Paper deep dive
Leveraging Large Language Models and Survival Analysis for Early Prediction of Chemotherapy Outcomes
Muhammad Faisal Shahid, Asad Afzal, Abdullah Faiz, Muhammad Siddiqui, Arbaz Khan Shehzad, Fatima Aftab, Muhammad Usamah Shahid, Muddassar Farooq
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 3/22/2026, 6:25:26 AM
Summary
This study presents a framework for the early prediction of chemotherapy outcomes by leveraging Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) to extract clinical phenotypes and treatment labels from unstructured electronic medical records. The extracted data, including vitals, biomarkers, and staging, are used in a Random Survival Forest (RSF) model to predict time-to-failure and treatment outcomes. The approach achieved a C-index of 0.731 for breast cancer and demonstrated consistent performance across four other cancer types, providing a scalable method for personalized oncology.
Entities (5)
Relation Signals (3)
Llama-3 → extracts → Cancer Phenotypes
confidence 95% · The two well-known LLMs utilized for information extraction are LLaMA-3 8B and Mistral v0.2.
Random Survival Forest → predicts → Chemotherapy Outcome
confidence 95% · Random Survival Forest was used to predict time-to-failure... to predict treatment outcomes
RAG → enhances → LLM
confidence 90% · we propose a smart and autonomous framework for the extraction and annotation of cancer phenotypes utilizing a Retrieval-Augmented Generation (RAG) model
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Chemotherapy for cancer treatment is costly and accompanied by severe side effects, highlighting the critical need for early prediction of treatment outcomes to improve patient management and informed decision-making. Predictive models for chemotherapy outcomes using real-world data face challenges, including the absence of explicit phenotypes and treatment outcome labels such as cancer progression and toxicity. This study addresses these challenges by employing Large Language Models (LLMs) and ontology-based techniques for phenotypes and outcome label extraction from patient notes. We focused on one of the most frequently occurring cancers, breast cancer, due to its high prevalence and significant variability in patient response to treatment, making it a critical area for improving predictive modeling. The dataset included features such as vitals, demographics, staging, biomarkers, and performance scales. Drug regimens and their combinations were extracted from the chemotherapy plans in the EMR data and shortlisted based on NCCN guidelines, verified with NIH standards, and analyzed through survival modeling. The proposed approach significantly reduced phenotypes sparsity and improved predictive accuracy. Random Survival Forest was used to predict time-to-failure, achieving a C-index of 73%, and utilized as a classifier at a specific time point to predict treatment outcomes, with accuracy and F1 scores above 70%. The outcome probabilities were validated for reliability by calibration curves. We extended our approach to four other cancer types. This research highlights the potential of early prediction of treatment outcomes using LLM-based clinical data extraction enabling personalized treatment plans with better patient outcomes.
Tags
Links
- Source: https://arxiv.org/abs/2603.11594v1
- Canonical: https://arxiv.org/abs/2603.11594v1
Trouble viewing inline? Open PDF directly →
Full Text
25,334 characters extracted from source content.
Expand or collapse full text
Leveraging Large Language Models and Survival Analysis for Early Prediction of Chemotherapy Outcomes Muhammad Faisal Shahid 1 , Asad Afzal 1 , Abdullah Faiz 1 , Muhammad Siddiqui 1 , Arbaz Khan Shehzad 1 , Fatima Aftab 1 , Muhammad Usamah Shahid 1 , and Muddassar Farooq 1 CureMD Research, 80 Pine St 21st Floor, New York, NY 10005, United States http://w.curemd.com faisal.shahid, asad.afzal, abdullah.faiz, muhammad.siddiqui, arbaz.khan, fatima.aftab, muhammad.usamah, muddassar.farooq@curemd.com Abstract. Chemotherapy for cancer treatment is costly and accompa- nied by severe side effects, highlighting the critical need for early pre- diction of treatment outcomes to improve patient management and in- formed decision-making. Predictive models for chemotherapy outcomes using real-world data face challenges, including the absence of explicit phenotypes and treatment outcome labels such as cancer progression and toxicity. This study addresses these challenges by employing Large Lan- guage Models (LLMs) and ontology-based techniques for phenotypes and outcome label extraction from patient notes. We focused on one of the most frequently occurring cancers, breast cancer, due to its high preva- lence and significant variability in patient response to treatment, making it a critical area for improving predictive modeling. The dataset included features such as vitals, demographics, staging, biomarkers, and perfor- mance scales. Drug regimens and their combinations were extracted from the chemotherapy plans in the EMR data and shortlisted based on NCCN guidelines, verified with NIH standards, and analyzed through survival modeling. The proposed approach significantly reduced phenotypes spar- sity and improved predictive accuracy. Random Survival Forest was used to predict time-to-failure, achieving a C-index of 73%, and utilized as a classifier at a specific time point to predict treatment outcomes, with accuracy and F1 scores above 70%. The outcome probabilities were vali- dated for reliability by calibration curves. We extended our approach to four other cancer types. This research highlights the potential of early prediction of treatment outcomes using LLM-based clinical data extrac- tion enabling personalized treatment plans with better patient outcomes. Keywords: Survival Analysis· Information Extraction· Large Lan- guage Models· Chemotherapy Outcome Prediction. 1 Introduction Chemotherapy is widely regarded as a core treatment for breast cancer, yet it often poses significant physical, emotional, and financial burdens on patients. arXiv:2603.11594v1 [cs.AI] 12 Mar 2026 2F. Shahid et al. The uncertainty of chemotherapy success compounds this stress, underscoring the need for methods that can predict outcomes earlier in the treatment pro- cess. Early prediction of potential treatment outcomes can not only guide clinical decisions but can also reduce unnecessary expenses and alleviate patient anxi- ety. In this vein, recent research has explored the integration of biomarkers [3], imaging-based insight [13], and molecular characteristics of tumors [12] to refine chemotherapy prediction and personalize therapy strategies. While these studies mark substantial progress, they frequently highlight the complexity of capturing the full clinical context. Some focus on a limited set of biomarkers or imaging parameters, whereas others rely on machine learning (ML) methods but omit critical patient details such as comorbidities, staging, or performance scales [5]. Emerging approaches that employ Large Language Models (LLMs) to structure information from clinical notes indicate a promising direction, particularly to handle heterogeneous and unstructured real-world data and entity extraction more systematically [15]. Driven by these insights, our work aims to unify broader clinical features with outcome modeling, ultimately providing clinicians with an early signal of whether a given chemotherapy plan might fail. We draw on findings from earlier machine learning studies on survival analysis [8], focusing on robust data extraction and meaningful feature engineering. By prioritizing both model interpretability and reliable data integration, our goal is to create a framework that can help improve breast cancer treatment and serve as a blueprint for similar predictive tasks in other cancer types. 2 Cancer Phenotype Extraction from Oncological Notes 2.1 Introduction AI-centered electronic health applications utilize big data collected from EHR systems. Coupled with AI/ML models, they are at the core of Real-World Ev- idence (RWE) paradigm. However, much of the valuable information required for building such applications is stored inside clinical notes, owing to the profes- sion’s legacy. For example, in our partner oncology EMR, 97% of the oncologists record phenotypes in clinical notes. Furthermore, annotation of the phenotypes by expert oncologists would result in infeasible time and costs. Valuable RWE cannot be built around the true outcomes of effective cancer treatments. Building on the aforementioned issues, we propose a smart and autonomous framework for the extraction and annotation of cancer phenotypes utilizing a Retrieval-Augmented Generation (RAG) model and a Large Language Model (LLM) respectively. It runs using our on-premise, secure Nvidia A100 GPU clus- ters. We also perform a deep comparison with an earlier knowledge-driven system using the NCIt Ontology Annotator, while improving on the already existing on- tology system. Early Prediction of Chemotherapy Outcomes3 Data Lakehouse Oncological Notes Note SegmentationRedundancy Removal Notes Preprocessing Preprocessed Notes Chunked Notes Query Semantic Ranking Query Embedding Chunk Embeddings Lexical Ranking Cosine Similarity Re-ranked Chunks Top K Chunks Processed QueryProcessed ChunksBM25 Scores Re-ranked Chunks Top K Chunks Prompt Combined Chunks Semantic Chunks Lexical Chunks Language Model Extracted Phenotypes Fig. 1. The LLM Annotation System’s workflow. Both the lexical and semantic chunks are used to maximize information retrieval. 2.2 Research Literature Rule-based extractions are often reliable with a database housing relevant big data. The National Cancer Institute thesaurus (NCIt) [6], for instance, provides codes for over 171,000 classes and 500,000 relationships. Similarly, the Cancer Care Treatment Outcome Ontology (CCTOO) [10] consists of a total of 1,133 classes. Towards deep learning, BioBERT [9] and SciBERT [4], while boasting great F1 scores, extract limited entities. LLMs have been shown to be much more reliable for entity extraction. Huang et al. [7] used GPT-3.5 Turbo to extract lung and bone cancer information from pathological reports, showing the possibility of re-engineering an LLM for extracting information from provider notes with minimal human supervision. 2.3 Methodology The LLM Annotation System shown in Figure 1 contains three main steps: (1) preprocessing notes to segment and remove redundancies; (2) computing cosine similarity and BM25 Scores to extract topk note chunks both lexically and semantically; and (3) preparing a K-shot prompt to feed to the LLM for infor- mation extraction. The two well-known LLMs utilized for information extraction are LLaMA-3 8B and Mistral v0.2. The phenotypes important for breast cancer identification include TNM stag- ing, stage group, tumor size, cancer grade, performance metrics, and biomark- ers such as estrogen receptor (ER) and progesterone receptor (PR). To extract 4F. Shahid et al. the information properly while minimizing the post-processing steps, we use Python’s Jsonschema module to create a simple and effective schema that the LLM can adhere to. As LLMs are constrained by token limits, we apply a RAG system using the mxbai embeddings model that splits a note into chunks if the entire note exceeds 2,500 tokens. This ensures that the LLM has enough information to work with and maximize in-context retrieval ability. We find the topk note chunks having the required information to be best atk = 10. For the ontology system, primary improvements in the system included up- dated regex for better biomarkers extraction, referencing for metastatic cancer, and processing improvements from the NCIt API. 2.4 Results and Discussion A random sample of 150 admission and progress notes from different oncology practices was obtained to validate the results from both systems. Evaluating on five labels from each note, we evaluate on 750 labels in total, hand-annotated by a panel of five physicians headed by a senior resident oncologist in a partner university teaching hospital. Table 1 shows the metrics between the ontology system and the different LLMs used in our new system. The LLM Annotation system is rarely prone to (1) missing phenotypes due to vague semantics; and (2) hallucinating phenotypes not present in the clinical notes. The Ontology system’s point of contention is the NCIt API which has been observed to fail annotation on some clinical notes, further driving the need for an improved phenotype annotation system. ModelAccuracyPrecisionRecallF1-Score LLaMA 3 8B86.13%87.90%94.95%91.29% Mistral 7B v0.279.20%81.35%91.98%86.34% Ontology85.04%100.00%83.80%91.19% Table 1. Classification metrics across both systems and the different LLMs. The On- tology system does not hallucinate and therefore achieves a perfect score in Precision. The system has since been updated to use better LLMs for phenotype ex- traction, including LLaMA 3.1 8B and Qwen 2.5 32B, combined with a critic agent that minimizes information that has been skipped or hallucinated by the LLM. 3 Cancer Labels Extraction from Oncological Notes 3.1 Introduction Chemotherapy treatment outcomes contain valuable information about the treat- ment itself and how it affects the cancer and the patient. They are vital to ex- tract from clinical notes to effectively determine critical patient health and use Early Prediction of Chemotherapy Outcomes5 them in decision-making processes. Multiple labels are extracted for treatment outcomes, categorized into three distinct branches: (1) progression; (2) toxicity; and (3) death/hospice. As with our LLM Annotation system, we once again use the power of LLMs to extract and interpret data to enhance the efficacy of cancer patient care pathways. 3.2 Research Literature Wang et al. [14] introduced an entity extraction pipeline for medical text records using LLMs. Despite the remarkable results, the authors note that the hallucina- tion issues of LLMs require great attention. Monajatipoor et al. [11] also noted remarkable results in few-shot Named Entity Recognition (NER) for biomedical knowledge extraction using LLMs. The hallucination issues gave rise to thehe need of verifying results generated by the LLM, which is where our LLM Annotation system is refined not just for speedups and improvements in prompt engineering, but also an additional workflow after LLM generation that ensures that the generated response is valid and is contained in the note. 3.3 Methodology We first set up a JSON schema to capture critical information from the notes. For each branch of the treatment outcomes, here is what’s extracted as labels: (1) progression includes whether the cancer has progressed and details pertain- ing to the progression; (2) toxicity covers adverse effects, deterioration in qual- ity life, and whether the treatment has been discontinued or modified; and (3) death/hospice covers if the patient has died or transferred to hospice, and date and other details on either. The Annotation system is updated for this objective to include a Critic agent at the end that validates the answer generated by the LLM to check whether the outcomes are correctly picked from the note and placed in the JSON structure. If any of the labels are generated incorrectly in the JSON structure, the output is returned to the LLM along with the note chunks to be reused for proper extraction. This feedback loop ensures that the LLM does not hallucinate. The additional step in the system can be seen in Figure 2. Our motivation with the feedback loop is to prevent any false alarms with the progression, toxi- city, or hospice of death. As such, emphasis was given more so on correcting any false alarms. 3.4 Results and Discussion A random sample of 225 oncological notes was obtained. Of those, all 225 notes were compared for progression and toxicity, and 50 from the 225 notes were compared for death/hospice. These notes were hand-annotated to check for the aforementioned labels if they were given in the note. The same notes are then passed to the system label by label and is validated with the original hand annotations. 6F. Shahid et al. Language Model Extracted Labels Yes No Is label positive? Critic Agent No Yes Is output valid? Store JSON Resend chunks Fig. 2. The step after LLM inference. The critic agent sends the note chunks back to the LLM if the generated JSON object is invalid. For this problem, only true and false positives were gathered to assess the reliability of the critic agent. Table 2 shows the count of true positives, false positives, and the calculated precision across all three branches. CategoryTPFPPrecision Progression1794679.5% Toxicity1943186.2% Death42884.0% Table 2. Precision across Progression, Toxicity, and Death. LLMs are generally prone to hallucinations and failing to understand infor- mation in the prompt. This necessitated the requirement of a critic agent with a feedback loop to provide stability and improvement in responses. The primary target was to reduce false positives as the LLM often had a tendency to declare a label to be present in the notes even if the information regarding it was absent. Most of the incorrect labels were what the LLM assumed given the details in the note. Much like how it was handled for the LLM Annotation system, our system focuses on explicit mention of progression with discontinuation, toxicity resulted in discontinuation, and death/hospice details to reduce such occurrences. Currently, improvements in the critic agent and the language model have vastly improved the results. The system runs in a parallel model to exponentially speed up the process on a single language model. 4 Breast Cancer Chemotherapy Outcome Modeling The breast cancer dataset consists of 3,409 patients, each with their first recorded treatment plan in the EMR. It includes important clinical features such as pa- tient vitals, demographics, labs, comorbid conditions, and staging information, including overall stage, TNM staging, and cancer grade. Biomarker data in- Early Prediction of Chemotherapy Outcomes7 cludes ER, PR, and HER2 status, as well as performance metrics like ECOG and Karnofsky scores, all extracted using the Language Model. 4.1 Feature Vector The dataset comprises a variety of clinically relevant features to enhance predic- tive modeling. Key feature categories include: Vitals: Body surface area (BSA). Demographics: Age and gender. Labs: Serum Creatinine (SrCr). Comorbid Conditions: Diabetes, hypertension, kidney disease, and anemia, among others in ICD10 and elixhauser groups format along with readmission scores features. Biomarkers: Estrogen receptor (ER), Progesterone receptor (PR), and Human epidermal growth factor receptor 2 (HER2). Performance metrics: Eastern Cooperative Oncology Group (ECOG) and Karnofsky performance status scores. Dosage and treatment Duration: Standardized dosage per week and number of weeks in a treatment plan. Regimens and Combinations: Chemotherapy treatments include both indi- vidual drugs and combinations. Examples include: – Individual drugs: Docetaxel. – Combination regimens: Carboplatin+Docetaxel+Trastuzumab. For regimen analysis, we compiled a list of unique chemotherapy drugs from the NCCN[2] and NIH[1] guidelines approved for breast cancer treatment, using drug GPI codes up to length 8. A total of 553 drugs and drug combinations were extracted, but most were supported by a single patient, indicating that these regimens were highly personalized. To ensure the robustness and generalizabil- ity of our feature vectors, we applied a patient support threshold of 20, reducing the number of regimen combinations used as features to 22. Additionally, we in- cluded 63 unique individual chemotherapy drugs as features, capturing a broad spectrum of treatment options while maintaining model reliability. The number of positive occurrences in the outcome labels in the cohort include progression with treatment discontinuation (42.12%), treatment-related toxicity with discon- tinuation (25.01%), and death or hospice care (2.11%). By using these outcome labels, we establish the failure definitions used as targets in modeling and sur- vival analysis. The drugs with the highest failure percentages in our breast can- cer cohort include Denosumab+Fulvestrant (79.6%), Fulvestrant (70.7%), and Carboplatin+Paclitaxel (72%), indicating a substantial proportion of patients experiencing treatment failure with these regimens. On the other hand, Car- boplatin+Docetaxel+Trastuzumab (30.5%) and Zoledronic acid (34.5%) exhib- ited the lowest failure percentages, suggesting better treatment responses among these patients. 8F. Shahid et al. 5 Survival Analysis and Modeling Survival modeling is a key statistical method for analyzing time-to-event out- comes in cancer treatment. In this study, we used survival analysis to pre- dict chemotherapy treatment failures, defining time to failure from the start of chemotherapy to the observed event and incorporating censoring for patients without failure during the observation period. We employed Random Survival Forest (RSF) to calculate and visualize survival probabilities over time for both failure and non-failure groups by averaging the predicted survival functions. Model performance was evaluated using the concordance index (C-index), which measures how effectively the model ranks patients based on their risk scores. Classification from Survival Models Beyond estimating survival probabil- ities, we employed RSF as a classifier to predict whether a patient would expe- rience treatment failure. We strategically evaluated outcomes at different time points and selected the time point that yielded the optimal evaluation metrics, accuracy, and F1 scores for both classes separately. The survival curves presented in Figure 3 illustrate the mean survival probabilities for different failure states. Fig. 3. Mean survival curves for differ- ent failure states. The black vertical line marks the selected optimal time point for classification. Fig. 4. Calibration curve comparing the predicted probabilities with observed out- comes. Calibration Curves To ensure confidence in our predictions, we evaluated our model’s calibration by comparing the predicted probability distributions to the actual class distributions.We divided the predicted probability range into ten bins and computed the mean predicted probability within each bin. We then plotted this against the observed proportion of positive cases in that bin. The resulting calibration curve, shown in Figure 4, demonstrates how well our model’s predicted probabilities align with actual outcomes. The dashed blue line represents perfect calibration (i.e., an ideal classifier), while the solid orange line represents our model’s performance. A well-calibrated model should closely Early Prediction of Chemotherapy Outcomes9 follow the diagonal, indicating that predicted probabilities accurately reflect ob- served frequencies. 6 Results For the Breast Cancer (C50) cohort, the RSF model achieved a C-index of 0.731, indicating a strong ability to differentiate between high- and low-risk patients. The classifier achieved an accuracy of 0.723 and an F1 score of 0.724 at the op- timal time point of 431 days. The prevalence of treatment failure in this cohort was 50.3%. These results highlight the effectiveness of the RSF model in predict- ing chemotherapy treatment failure. The most influential features in the model include weekly dose, overall stage, and metastatic stage (m stage), indicating the critical role in predicting survival outcomes. Among the treatment-related fea- tures, Denosumab and Cyclophosphamide+Doxorubicin HCL+Paclitaxel also showed significant importance, highlighting their potential impact on patient prognosis. We further extended this approach to four additional prevalent can- cer types, including Colon Cancer (C18), Lung Cancer (C34), Prostate Cancer (C61) and Multiple Myeloma (C90), by considering their relevant phenotypes extracted from the oncological notes with the same methodology. RSF demon- strated consistent performance across these cancer types, effectively capturing survival trends and treatment failure risks within each cohort. Table 3 summa- rizes the model’s performance metrics across all cohorts, highlighting its predic- tive strength in classifying treatment failure. Table 3. Performance Metrics of RSF as a Classifier for Treatment Failure CohortCohort SizeC-indexAccuracyF1 ScoreTime PointFailure C5034090.7310.7230.7244310.503 C1816850.7140.7000.7131090.705 C6120790.7570.7310.6782380.496 C9010720.6750.7350.8181740.646 C3423660.6600.6770.7661220.609 7 Conclusion In this study, we leveraged Large Language Models (LLMs) and survival anal- ysis for the early prediction of chemotherapy outcomes. By extracting pheno- types and treatment labels using a RAG and critic-agent loop, we improved the overall predictive accuracy. Random Survival Forest (RSF) achieved a C-index of 0.731 for breast cancer, with strong classification performance. The approach was also validated across four additional cancers, presenting our approach adapt- ability. By enabling early risk assessment, this framework supports personalized treatment planning. Future work will refine extraction methods,incorporate ad- vanced survival models, undergo clinical validation, and extend applicability to additional cancer types. 10F. Shahid et al. References 1. Drugs Approved for Different Types of Cancer — cancer.gov. https:// w.cancer.gov/about-cancer/treatment/drugs/cancer-type, [Accessed 03-02- 2025] 2. Guidelines for Patients — nccn.org. https://w.nccn.org/patientresources/ patient-resources/guidelines-for-patients, [Accessed 03-02-2025] 3. Allen, W.L., Coyle, V.M., Johnston, P.G.: Predicting the outcome of chemother- apy for colorectal cancer. Current Opinion in Pharmacology 6(4), 332–336 (2006). https://doi.org/https://doi.org/10.1016/j.coph.2006.02.005, https://w.sciencedirect.com/science/article/pii/S1471489206000853, cancer/Immunomodulation 4. Beltagy, I., Lo, K., Cohan, A.: Scibert: A pretrained language model for scientific text. arXiv preprint arXiv:1903.10676 (2019) 5. Froicu, E.M., et al., O.: The use of artificial intelligence in predicting chemotherapy- induced toxicities in metastatic colorectal cancer: A data-driven approach for personalized oncology. Diagnostics 14(18) (2024), https://w.mdpi.com/ 2075-4418/14/18/2074 6. Golbeck, J., Fragoso, G., Hartel, F., Hendler, J., Oberthaler, J., Parsia, B.: The national cancer institute’s thesaurus and ontology. Journal of web semantics 1(1), 75–80 (2003) 7. Huang, J., Yang, D.M., Rong, R., Nezafati, K., Treager, C., Chi, Z., Wang, S., Cheng, X., Guo, Y., Klesse, L.J., et al.: A critical assessment of using chatgpt for extracting structured data from clinical notes. npj Digital Medicine 7(1), 106 (2024) 8. Huang, Y., Li, J., Li, M., Aparasu, R.R.: Application of machine learning in pre- dicting survival outcomes involving real-world data: a scoping review. BMC Med. Res. Methodol. 23(1), 268 (Nov 2023) 9. Lee, J., Yoon, W., Kim, S., Kim, D., Kim, S., So, C.H., Kang, J.: Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics 36(4), 1234–1240 (2020) 10. Lin, F.P., Groza, T., Kocbek, S., Antezana, E., Epstein, R.J.: Cancer care treat- ment outcome ontology: a novel computable ontology for profiling treatment out- comes in patients with solid tumors. JCO clinical cancer informatics 2, 1–14 (2018) 11. Monajatipoor, M., Yang, J., Stremmel, J., Emami, M., Mohaghegh, F., Rouhsedaghat, M., Chang, K.W.: Llms in biomedicine: A study on clinical named entity recognition. arXiv preprint arXiv:2404.07376 (2024) 12. Prat, A., Lluch, A.e.a.: Predicting response and survival in chemotherapy-treated triple-negative breast cancer. Br. J. Cancer 111(8), 1532–1541 (Oct 2014) 13. Tahmassebi, A., et al., W.: Impact of machine learning with multiparametric mag- netic resonance imaging of the breast for early prediction of response to neoadju- vant chemotherapy and survival outcomes in breast cancer patients. Invest. Radiol. 54(2), 110–117 (Feb 2019) 14. Wang, L., Ma, Y., Bi, W., Lv, H., Li, Y.: An entity extraction pipeline for medical text records using large language models: Analytical study. Journal of Medical Internet Research 26, e54580 (2024) 15. Yang, J., Liu, C., Deng, W., Wu, D., Weng, C., Zhou, Y., Wang, K.: Enhanc- ing phenotype recognition in clinical notes using large language models: Phe- nobcbert and phenogpt. Patterns 5(1), 100887 (2024). https://doi.org/https: //doi.org/10.1016/j.patter.2023.100887, https://w.sciencedirect.com/ science/article/pii/S266638992300288X