A 0.22 -0.07 -0.21 -0.01 Abnormality of the orbital region N/A 0.46 N/A N/A N/A 0.29 0.29 -0.14 N/A N/A Abnormality of the ocular adnexa N/A 0.52 N/A N/A N/A 0.28 0.28 -0.15 N/A N/A Abnormal eyelid morphology N/A 0.54 N/A N/A N/A 0.54 0.30 -0.13 N/A N/A Abnormal morphology of the nasal alae N/A 0.56 N/A N/A -0.11 N/A -0.11 N/A N/A 0.22 Abnormal ocular adnexa morphology N/A 0.53 N/A N/A N/A 0.53 0.19 0.00 N/A N/A Abnormal conjugate eye movement N/A 0.35 N/A -0.15 0.85 N/A 0.10 -0.01 N/A 0.85 Abnormality of eye movement N/A 0.47 N/A -0.13 0.87 N/A 0.12 0.01 N/A 0.87 Abnormal upper lip morphology N/A N/A 0.08 N/A 0.38 N/A 0.31 N/A 0.21 N/A Abnormal eyebrow morphology N/A 0.55 N/A N/A -0.12 -0.01 N/A 0.21 N/A 0.88 Strabismus N/A 0.16 N/A -0.20 0.66 N/A -0.09 -0.01 N/A 0.66 Ptosis N/A -0.13 N/A N/A 0.54 0.21 N/A -0.13 N/A -0.46 Abnormality of the philtrum N/A N/A 0.07 N/A 0.37 N/A 0.30 N/A 0.20 N/A Continued on next page 33 Table D4 : An overview of the mean F1-Score performance dif- ference (test set vs. validation set) of the HPO models with aprevalence greater than zero. A number below zero means the per-formance was better on the test set, while a number above zeromeans the performance on the validation set was better. The labelsin bold letters are leaf nodes of the Human Phenotype Ontology.The names of the unseen disorders are as follows: Susceptibilityto Autism (StA), FBXW7 syndrome (FBXW7S), Floating-Harborsyndrome (FLHS), Mowat-Wilson syndrome (MOWS), Nicolaides-Baraitser syndrome (NCBRS), Ohdo syndrome, SBBYS-variant(SBBYSS), Opitz GBBB syndrome (OGBBBS), Sotos syndrome(SOTOS), Seckel syndrome (S), and White-Sutton syndrome(WHSUS). HPO StA FBXW7S FLHS MOWS NCBRS SBBYSS OGBBBS SOTOS S WHSUS Abnormal nasal bridge morphology N/A 0.30 0.54 0.20 0.47 -0.02 -0.13 N/A N/A N/A Thin vermilion border N/A 0.79 N/A N/A 0.12 N/A N/A N/A 0.12 0.29 Anteverted nares N/A N/A N/A N/A -0.08 N/A -0.28 N/A N/A 0.72 Abnormal lower lip morphology 0.20 0.54 N/A 0.30 0.20 N/A N/A N/A N/A 0.20 Abnormality of the palpebral fissures N/A 0.54 N/A N/A N/A N/A 0.30 -0.13 N/A N/A Thick lower lip vermilion 0.00 0.67 N/A 0.10 N/A N/A N/A N/A N/A 0.00 Long philtrum N/A N/A N/A N/A 0.22 N/A 0.39 N/A -0.28 N/A Abnormal forehead morphology N/A 0.47 N/A 0.80 N/A N/A N/A N/A N/A N/A Slanting of the palpebral fissure N/A N/A N/A N/A N/A N/A 0.30 -0.13 N/A N/A Abnormality of mouth size N/A 0.47 0.87 0.47 0.20 0.12 N/A N/A N/A N/A Abnormal eye morphology 0.21 N/A N/A N/A N/A -0.12 -0.12 N/A N/A N/A Abnormality of globe location 0.21 N/A N/A N/A N/A -0.12 -0.12 N/A N/A N/A Epicanthus N/A N/A N/A N/A N/A 0.36 N/A 0.19 N/A N/A Continued on next page 34 Table D4 : An overview of the mean F1-Score performance dif- ference (test set vs. validation set) of the HPO models with aprevalence greater than zero. A number below zero means the per-formance was better on the test set, while a number above zeromeans the performance on the validation set was better. The labelsin bold letters are leaf nodes of the Human Phenotype Ontology.The names of the unseen disorders are as follows: Susceptibilityto Autism (StA), FBXW7 syndrome (FBXW7S), Floating-Harborsyndrome (FLHS), Mowat-Wilson syndrome (MOWS), Nicolaides-Baraitser syndrome (NCBRS), Ohdo syndrome, SBBYS-variant(SBBYSS), Opitz GBBB syndrome (OGBBBS), Sotos syndrome(SOTOS), Seckel syndrome (S), and White-Sutton syndrome(WHSUS). HPO StA FBXW7S FLHS MOWS NCBRS SBBYSS OGBBBS SOTOS S WHSUS Coarse facial features N/A 0.18 N/A N/A N/A N/A N/A 0.01 N/A N/A Wide nasal bridge N/A N/A N/A 0.14 0.81 -0.08 -0.19 N/A N/A N/A Downslanted palpebral fissures N/A N/A N/A N/A N/A N/A 0.17 -0.26 N/A N/A Thin upper lip vermilion N/A 0.78 N/A N/A N/A N/A N/A N/A 0.78 0.28 Hypertelorism 0.13 N/A N/A N/A N/A -0.09 -0.20 N/A N/A N/A Thick eyebrow N/A 0.24 N/A N/A N/A N/A N/A 0.64 N/A N/A Wide mouth N/A N/A -0.05 N/A 0.08 0.18 N/A N/A N/A N/A Short nose N/A 0.81 N/A N/A N/A N/A N/A N/A N/A N/A Everted lower lip vermilion N/A N/A N/A 0.69 0.02 N/A N/A N/A N/A 0.02 Broad nasal tip N/A N/A 0.40 0.33 N/A N/A N/A N/A N/A 0.73 Abnormal nasal tip morphology N/A N/A 0.54 0.47 N/A N/A N/A N/A N/A 0.87 Aplasia/Hypoplasia of the eyebrow N/A N/A N/A N/A N/A 0.13 N/A 0.48 N/A 0.88 Abnormal size of the palpebral fissures N/A 0.47 N/A N/A N/A N/A N/A N/A N/A N/A Continued on next page 35 Table D4 : An overview of the mean F1-Score performance dif- ference (test set vs. validation set) of the HPO models with aprevalence greater than zero. A number below zero means the per-formance was better on the test set, while a number above zeromeans the performance on the validation set was better. The labelsin bold letters are leaf nodes of the Human Phenotype Ontology.The names of the unseen disorders are as follows: Susceptibilityto Autism (StA), FBXW7 syndrome (FBXW7S), Floating-Harborsyndrome (FLHS), Mowat-Wilson syndrome (MOWS), Nicolaides-Baraitser syndrome (NCBRS), Ohdo syndrome, SBBYS-variant(SBBYSS), Opitz GBBB syndrome (OGBBBS), Sotos syndrome(SOTOS), Seckel syndrome (S), and White-Sutton syndrome(WHSUS). HPO StA FBXW7S FLHS MOWS NCBRS SBBYSS OGBBBS SOTOS S WHSUS Long palpebral fissure N/A 0.29 N/A N/A N/A N/A N/A N/A N/A N/A Depressed nasal bridge N/A -0.02 N/A N/A N/A N/A N/A N/A N/A N/A Abnormality of the frontal hairline N/A 0.35 N/A N/A N/A N/A N/A N/A N/A N/A Narrow mouth N/A 0.26 0.76 0.36 N/A 0.76 N/A N/A N/A N/A Downturned corners of mouth N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Abnormality of mouth shape N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A High forehead N/A N/A N/A 0.70 N/A N/A N/A N/A N/A N/A Triangular face N/A N/A N/A N/A 0.19 N/A N/A N/A N/A N/A Highly arched eyebrow N/A N/A N/A N/A -0.28 N/A N/A 0.72 N/A N/A Narrow face N/A N/A N/A N/A N/A N/A N/A N/A -0.35 N/A Low anterior hairline N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Small face N/A N/A N/A N/A N/A N/A N/A N/A -0.15 N/A Underdeveloped nasal alae N/A N/A N/A N/A N/A N/A 0.08 N/A N/A N/A Continued on next page 36 Table D4 : An overview of the mean F1-Score performance dif- ference (test set vs. validation set) of the HPO models with aprevalence greater than zero. A number below zero means the per-formance was better on the test set, while a number above zeromeans the performance on the validation set was better. The labelsin bold letters are leaf nodes of the Human Phenotype Ontology.The names of the unseen disorders are as follows: Susceptibilityto Autism (StA), FBXW7 syndrome (FBXW7S), Floating-Harborsyndrome (FLHS), Mowat-Wilson syndrome (MOWS), Nicolaides-Baraitser syndrome (NCBRS), Ohdo syndrome, SBBYS-variant(SBBYSS), Opitz GBBB syndrome (OGBBBS), Sotos syndrome(SOTOS), Seckel syndrome (S), and White-Sutton syndrome(WHSUS). HPO StA FBXW7S FLHS MOWS NCBRS SBBYSS OGBBBS SOTOS S WHSUS Wide nasal base N/A 0.23 N/A N/A N/A N/A N/A N/A N/A N/A Midface retrusion N/A 0.16 N/A N/A N/A N/A N/A N/A N/A N/A Abnormal nasal base norphology N/A 0.37 N/A N/A N/A N/A N/A N/A N/A N/A Abnormal midface morphology N/A 0.48 N/A N/A N/A N/A N/A N/A N/A N/A Proptosis N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Short philtrum N/A N/A -0.03 N/A N/A N/A N/A N/A N/A N/A Narrow nasal bridge N/A N/A -0.08 N/A 0.27 N/A N/A N/A N/A N/A Broad forehead N/A 0.22 N/A N/A N/A N/A N/A N/A N/A N/A Synophrys N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Thick nasal alae N/A 0.18 N/A N/A N/A N/A N/A N/A N/A N/A Abnormality of the chin N/A N/A N/A 0.10 N/A N/A N/A N/A N/A 0.85 Pointed chin N/A N/A N/A -0.05 N/A N/A N/A N/A N/A 0.70 Narrow forehead N/A 0.60 N/A N/A N/A N/A N/A N/A N/A N/A Continued on next page 37 Table D4 : An overview of the mean F1-Score performance dif- ference (test set vs. validation set) of the HPO models with aprevalence greater than zero. A number below zero means the per-formance was better on the test set, while a number above zeromeans the performance on the validation set was better. The labelsin bold letters are leaf nodes of the Human Phenotype Ontology.The names of the unseen disorders are as follows: Susceptibilityto Autism (StA), FBXW7 syndrome (FBXW7S), Floating-Harborsyndrome (FLHS), Mowat-Wilson syndrome (MOWS), Nicolaides-Baraitser syndrome (NCBRS), Ohdo syndrome, SBBYS-variant(SBBYSS), Opitz GBBB syndrome (OGBBBS), Sotos syndrome(SOTOS), Seckel syndrome (S), and White-Sutton syndrome(WHSUS). HPO StA FBXW7S FLHS MOWS NCBRS SBBYSS OGBBBS SOTOS S WHSUS Broad philtrum N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Short columella N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Abnormal columella morphology N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Elfin facies N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Abnormality of upper lip vermillion N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Everted upper lip vermilion N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Prominent eyelashes N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Upslanted palpebral fissure N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A Abnormal eyelash morphology N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A High anterior hairline N/A 0.19 N/A N/A N/A N/A N/A N/A N/A N/A Long nose N/A N/A 0.29 N/A N/A N/A N/A N/A N/A N/A 38 Table D5: An overview of the mean F1-Score performance dif- ferences (test set vs. validation set) of the HPO models with a prevalence greater than zero. A number below zero means the performance was better on the test set, while a number above zero means the performance on the validation set was better. The labels in bold letters are leaf nodes of the Human Pheno- type Ontology. The names of the seen disorders are as follows: Coffin-Siris syndrome (CSS), Cornelia de Lange syndrome (CdLS), Hyperphosphatasia with mental retardation syndrome (HPMRS), KBG syndrome (KBGS), Noonan syndrome (NS), Ogden syndrome (OGDNS), and Williams-Beuren syndrome (WBS). HPOCSSCdLSHPMRSKBGSNSOGDNSWBS Abnormality of the face-0.07-0.12-0.120.080.02-0.12-0.12 Abnormality of the mouth-0.07-0.12-0.010.380.080.02-0.12 Abnormal oral morphology-0.06-0.110.000.390.090.03-0.11 Abnormal oral cavity morphology-0.09-0.14-0.030.360.060.00-0.14 Abnormal nasal morphology -0.120.13-0.010.210.480.020.02 Abnormal lip morphology-0.07-0.120.550.380.080.02-0.12 Abnormality of the nose -0.070.13-0.010.210.380.020.02 Abnormality of the eye0.32N/A0.32N/A0.030.22-0.11 Abnormal eye physiology0.29N/A0.29N/A0.000.19-0.14 Abnormal external nose morphology -0.060.14N/A0.220.390.030.49 Abnormal facial shape-0.01N/AN/A0.290.12-0.07-0.21 Abnormality of the orbital region0.06N/A-0.14-0.140.190.190.19 Abnormality of the ocular adnexa0.10N/A-0.15-0.150.180.180.18 Abnormal eyelid morphology0.07N/A-0.13-0.130.200.200.20 Abnormal morphology of the nasal alae0.020.14N/A0.220.490.030.49 Abnormal ocular adnexa morphology0.06N/A-0.140.190.190.190.19 Abnormal conjugate eye movement0.39N/AN/AN/A0.180.45-0.15 Abnormality of eye movement0.41N/AN/AN/A0.200.47-0.13 Abnormal upper lip morphology -0.120.13N/AN/A0.380.020.02 Abnormal eyebrow morphology0.130.13N/A0.210.880.48N/A Strabismus0.16N/AN/AN/A-0.010.26-0.34 Ptosis0.29N/A-0.03N/A-0.320.040.14 Abnormality of the philtrum -0.130.12N/AN/A0.370.470.01 Abnormal nasal bridge morphology0.12N/A-0.02N/AN/AN/A0.47 Thin vermilion border0.12-0.210.460.29N/A0.39N/A Anteverted nares0.05-0.03N/A0.720.220.220.32 Abnormal lower lip morphology-0.02N/AN/AN/A0.07N/A-0.13 Abnormality of the palpebral fissures0.51N/A-0.02-0.130.200.20N/A Thick lower lip vermilion-0.22N/AN/AN/A-0.13N/A-0.33 Long philtrum0.500.15N/AN/AN/A0.32-0.14 Continued on next page 39 Table D5: An overview of the mean F1-Score performance dif- ferences (test set vs. validation set) of the HPO models with a prevalence greater than zero. A number below zero means the performance was better on the test set, while a number above zero means the performance on the validation set was better. The labels in bold letters are leaf nodes of the Human Pheno- type Ontology. The names of the seen disorders are as follows: Coffin-Siris syndrome (CSS), Cornelia de Lange syndrome (CdLS), Hyperphosphatasia with mental retardation syndrome (HPMRS), KBG syndrome (KBGS), Noonan syndrome (NS), Ogden syndrome (OGDNS), and Williams-Beuren syndrome (WBS). HPOCSSCdLSHPMRSKBGSNSOGDNSWBS Abnormal forehead morphology0.050.05N/AN/A-0.20-0.20N/A Slanting of the palpebral fissure0.54N/A0.30N/A0.200.20N/A Abnormality of mouth size0.07N/AN/AN/AN/AN/AN/A Abnormal eye morphologyN/AN/AN/AN/A0.020.48N/A Abnormality of globe locationN/AN/AN/AN/A0.020.48N/A Epicanthus0.69N/A-0.06N/A0.19N/A0.02 Coarse facial features-0.12N/AN/AN/AN/A-0.180.01 Wide nasal bridgeN/AN/A-0.08N/AN/AN/AN/A Downslanted palpebral fissures0.34N/AN/AN/A0.070.07N/A Thin upper lip vermilion0.07N/AN/A0.28N/A0.38N/A HypertelorismN/AN/AN/AN/A-0.06N/AN/A Thick eyebrow -0.220.07N/A0.64N/AN/AN/A Wide mouth0.00N/AN/AN/AN/AN/AN/A Short nose0.060.06N/AN/AN/AN/A-0.05 Everted lower lip vermilionN/AN/AN/AN/AN/AN/A-0.31 Broad nasal tip -0.15N/AN/AN/AN/AN/AN/A Abnormal nasal tip morphology0.05N/AN/AN/AN/AN/AN/A Aplasia/Hypoplasia of the eyebrowN/AN/AN/AN/AN/A0.48N/A Abnormal size of the palpebral fissuresN/AN/A-0.02-0.13N/AN/AN/A Long palpebral fissureN/AN/A-0.20-0.31N/AN/AN/A Depressed nasal bridge -0.04N/AN/AN/AN/AN/A0.38 Abnormality of the frontal hairline -0.070.00N/AN/AN/AN/AN/A Narrow mouthN/AN/AN/AN/AN/AN/AN/A Downturned corners of mouth0.530.210.03N/AN/AN/AN/A Abnormality of mouth shape0.670.300.12N/AN/AN/AN/A High foreheadN/AN/AN/AN/A-0.300.30N/A Triangular faceN/AN/AN/A0.69-0.11N/AN/A Highly arched eyebrowN/AN/AN/AN/A0.72N/AN/A Narrow faceN/AN/AN/AN/AN/AN/A-0.21 Low anterior hairline -0.09-0.04N/AN/AN/AN/AN/A Continued on next page 40 Table D5: An overview of the mean F1-Score performance dif- ferences (test set vs. validation set) of the HPO models with a prevalence greater than zero. A number below zero means the performance was better on the test set, while a number above zero means the performance on the validation set was better. The labels in bold letters are leaf nodes of the Human Pheno- type Ontology. The names of the seen disorders are as follows: Coffin-Siris syndrome (CSS), Cornelia de Lange syndrome (CdLS), Hyperphosphatasia with mental retardation syndrome (HPMRS), KBG syndrome (KBGS), Noonan syndrome (NS), Ogden syndrome (OGDNS), and Williams-Beuren syndrome (WBS). HPOCSSCdLSHPMRSKBGSNSOGDNSWBS Small faceN/AN/AN/AN/AN/AN/A-0.34 Underdeveloped nasal alaeN/AN/AN/AN/AN/A-0.21N/A Wide nasal base -0.09N/AN/AN/AN/AN/AN/A Midface retrusionN/AN/AN/AN/A-0.14N/AN/A Abnormal nasal base norphology0.12N/AN/AN/AN/AN/AN/A Abnormal midface morphologyN/AN/AN/AN/A0.08N/AN/A ProptosisN/AN/AN/AN/A-0.130.27N/A Short philtrumN/AN/AN/AN/A0.14N/AN/A Narrow nasal bridgeN/AN/AN/AN/AN/AN/AN/A Broad foreheadN/AN/AN/AN/A0.22N/AN/A SynophrysN/A0.25N/A-0.09N/AN/AN/A Thick nasal alae0.18N/AN/AN/AN/AN/AN/A Abnormality of the chinN/AN/AN/AN/AN/AN/AN/A Pointed chinN/AN/AN/AN/AN/AN/AN/A Narrow foreheadN/AN/AN/AN/AN/A-0.07N/A Broad philtrum -0.25N/AN/AN/AN/AN/AN/A Short columellaN/AN/AN/AN/AN/A-0.15N/A Abnormal columella morphologyN/AN/AN/AN/AN/A0.01N/A Elfin faciesN/AN/AN/AN/AN/AN/A-0.58 Abnormality of upper lip vermillionN/AN/AN/AN/AN/A0.16N/A Everted upper lip vermilionN/AN/AN/AN/AN/A0.02N/A Prominent eyelashes0.04N/AN/AN/AN/AN/AN/A Upslanted palpebral fissureN/AN/A0.01N/AN/AN/AN/A Abnormal eyelash morphology0.22N/AN/AN/AN/AN/AN/A High anterior hairlineN/AN/AN/AN/AN/AN/AN/A Long noseN/AN/AN/AN/AN/AN/AN/A 41 Table D6: Prediction results on the test set for Susceptibility to Autism. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0021.001.00 Abnormality of the mouth0.500.670.501.0020.501.00 Abnormal oral morphology0.500.670.501.0020.501.00 Abnormal oral cavity morphology0.500.670.501.0020.501.00 Abnormal lip morphology0.500.670.501.0020.501.00 Abnormal lower lip morphology0.500.670.501.0020.501.00 Thick lower lip vermilion0.500.670.501.0020.501.00 Abnormality of the eye0.500.670.501.0020.501.00 Abnormal eye morphology0.500.670.501.0020.501.00 Abnormality of globe location0.500.670.501.0020.501.00 Hypertelorism0.500.670.501.0020.501.00 Table D7: Prediction results on the test set for FBXW7 syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0051.001.00 Abnormality of the mouth0.500.570.401.0050.401.00 Abnormal oral morphology0.500.570.401.0050.401.00 Abnormal oral cavity morphology0.500.570.401.0050.401.00 Abnormal lip morphology0.500.330.201.0050.201.00 Abnormal lower lip morphology0.500.330.201.0050.201.00 Thick lower lip vermilion0.000.000.000.0050.200.80 Thin vermilion border0.120.000.000.0050.200.60 Thin upper lip vermilion0.120.000.000.0050.200.60 Abnormality of mouth size0.620.400.251.0050.200.80 Narrow mouth0.750.500.331.0050.200.60 Abnormal forehead morphology0.250.330.250.5050.400.80 Broad forehead0.750.500.331.0050.200.60 Narrow forehead0.120.000.000.0050.200.60 Abnormality of the frontal hairline0.620.400.251.0050.200.80 High anterior hairline0.750.500.331.0050.200.60 Abnormal midface morphology0.620.400.251.0050.200.80 Continued on next page 42 Table D7: Prediction results on the test set for FBXW7 syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Midface retrusion0.750.500.331.0050.200.60 Abnormality of the orbital region0.620.400.251.0050.200.80 Abnormality of the ocular adnexa0.500.330.201.0050.201.00 Abnormal ocular adnexa morphology0.500.330.201.0050.201.00 Abnormal eyelid morphology0.500.330.201.0050.201.00 Abnormality of the palpebral fissures0.500.330.201.0050.201.00 Abnormal size of the palpebral fissures0.620.400.251.0050.200.80 Long palpebral fissure0.620.400.251.0050.200.80 Abnormality of the nose0.500.890.801.0050.801.00 Abnormal nasal morphology0.500.890.801.0050.801.00 Abnormal nasal bridge morphology0.500.570.401.0050.401.00 Depressed nasal bridge0.830.800.671.0050.400.60 Short nose0.170.000.000.0050.400.40 Abnormal external nose morphology0.500.330.201.0050.201.00 Abnormal morphology of the nasal alae0.500.330.201.0050.201.00 Thick nasal alae0.750.500.331.0050.200.60 Abnormal nasal base norphology0.750.500.331.0050.200.60 Wide nasal base0.750.500.331.0050.200.60 Abnormality of the eye0.500.570.401.0050.401.00 Abnormal eye physiology0.670.670.501.0050.400.80 Abnormality of eye movement0.620.400.251.0050.200.80 Abnormal conjugate eye movement0.750.500.331.0050.200.60 Strabismus0.750.500.331.0050.200.60 Ptosis0.880.670.501.0050.200.40 Abnormal eyebrow morphology0.500.330.201.0050.201.00 Thick eyebrow0.620.400.251.0050.200.80 Abnormal facial shape0.500.330.201.0050.201.00 Coarse facial features0.750.500.331.0050.200.60 43 Table D8: Prediction results on the test set for Floating-Harbor syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A0.571.000.4051.000.40 Abnormality of the mouth0.750.671.000.5050.800.40 Abnormal oral morphology0.750.671.000.5050.800.40 Abnormal oral cavity morphology0.750.671.000.5050.800.40 Abnormal lip morphology0.830.801.000.6750.600.40 Abnormal upper lip morphology0.830.801.000.6750.600.40 Abnormality of the philtrum0.830.801.000.6750.600.40 Short philtrum0.580.670.670.6750.600.60 Abnormality of mouth size0.250.000.000.0050.600.20 Wide mouth0.830.800.671.0050.400.60 Narrow mouth0.380.000.000.0050.200.20 Abnormality of the nose0.120.330.500.2550.800.40 Abnormal nasal morphology0.380.750.750.7550.800.80 Abnormal nasal bridge morphology0.120.330.500.2550.800.40 Narrow nasal bridge0.380.750.750.7550.800.80 Long nose0.250.330.250.5050.400.80 Abnormal external nose morphology0.120.330.500.2550.800.40 Abnormal nasal tip morphology0.120.330.500.2550.800.40 Broad nasal tip0.120.330.500.2550.800.40 Table D9: Prediction results on the test set for Mowat-Wilson syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A0.891.000.8051.000.80 Abnormality of the mouth0.380.750.750.7550.800.80 Abnormal oral morphology0.380.750.750.7550.800.80 Abnormal oral cavity morphology0.380.750.750.7550.800.80 Abnormal lip morphology0.330.570.500.6750.600.80 Abnormal lower lip morphology0.330.570.500.6750.600.80 Thick lower lip vermilion0.330.570.500.6750.600.80 Everted lower lip vermilion0.000.000.000.0050.400.60 Abnormality of mouth size0.620.400.251.0050.200.80 Continued on next page 44 Table D9: Prediction results on the test set for Mowat-Wilson syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Narrow mouth0.620.400.251.0050.200.80 Abnormal forehead morphology0.000.000.000.0050.200.80 High forehead0.250.000.000.0050.200.40 Abnormality of the chin0.380.750.750.7550.800.80 Pointed chin0.380.750.750.7550.800.80 Abnormality of the nose0.670.670.501.0050.400.80 Abnormal nasal morphology0.670.670.501.0050.400.80 Abnormal nasal bridge morphology0.670.670.501.0050.400.80 Wide nasal bridge0.670.670.501.0050.400.80 Abnormal external nose morphology0.620.400.251.0050.200.80 Abnormal nasal tip morphology0.620.400.251.0050.200.80 Broad nasal tip0.620.400.251.0050.200.80 Abnormality of the eye1.001.001.001.0050.800.80 Abnormal eye physiology1.001.001.001.0050.800.80 Abnormality of eye movement1.001.001.001.0050.800.80 Abnormal conjugate eye movement1.001.001.001.0050.800.80 Strabismus0.880.861.000.7550.800.60 Table D10: Prediction results on the test set for Nicolaides- Baraitser syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A0.861.000.7541.000.75 Abnormality of the mouth0.330.670.670.6740.750.75 Abnormal oral morphology0.330.670.670.6740.750.75 Abnormal oral cavity morphology0.330.670.670.6740.750.75 Abnormal lip morphology0.330.670.670.6740.750.75 Abnormal upper lip morphology0.670.500.331.0040.250.75 Abnormality of the philtrum0.670.500.331.0040.250.75 Long philtrum0.670.500.331.0040.250.75 Abnormal lower lip morphology0.330.670.670.6740.750.75 Everted lower lip vermilion0.330.670.670.6740.750.75 Thin vermilion border0.330.670.670.6740.750.75 Continued on next page 45 Table D10: Prediction results on the test set for Nicolaides- Baraitser syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of mouth size0.330.670.670.6740.750.75 Wide mouth0.330.670.670.6740.750.75 Abnormality of the noseN/A0.861.000.7541.000.75 Abnormal nasal morphologyN/A0.861.000.7541.000.75 Abnormal nasal bridge morphology0.250.400.330.5040.500.75 Wide nasal bridge0.000.000.000.0040.250.75 Narrow nasal bridge0.250.400.330.5040.500.75 Abnormal external nose morphology1.001.001.001.0040.750.75 Abnormal morphology of the nasal alae1.001.001.001.0040.750.75 Anteverted nares0.830.801.000.6740.750.50 Abnormality of the eye0.000.000.000.0040.250.75 Abnormal eye physiology0.000.000.000.0040.250.75 Abnormality of eye movement0.000.000.000.0040.250.75 Abnormal conjugate eye movement0.000.000.000.0040.250.75 Strabismus0.000.000.000.0040.250.75 Ptosis0.000.000.000.0040.250.75 Abnormal eyebrow morphology1.001.001.001.0040.750.75 Highly arched eyebrow1.001.001.001.0040.750.75 Abnormal facial shape0.750.800.671.0040.500.75 Triangular face0.500.500.500.5040.500.50 Table D11: Prediction results on the test set for Ohdo syndrome, SBBYS-variant. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0051.001.00 Abnormality of the mouth0.500.750.601.0050.601.00 Abnormal oral morphology0.500.750.601.0050.601.00 Abnormal oral cavity morphology0.500.750.601.0050.601.00 Abnormality of mouth size0.500.750.601.0050.601.00 Wide mouth0.500.570.401.0050.401.00 Narrow mouth0.120.000.000.0050.200.60 Abnormality of the orbital region0.500.570.401.0050.401.00 Continued on next page 46 Table D11: Prediction results on the test set for Ohdo syndrome, SBBYS-variant. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the ocular adnexa0.500.570.401.0050.401.00 Abnormal ocular adnexa morphology0.250.330.250.5050.400.80 Abnormal eyelid morphology0.250.330.250.5050.400.80 Epicanthus0.250.330.250.5050.400.80 Abnormality of the nose0.500.890.801.0050.801.00 Abnormal nasal morphology0.500.890.801.0050.801.00 Abnormal nasal bridge morphology0.500.890.801.0050.801.00 Wide nasal bridge0.500.890.801.0050.801.00 Abnormality of the eyeN/A1.001.001.0051.001.00 Abnormal eye morphologyN/A1.001.001.0051.001.00 Abnormality of globe locationN/A1.001.001.0051.001.00 HypertelorismN/A0.891.000.8051.000.80 Abnormal eye physiology0.500.570.401.0050.401.00 Ptosis0.250.330.250.5050.400.80 Abnormal eyebrow morphology0.500.890.801.0050.801.00 Aplasia/Hypoplasia of the eyebrow0.380.750.750.7550.800.80 Table D12: Prediction results on the test set for Opitz GBBB syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0051.001.00 Abnormality of the mouth0.500.570.401.0050.401.00 Abnormal oral morphology0.500.570.401.0050.401.00 Abnormal oral cavity morphology0.500.570.401.0050.401.00 Abnormal lip morphology0.500.570.401.0050.401.00 Abnormal upper lip morphology0.500.570.401.0050.401.00 Abnormality of the philtrum0.500.570.401.0050.401.00 Long philtrum0.250.330.250.5050.400.80 Abnormality of the orbital region0.500.570.401.0050.401.00 Abnormality of the ocular adnexa0.500.570.401.0050.401.00 Abnormal ocular adnexa morphology0.670.670.501.0050.400.80 Abnormal eyelid morphology0.500.570.401.0050.401.00 Continued on next page 47 Table D12: Prediction results on the test set for Opitz GBBB syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the palpebral fissures0.500.570.401.0050.401.00 Slanting of the palpebral fissure0.500.570.401.0050.401.00 Downslanted palpebral fissures0.500.570.401.0050.401.00 Abnormality of the noseN/A1.001.001.0051.001.00 Abnormal nasal morphologyN/A1.001.001.0051.001.00 Abnormal nasal bridge morphologyN/A1.001.001.0051.001.00 Wide nasal bridgeN/A1.001.001.0051.001.00 Abnormal external nose morphologyN/A1.001.001.0051.001.00 Abnormal morphology of the nasal alaeN/A1.001.001.0051.001.00 Underdeveloped nasal alae0.500.570.401.0050.401.00 Anteverted naresN/A1.001.001.0051.001.00 Abnormality of the eyeN/A1.001.001.0051.001.00 Abnormal eye morphologyN/A1.001.001.0051.001.00 Abnormality of globe locationN/A1.001.001.0051.001.00 HypertelorismN/A1.001.001.0051.001.00 Abnormal eye physiology0.500.750.601.0050.601.00 Abnormality of eye movement0.500.750.601.0050.601.00 Abnormal conjugate eye movement0.500.750.601.0050.601.00 Strabismus0.500.750.601.0050.601.00 Abnormal facial shape0.500.570.401.0050.401.00 Table D13: Prediction results on the test set for Sotos syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0041.001.00 Abnormality of the orbital regionN/A1.001.001.0041.001.00 Abnormality of the ocular adnexaN/A1.001.001.0041.001.00 Abnormal ocular adnexa morphologyN/A0.861.000.7541.000.75 Abnormal eyelid morphologyN/A1.001.001.0041.001.00 Epicanthus0.670.500.331.0040.250.75 Abnormality of the palpebral fissuresN/A1.001.001.0041.001.00 Slanting of the palpebral fissureN/A1.001.001.0041.001.00 Continued on next page 48 Table D13: Prediction results on the test set for Sotos syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Downslanted palpebral fissuresN/A1.001.001.0041.001.00 Abnormality of the eyeN/A1.001.001.0041.001.00 Abnormal eye physiologyN/A1.001.001.0041.001.00 Abnormality of eye movement0.500.860.751.0040.751.00 Abnormal conjugate eye movement0.500.860.751.0040.751.00 Strabismus0.330.670.670.6740.750.75 Ptosis0.330.670.670.6740.750.75 Abnormal eyebrow morphology0.500.670.501.0040.501.00 Thick eyebrow0.000.000.000.0040.250.75 Highly arched eyebrow0.000.000.000.0040.250.75 Aplasia/Hypoplasia of the eyebrow0.500.400.251.0040.251.00 Abnormal facial shape0.500.860.751.0040.751.00 Coarse facial features0.330.670.670.6740.750.75 Table D14: Prediction results on the test set for Seckel syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A0.671.000.5021.000.50 Abnormality of the mouthN/A1.001.001.0021.001.00 Abnormal oral morphologyN/A1.001.001.0021.001.00 Abnormal oral cavity morphologyN/A1.001.001.0021.001.00 Abnormal lip morphologyN/A0.671.000.5021.000.50 Abnormal upper lip morphologyN/A0.671.000.5021.000.50 Abnormality of the philtrumN/A0.671.000.5021.000.50 Long philtrumN/A1.001.001.0021.001.00 Thin vermilion border0.500.670.501.0020.501.00 Thin upper lip vermilion0.000.000.000.0020.500.50 Abnormal facial shapeN/A1.001.001.0021.001.00 Small faceN/A0.671.000.5021.000.50 Narrow faceN/A1.001.001.0021.001.00 49 Table D15: Prediction results on the test set for White-Sutton syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A0.501.000.3331.000.33 Abnormality of the mouthN/A0.501.000.3331.000.33 Abnormal oral morphologyN/A0.501.000.3331.000.33 Abnormal oral cavity morphologyN/A0.501.000.3331.000.33 Abnormal lip morphologyN/A0.501.000.3331.000.33 Abnormal lower lip morphology0.750.671.000.5030.670.33 Thick lower lip vermilion0.750.670.501.0030.330.67 Everted lower lip vermilion0.750.671.000.5030.670.33 Thin vermilion borderN/A0.501.000.3331.000.33 Thin upper lip vermilionN/A0.501.000.3331.000.33 Abnormality of the chin0.000.000.000.0030.670.33 Pointed chin0.000.000.000.0030.670.33 Abnormality of the noseN/A0.501.000.3331.000.33 Abnormal nasal morphologyN/A0.801.000.6731.000.67 Abnormal external nose morphologyN/A0.501.000.3331.000.33 Abnormal morphology of the nasal alae0.750.671.000.5030.670.33 Anteverted nares0.500.00N/A0.0030.670.00 Abnormal nasal tip morphology0.000.000.000.0030.670.33 Broad nasal tip0.000.000.000.0030.670.33 Abnormality of the eye0.750.671.000.5030.670.33 Abnormal eye physiology0.750.671.000.5030.670.33 Abnormality of eye movement0.250.000.000.0030.330.33 Abnormal conjugate eye movement0.250.000.000.0030.330.33 Strabismus0.250.000.000.0030.330.33 Ptosis1.001.001.001.0030.330.33 Abnormal eyebrow morphology0.250.000.000.0030.330.33 Aplasia/Hypoplasia of the eyebrow0.250.000.000.0030.330.33 Abnormal facial shapeN/A0.801.000.6731.000.67 50 Table D16: Prediction results on the test set for Coffin-Siris syn- drome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A0.951.000.90101.000.90 Abnormality of the mouthN/A0.951.000.90101.000.90 Abnormal oral morphologyN/A0.951.000.90101.000.90 Abnormal oral cavity morphologyN/A0.951.000.90101.000.90 Abnormal lip morphologyN/A0.951.000.90101.000.90 Abnormal upper lip morphology1.001.001.001.00100.900.90 Abnormality of the philtrum1.001.001.001.00100.900.90 Broad philtrum1.001.001.001.00100.900.90 Long philtrum0.610.220.121.00100.100.80 Abnormal lower lip morphology0.440.890.890.89100.900.90 Thick lower lip vermilion0.440.890.890.89100.900.90 Thin vermilion border0.500.670.501.00100.501.00 Thin upper lip vermilion0.600.710.561.00100.500.90 Abnormality of mouth size0.620.800.671.00100.600.90 Wide mouth0.500.750.601.00100.601.00 Abnormality of mouth shape0.560.200.111.00100.100.90 Downturned corners of mouth0.670.250.141.00100.100.70 Abnormal forehead morphology0.430.750.670.86100.700.90 Abnormality of the frontal hairline0.500.820.701.00100.701.00 Low anterior hairline0.600.800.750.86100.700.80 Abnormality of the orbital region0.620.800.671.00100.600.90 Abnormality of the ocular adnexa0.500.750.601.00100.601.00 Abnormal ocular adnexa morphology0.620.800.671.00100.600.90 Abnormal eyelid morphology0.620.800.671.00100.600.90 Epicanthus0.060.000.000.00100.100.80 Abnormal eyelash morphology0.580.620.441.00100.400.90 Prominent eyelashes0.670.670.501.00100.400.80 Abnormality of the palpebral fissures0.560.360.221.00100.200.90 Slanting of the palpebral fissure0.500.330.201.00100.201.00 Downslanted palpebral fissures0.620.400.251.00100.200.80 Abnormality of the noseN/A0.951.000.90101.000.90 Abnormal nasal morphologyN/A1.001.001.00101.001.00 Abnormal nasal bridge morphology0.430.750.670.86100.700.90 Depressed nasal bridge0.500.820.701.00100.701.00 Short nose0.380.750.750.75100.800.80 Abnormal external nose morphology0.500.950.901.00100.901.00 Abnormal morphology of the nasal alae0.670.880.781.00100.700.90 Continued on next page 51 Table D16: Prediction results on the test set for Coffin-Siris syn- drome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Anteverted nares0.600.670.570.80100.500.70 Thick nasal alae0.460.500.380.75100.400.80 Abnormal nasal tip morphology0.500.820.701.00100.701.00 Broad nasal tip0.670.880.781.00100.700.90 Abnormal nasal base norphology0.430.750.670.86100.700.90 Wide nasal base0.500.820.701.00100.701.00 Abnormality of the eye0.500.570.401.00100.401.00 Abnormal eye physiology0.500.570.401.00100.401.00 Abnormality of eye movement0.500.460.301.00100.301.00 Abnormal conjugate eye movement0.500.460.301.00100.301.00 Strabismus0.570.500.331.00100.300.90 Ptosis0.440.250.170.50100.200.60 Abnormal eyebrow morphology0.500.750.601.00100.601.00 Thick eyebrow0.750.860.751.00100.600.80 Abnormal facial shape0.620.800.671.00100.600.90 Coarse facial features0.620.800.671.00100.600.90 Table D17: Prediction results on the test set for Cornelia de Lange syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0051.001.00 Abnormality of the mouthN/A1.001.001.0051.001.00 Abnormal oral morphologyN/A1.001.001.0051.001.00 Abnormal oral cavity morphologyN/A1.001.001.0051.001.00 Abnormal lip morphologyN/A1.001.001.0051.001.00 Abnormal upper lip morphology0.500.750.601.0050.601.00 Abnormality of the philtrum0.500.750.601.0050.601.00 Long philtrum0.330.570.500.6750.600.80 Thin vermilion borderN/A1.001.001.0051.001.00 Abnormality of mouth shape0.500.570.401.0050.401.00 Downturned corners of mouth0.500.570.401.0050.401.00 Abnormal forehead morphology0.500.750.601.0050.601.00 Continued on next page 52 Table D17: Prediction results on the test set for Cornelia de Lange syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the frontal hairline0.500.750.601.0050.601.00 Low anterior hairline0.500.750.601.0050.601.00 Abnormality of the nose0.500.750.601.0050.601.00 Abnormal nasal morphology0.500.750.601.0050.601.00 Short nose0.500.750.601.0050.601.00 Abnormal external nose morphology0.500.750.601.0050.601.00 Abnormal morphology of the nasal alae0.500.750.601.0050.601.00 Anteverted nares0.500.750.601.0050.601.00 Abnormal eyebrow morphology0.500.750.601.0050.601.00 Thick eyebrow0.330.570.500.6750.600.80 Synophrys0.250.330.250.5050.400.80 Table D18: Prediction results on the test set for Hyperphosphata- sia with mental retardation syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0051.001.00 Abnormality of the mouth0.500.890.801.0050.801.00 Abnormal oral morphology0.500.890.801.0050.801.00 Abnormal oral cavity morphology0.500.890.801.0050.801.00 Abnormal lip morphology0.500.330.201.0050.201.00 Thin vermilion border0.500.330.201.0050.201.00 Abnormality of mouth shape0.500.750.601.0050.601.00 Downturned corners of mouth0.500.750.601.0050.601.00 Abnormality of the orbital regionN/A1.001.001.0051.001.00 Abnormality of the ocular adnexaN/A1.001.001.0051.001.00 Abnormal ocular adnexa morphologyN/A1.001.001.0051.001.00 Abnormal eyelid morphologyN/A1.001.001.0051.001.00 Epicanthus0.500.750.601.0050.601.00 Abnormality of the palpebral fissures0.500.890.801.0050.801.00 Slanting of the palpebral fissure0.500.570.401.0050.401.00 Upslanted palpebral fissure0.670.670.501.0050.400.80 Continued on next page 53 Table D18: Prediction results on the test set for Hyperphosphata- sia with mental retardation syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormal size of the palpebral fissures0.500.890.801.0050.801.00 Long palpebral fissure0.500.890.801.0050.801.00 Abnormality of the nose0.500.890.801.0050.801.00 Abnormal nasal morphology0.500.890.801.0050.801.00 Abnormal nasal bridge morphology0.500.890.801.0050.801.00 Wide nasal bridge0.500.890.801.0050.801.00 Abnormality of the eye0.500.570.401.0050.401.00 Abnormal eye physiology0.500.570.401.0050.401.00 Ptosis0.500.570.401.0050.401.00 Table D19: Prediction results on the test set for KBG syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A0.801.000.6731.000.67 Abnormality of the mouth0.250.500.500.5030.670.67 Abnormal oral morphology0.250.500.500.5030.670.67 Abnormal oral cavity morphology0.250.500.500.5030.670.67 Abnormal lip morphology0.250.500.500.5030.670.67 Thin vermilion border0.250.500.500.5030.670.67 Thin upper lip vermilion0.250.500.500.5030.670.67 Abnormality of the orbital region1.001.001.001.0030.670.67 Abnormality of the ocular adnexa1.001.001.001.0030.670.67 Abnormal ocular adnexa morphology0.750.671.000.5030.670.33 Abnormal eyelid morphology1.001.001.001.0030.670.67 Abnormality of the palpebral fissures1.001.001.001.0030.670.67 Abnormal size of the palpebral fissures1.001.001.001.0030.670.67 Long palpebral fissure1.001.001.001.0030.670.67 Abnormality of the nose0.750.670.501.0030.330.67 Abnormal nasal morphology0.750.670.501.0030.330.67 Abnormal external nose morphology0.750.670.501.0030.330.67 Abnormal morphology of the nasal alae0.750.670.501.0030.330.67 Continued on next page 54 Table D19: Prediction results on the test set for KBG syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Anteverted nares0.250.000.000.0030.330.33 Abnormal eyebrow morphology0.750.670.501.0030.330.67 Thick eyebrow0.250.000.000.0030.330.33 Synophrys0.750.670.501.0030.330.67 Abnormal facial shape0.250.500.500.5030.670.67 Triangular face0.500.00N/A0.0030.670.00 Table D20: Prediction results on the test set for Noonan syn- drome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A0.861.000.7541.000.75 Abnormality of the mouth0.750.800.671.0040.500.75 Abnormal oral morphology0.750.800.671.0040.500.75 Abnormal oral cavity morphology0.750.800.671.0040.500.75 Abnormal lip morphology0.750.800.671.0040.500.75 Abnormal upper lip morphology0.670.500.331.0040.250.75 Abnormality of the philtrum0.670.500.331.0040.250.75 Short philtrum0.670.500.331.0040.250.75 Abnormal lower lip morphology0.750.800.671.0040.500.75 Thick lower lip vermilion0.750.800.671.0040.500.75 Abnormal forehead morphology1.001.001.001.0040.750.75 Broad forehead0.670.500.331.0040.250.75 High forehead1.001.001.001.0040.750.75 Abnormal midface morphology0.750.800.671.0040.500.75 Midface retrusion0.750.800.671.0040.500.75 Abnormality of the orbital region0.330.670.670.6740.750.75 Abnormality of the ocular adnexa0.330.670.670.6740.750.75 Abnormal ocular adnexa morphology0.330.670.670.6740.750.75 Abnormal eyelid morphology0.330.670.670.6740.750.75 Epicanthus0.670.500.331.0040.250.75 Abnormality of the palpebral fissures0.330.670.670.6740.750.75 Slanting of the palpebral fissure0.330.670.670.6740.750.75 Continued on next page 55 Table D20: Prediction results on the test set for Noonan syn- drome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Downslanted palpebral fissures0.330.670.670.6740.750.75 Abnormality of the nose0.670.500.331.0040.250.75 Abnormal nasal morphology0.500.400.251.0040.251.00 Abnormal external nose morphology0.670.500.331.0040.250.75 Abnormal morphology of the nasal alae0.500.400.251.0040.251.00 Anteverted nares0.670.500.331.0040.250.75 Abnormality of the eyeN/A0.861.000.7541.000.75 Abnormal eye morphologyN/A0.861.000.7541.000.75 Abnormality of globe locationN/A0.861.000.7541.000.75 HypertelorismN/A0.861.000.7541.000.75 Proptosis0.750.800.671.0040.500.75 Abnormal eye physiologyN/A0.861.000.7541.000.75 Abnormality of eye movement0.330.670.670.6740.750.75 Abnormal conjugate eye movement0.330.670.670.6740.750.75 Strabismus0.330.670.670.6740.750.75 PtosisN/A0.861.000.7541.000.75 Abnormal eyebrow morphology0.000.000.000.0040.250.75 Highly arched eyebrow0.000.000.000.0040.250.75 Abnormal facial shape0.500.670.501.0040.501.00 Triangular face0.750.800.671.0040.500.75 Table D21: Prediction results on the test set for Ogden syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0041.001.00 Abnormality of the mouth0.500.860.751.0040.751.00 Abnormal oral morphology0.500.860.751.0040.751.00 Abnormal oral cavity morphology0.500.860.751.0040.751.00 Abnormal lip morphology0.500.860.751.0040.751.00 Abnormal upper lip morphology0.500.860.751.0040.751.00 Abnormality of the philtrum0.500.400.251.0040.251.00 Long philtrum0.500.400.251.0040.251.00 Continued on next page 56 Table D21: Prediction results on the test set for Ogden syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of upper lip vermillion0.500.670.501.0040.501.00 Everted upper lip vermilion0.500.670.501.0040.501.00 Thin vermilion border0.500.400.251.0040.251.00 Thin upper lip vermilion0.500.400.251.0040.251.00 Abnormal forehead morphologyN/A1.001.001.0041.001.00 Narrow forehead0.500.670.501.0040.501.00 High forehead0.500.400.251.0040.251.00 Abnormality of the orbital region0.500.670.501.0040.501.00 Abnormality of the ocular adnexa0.500.670.501.0040.501.00 Abnormal ocular adnexa morphology0.500.670.501.0040.501.00 Abnormal eyelid morphology0.500.670.501.0040.501.00 Abnormality of the palpebral fissures0.500.670.501.0040.501.00 Slanting of the palpebral fissure0.500.670.501.0040.501.00 Downslanted palpebral fissures0.500.670.501.0040.501.00 Abnormality of the nose0.500.860.751.0040.751.00 Abnormal nasal morphology0.500.860.751.0040.751.00 Abnormal external nose morphology0.500.860.751.0040.751.00 Abnormal morphology of the nasal alae0.500.860.751.0040.751.00 Underdeveloped nasal alae0.500.860.751.0040.751.00 Anteverted nares0.670.500.331.0040.250.75 Abnormal columella morphology0.500.860.751.0040.751.00 Short columella0.500.860.751.0040.751.00 Abnormality of the eye0.500.670.501.0040.501.00 Abnormal eye morphology0.500.400.251.0040.251.00 Abnormality of globe location0.500.400.251.0040.251.00 Proptosis0.500.400.251.0040.251.00 Abnormal eye physiology0.500.670.501.0040.501.00 Abnormality of eye movement0.500.400.251.0040.251.00 Abnormal conjugate eye movement0.500.400.251.0040.251.00 Strabismus0.500.400.251.0040.251.00 Ptosis0.670.500.331.0040.250.75 Abnormal eyebrow morphology0.500.400.251.0040.251.00 Aplasia/Hypoplasia of the eyebrow0.500.400.251.0040.251.00 Abnormal facial shape0.500.860.751.0040.751.00 Coarse facial features0.500.860.751.0040.751.00 57 Table D22: Prediction results on the test set for Williams-Beuren syndrome. The HPO terms in bold letters are leafs of the human phenotype ontology. F1: F1-Score, Prec: Precision, Reca: Recall, N: Samples, Prev: Prevalence, Det. P: Detection Prevalence. AUROCF1PrecRecaNPrevDet. P Abnormality of the faceN/A1.001.001.0041.001.00 Abnormality of the mouthN/A1.001.001.0041.001.00 Abnormal oral morphologyN/A1.001.001.0041.001.00 Abnormal oral cavity morphologyN/A1.001.001.0041.001.00 Abnormal lip morphologyN/A1.001.001.0041.001.00 Abnormal upper lip morphology0.500.860.751.0040.751.00 Abnormality of the philtrum0.500.860.751.0040.751.00 Long philtrum0.500.860.751.0040.751.00 Abnormal lower lip morphologyN/A1.001.001.0041.001.00 Thick lower lip vermilionN/A1.001.001.0041.001.00 Everted lower lip vermilionN/A1.001.001.0041.001.00 Abnormality of the orbital region0.500.670.501.0040.501.00 Abnormality of the ocular adnexa0.500.670.501.0040.501.00 Abnormal ocular adnexa morphology0.500.670.501.0040.501.00 Abnormal eyelid morphology0.500.670.501.0040.501.00 Epicanthus0.500.670.501.0040.501.00 Abnormality of the nose0.500.860.751.0040.751.00 Abnormal nasal morphology0.500.860.751.0040.751.00 Abnormal nasal bridge morphology0.500.400.251.0040.251.00 Depressed nasal bridge0.500.400.251.0040.251.00 Short nose0.500.860.751.0040.751.00 Abnormal external nose morphology0.500.400.251.0040.251.00 Abnormal morphology of the nasal alae0.500.400.251.0040.251.00 Anteverted nares0.500.400.251.0040.251.00 Abnormality of the eyeN/A1.001.001.0041.001.00 Abnormal eye physiologyN/A1.001.001.0041.001.00 Abnormality of eye movementN/A1.001.001.0041.001.00 Abnormal conjugate eye movementN/A1.001.001.0041.001.00 StrabismusN/A1.001.001.0041.001.00 Ptosis0.500.400.251.0040.251.00 Abnormal facial shapeN/A1.001.001.0041.001.00 Small face0.500.860.751.0040.751.00 Narrow face0.500.860.751.0040.751.00 Coarse facial features0.500.670.501.0040.501.00 Elfin facies0.500.860.751.0040.751.00 References [1] Bauskis, A., Strange, C., Molster, C., Fisher, C.: The diagnostic odyssey: insights 58 from parents of children living with an undiagnosed condition. Orphanet Journal of Rare Diseases 17(1), 233 (2022) https://doi.org/10.1186/s13023-022-02358-x . Accessed 2026-06-07 [2] Michaels-Igbokwe, C., McInnes, B., MacDonald, K.V., Currie, G.R., Omar, F., Shewchuk, B., Bernier, F.P., Marshall, D.A.: (Un)standardized testing: the diag- nostic odyssey of children with rare genetic disorders in Alberta, Canada. Genetics in Medicine 23(2), 272–279 (2021) https://doi.org/10.1038/s41436-020-00975-0 . Accessed 2026-06-07 [3] Hsieh, T., Krawitz, P.M.: Computational facial analysis for rare Mendelian disor- ders. American Journal of Medical Genetics Part C: Seminars in Medical Genetics 193(3), 32061 (2023) https://doi.org/10.1002/ajmg.c.32061 [4] Hsieh, T.-C., Bar-Haim, A., Moosa, S., Ehmke, N., Gripp, K.W., Pantel, J.T., Danyel, M., Mensah, M.A., Horn, D., Rosnev, S., Fleischer, N., Bonini, G., Hustinx, A., Schmid, A., Knaus, A., Javanmardi, B., Klinkhammer, H., Les- mann, H., Sivalingam, S., Kamphans, T., Meiswinkel, W., Ebstein, F., Kr ̈uger, E., K ̈ury, S., B ́ezieau, S., Schmidt, A., Peters, S., Engels, H., Mangold, E., Kreiß, M., Cremer, K., Perne, C., Betz, R.C., Bender, T., Grundmann-Hauser, K., Haack, T.B., Wagner, M., Brunet, T., Bentzen, H.B., Averdunk, L., Coet- zer, K.C., Lyon, G.J., Spielmann, M., Schaaf, C.P., Mundlos, S., N ̈othen, M.M., Krawitz, P.M.: GestaltMatcher facilitates rare disease matching using facial phe- notype descriptors. Nature Genetics 54(3), 349–357 (2022) https://doi.org/10. 1038/s41588-021-01010-x [5] Hustinx, A., Hellmann, F., S ̈umer, O., Javanmardi, B., Andr ́e, E., Krawitz, P., Hsieh, T.-C.: Improving deep facial phenotyping for ultra-rare disorder verifica- tion using model ensembles. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), p. 5018–5028 (2023) [6] Van Der Donk, R., Jansen, S., Schuurs-Hoeijmakers, J.H.M., Koolen, D.A., Goltstein, L.C.M.J., Hoischen, A., Brunner, H.G., Kemmeren, P., Nell ̊aker, C., Vissers, L.E.L.M., De Vries, B.B.A., Hehir-Kwa, J.Y.: Next-generation pheno- typing using computer vision algorithms in rare genomic neurodevelopmental disorders. Genetics in Medicine 21(8), 1719–1725 (2019) https://doi.org/10.1038/ s41436-018-0404-y . Accessed 2026-06-07 [7] Dingemans, A.J.M., Hinne, M., Truijen, K.M.G., Goltstein, L., Reeuwijk, J., Leeuw, N., Schuurs-Hoeijmakers, J., Pfundt, R., Diets, I.J., Hoed, J., Boer, E., Spek, J., Jansen, S., Bon, B.W., Jonis, N., Ockeloen, C.W., Vulto-van Silfhout, A.T., Kleefstra, T., Koolen, D.A., Campeau, P.M., Palmer, E.E., Van Esch, H., Lyon, G.J., Alkuraya, F.S., Rauch, A., Marom, R., Baralle, D., Sluijs, P.J., Santen, G.W.E., Kooy, R.F., Gerven, M.A.J., Vissers, L.E.L.M., Vries, B.B.A.: PhenoScore quantifies phenotypic variation for rare genetic diseases by combining facial analysis with other clinical features using a machine-learning 59 framework. Nature Genetics 55(9), 1598–1607 (2023) https://doi.org/10.1038/ s41588-023-01469-w [8] Duong, D., Johny, A.R., Ledgister Hanchard, S., Fortney, C., Flaharty, K., Hell- mann, F., Hu, P., Javanmardi, B., Moosa, S., Patel, T., Persky, S., S ̈umer, O., Tekendo-Ngongang, C., Lesmann, H., Hsieh, T.-C., Waikel, R.L., Andr ́e, E., Krawitz, P., Solomon, B.D.: Comparison of clinical geneticist and computer visual attention in assessing genetic conditions. PLOS Genetics 20(2), 1011168 (2024) https://doi.org/10.1371/journal.pgen.1011168 [9] Robinson, P.N., K ̈ohler, S., Bauer, S., Seelow, D., Horn, D., Mundlos, S.: The human phenotype ontology: A tool for annotating and analyzing human hered- itary disease. The American Journal of Human Genetics 83(5), 610–615 (2008) https://doi.org/10.1016/j.ajhg.2008.09.017 [10] Qiao, H., Tan, J., Yan, J., Sun, C., Yin, X., Li, Z., Wu, J., Guan, H., Wen, S., Zhang, M., Xu, S., Jin, L.: A comprehensive evaluation of the phenotype-first and data-driven approaches in analyzing facial morphological traits. iScience 27(3), 109325 (2024) https://doi.org/10.1016/j.isci.2024.109325 [11] Lesmann, H., Hustinx, A., Moosa, S., Klinkhammer, H., Marchi, E., Caro, P., Abdelrazek, I.M., Pantel, J.T., Hagen, M.t., Thong, M.-K., Binti Mazlan, R.A., Tae, S.K., Kamphans, T., Meiswinkel, W., Li, J.-M., Javanmardi, B., Knaus, A., Uwineza, A., Knopp, C., Tkemaladze, T., Elbracht, M., Mattern, L., Jamra, R.A., Velmans, C., Strehlow, V., Jacob, M., Peron, A., Dias, C., Nunes, B.C., Vilella, T., Pinheiro, I.F., Kim, C.A., Melaragno, M.I., Weiland, H., Kaptain, S., Chwia lkowska, K., Kwasniewski, M., Saad, R., Wiethoff, S., Goel, H., Tang, C., Hau, A., Barakat, T.S., Panek, P., Nabil, A., Suh, J., Braun, F., Gomy, I., Averdunk, L., Ekure, E., Bergant, G., Peterlin, B., Graziano, C., Gaboon, N., Fiesco-Roa, M., Spinelli, A.M., Wilpert, N.-M., Phowthongkum, P., G ̈uzel, N., Haack, T.B., Bitar, R., Tzschach, A., Rodriguez-Palmero, A., Brunet, T., Rudnik- Sch ̈oneborn, S., Contreras-Capetillo, S.N., Oberlack, A., Samango-Sprouse, C., Sadeghin, T., Olaya, M., Platzer, K., Borovikov, A., Schnabel, F., Heuft, L., Herrmann, V., Oegema, R., Elkhateeb, N., Kumar, S., Komlosi, K., Mohamed, K., Kalantari, S., Sirchia, F., Martinez-Monseny, A.F., H ̈oller, M., Toutouna, L., Mohamed, A., Lasa-Aranzasti, A., Sayer, J.A., Ehmke, N., Danyel, M., Sczakiel, H., Schwartzmann, S., Boschann, F., Zhao, M., Adam, R., Einicke, L., Horn, D., Chew, K.S., KAM, C.C., Karakoyun, M., Pode-Shakked, B., Eliyahu, A., Rock, R., Carrion, T., Chorin, O., Zarate, Y.A., Conti, M.M., Karakaya, M., Tung, M.L., Chandra, B., Bouman, A., Lumaka, A., Wasif, N., Shinawi, M., Blackburn, P.R., Wang, T., Niehues, T., Schmidt, A., Roth, R.R., Wieczorek, D., Hu, P., Waikel, R.L., Ledgister Hanchard, S.E., Elmakkawy, G., Safwat, S., Ebstein, F., Kr ̈uger, E., K ̈ury, S., B ́ezieau, S., Arlt, A., Olinger, E., Marbach, F., Li, D., Dupuis, L., Mendoza-Londono, R., Houge, S.D., Weis, D., Chung, B.H.-Y., Mak, C.C.Y., Kayserili, H., Elcioglu, N., Aykut, A., S ̧im ̧sek-Kiper, P. ̈ O., 60 B ̈ogershausen, N., Wollnik, B., Bentzen, H.B., Kurth, I., Netzer, C., Jezela- Stanek, A., Devriendt, K., Gripp, K.W., M ̈ucke, M., Verloes, A., Schaaf, C.P., Nell ̊a ker, C., Solomon, B.D., N ̈othen, M.M., Abdalla, E., Lyon, G.J., Krawitz, P.M., Hsieh, T.-C.: GestaltMatcher Database - A global reference for facial phe- notypic variability in rare human diseases. Research Square, 3–4438861 (2024) https://doi.org/10.21203/rs.3.rs-4438861/v1 [12] Islam, T.U., Shaikh, T.A.: A lightweight attention-based deep learning facial recognition system for multiple genetic syndromes. International Journal of Data Science and Analytics 20(3), 2871–2889 (2025) https://doi.org/10.1007/ s41060-024-00658-y [13] Hallgr ́ımsson, B., Aponte, J.D., Katz, D.C., Bannister, J.J., Riccardi, S.L., Mahasuwan, N., McInnes, B.L., Ferrara, T.M., Lipman, D.M., Neves, A.B., Spitz- macher, J.A.J., Larson, J.R., Bellus, G.A., Pham, A.M., Aboujaoude, E., Benke, T.A., Chatfield, K.C., Davis, S.M., Elias, E.R., Enzenauer, R.W., French, B.M., Pickler, L.L., Shieh, J.T.C., Slavotinek, A., Harrop, A.R., Innes, A.M., McCand- less, S.E., McCourt, E.A., Meeks, N.J.L., Tartaglia, N.R., Tsai, A.C.-H., Wyse, J.P.H., Bernstein, J.A., Sanchez-Lara, P.A., Forkert, N.D., Bernier, F.P., Spritz, R.A., Klein, O.D.: Automated syndrome diagnosis by three-dimensional facial imaging. Genetics in Medicine 22(10), 1682–1693 (2020) https://doi.org/10.1038/ s41436-020-0845-y [14] Bannister, J.J., Wilms, M., Aponte, J.D., Katz, D.C., Klein, O.D., Bernier, F.P.J., Spritz, R.A., Hallgr ́ımsson, B., Forkert, N.D.: A Deep Invertible 3-D Facial Shape Model for Interpretable Genetic Syndrome Diagnosis. IEEE Journal of Biomedical and Health Informatics 26(7), 3229–3239 (2022) https://doi.org/10.1109/JBHI. 2022.3164848 [15] Mahdi, S.S., Matthews, H., Nauwelaers, N., Vanneste, M., Gong, S., Bouritsas, G., Baynam, G.S., Hammond, P., Spritz, R., Klein, O.D., Hallgr ́ımsson, B., Peeters, H., Bronstein, M., Claes, P.: Multi-Scale Part-Based Syndrome Classification of 3D Facial Images. IEEE Access 10, 23450–23462 (2022) https://doi.org/10.1109/ ACCESS.2022.3153357 [16] Dingemans, A.J.M., Vries, B.B.A.d., Vissers, L.E.L.M., Gerven, M.A.J.v., Hinne, M.: Comparing facial feature extraction methods in the diagnosis of rare genetic syndromes. medRxiv. Pages: 2022.08.26.22279217 (2022). https://doi.org/10. 1101/2022.08.26.22279217 [17] Robinson, P.N., Ravanmehr, V., Jacobsen, J.O.B., Danis, D., Zhang, X.A., Car- mody, L.C., Gargano, M.A., Thaxton, C.L., Karlebach, G., Reese, J., Holtgrewe, M., K ̈ohler, S., McMurry, J.A., Haendel, M.A., Smedley, D.: Interpretable Clin- ical Genomics with a Likelihood Ratio Paradigm. American Journal of Human Genetics 107(3), 403–417 (2020) https://doi.org/10.1016/j.ajhg.2020.06.021 61 [18] Ullah, M.Z., Aono, M., Seddiqui, M.H.: Estimating a ranked list of human heredi- tary diseases for clinical phenotypes by using weighted bipartite network. In: 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), p. 3475–3478 (2013). https://doi.org/10.1109/EMBC. 2013.6610290 . ISSN: 1558-4615 [19] Hamosh, A.: Online mendelian inheritance in man (omim), a knowledgebase of human genes and genetic disorders. Nucleic Acids Research 33(Database issue), 514–517 (2004) https://doi.org/10.1093/nar/gki033 [20] Nguengang Wakap, S., Lambert, D.M., Olry, A., Rodwell, C., Gueydan, C., Lanneau, V., Murphy, D., Le Cam, Y., Rath, A.: Estimating cumulative point prevalence of rare diseases: analysis of the orphanet database. Euro- pean Journal of Human Genetics 28(2), 165–173 (2019) https://doi.org/10.1038/ s41431-019-0508-0 [21] INSERM: Orphadata: Free access data from Orphanet. https://w.orphadata. com/. INSERM, Orphanet. Available at: https://w.orphadata.com/. Data accessed: 22 April 2026 (1999) [22] Fleiss, J.L.: Measuring nominal scale agreement among many raters. Psychologi- cal Bulletin 76(5), 378–382 (1971) https://doi.org/10.1037/h0031619 . Accessed 2026-06-30 [23] Kirchhoff, A., Hustinx, A., Javanmardi, B., Hsieh, T.-C., Brand, F., Hellmann, F., Mertes, S., Andr ́e, E., Moosa, S., Schultz, T., Solomon, B.D., Krawitz, P.: Gestaltgan: synthetic photorealistic portraits of individuals with rare genetic disorders. European Journal of Human Genetics 33(3), 377–382 (2025) https: //doi.org/10.1038/s41431-025-01787-z [24] Kartynnik, Y., Ablavatski, A., Grishchenko, I., Grundmann, M.: Real-time Facial Surface Geometry from Monocular Video on Mobile GPUs (2019). https://arxiv. org/abs/1907.06724v1 Accessed 2026-04-08 [25] Chicco, D., Jurman, G.: The matthews correlation coefficient (mcc) should replace the roc auc as the standard metric for assessing binary classification. BioData Mining 16(1) (2023) https://doi.org/10.1186/s13040-023-00322-4 [26] Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. arXiv. Version Number: 2 (2016). https: //doi.org/10.48550/ARXIV.1612.00593 [27] Zhang, Z., Song, Y., Qi, H.: Age progression/regression by conditional adversarial autoencoder. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017). IEEE 62 [28] Sundararajan, M., Taly, A., Yan, Q.: Axiomatic Attribution for Deep Net- works. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th Interna- tional Conference on Machine Learning, p. 3319–3328. PMLR, ??? (2017). https://proceedings.mlr.press/v70/sundararajan17a.html Accessed 2026-06-07 [29] Nguyen, Q., Valizadegan, H., Hauskrecht, M.: Learning classification models with soft-label information. Journal of the American Medical Informatics Association 21(3), 501–508 (2014) https://doi.org/10.1136/amiajnl-2013-001964 [30] Matthews, B.W.: Comparison of the predicted and observed secondary struc- ture of T4 phage lysozyme. Biochimica et Biophysica Acta (BBA) - Protein Structure 405(2), 442–451 (1975) https://doi.org/10.1016/0005-2795(75)90109-9 . Accessed 2026-07-04 [31] Kull, M., Filho, T.S., Flach, P.: Beta calibration: a well-founded and easily imple- mented improvement on logistic calibration for binary classifiers. In: Singh, A., Zhu, J. (eds.) Proceedings of the 20th International Conference on Artificial Intel- ligence and Statistics. Proceedings of Machine Learning Research, vol. 54, p. 623–631. PMLR, ??? (2017). https://proceedings.mlr.press/v54/kull17a.html [32] Balanya, S.A., Maro ̃nas, J., Ramos, D.: Adaptive temperature scaling for Robust calibration of deep neural networks. Neural Computing and Applications 36(14), 8073–8095 (2024) https://doi.org/10.1007/s00521-024-09505-4 . Accessed 2026- 07-03 [33] Brier, G.W.: VERIFICATION OF FORECASTS EXPRESSED IN TERMS OF PROBABILITY. Monthly Weather Review 78(1), 1–3 (1950) https://doi.org/10. 1175/1520-0493(1950)078⟨0001:VOFEIT⟩2.0.CO;2 . Accessed 2026-07-04 [34] Hellmann, F., Mertes, S., Benouis, M., Hustinx, A., Hsieh, T.-C., Conati, C., Krawitz, P., Andr ́e, E.: GANonymization: A GAN-Based Face Anonymization Framework for Preserving Emotional Expressions. ACM Transactions on Multi- media Computing, Communications, and Applications 21(1), 1–27 (2025) https: //doi.org/10.1145/3641107 63

Instant research discovery