Paper deep dive
Siam-ALS: An Explainable Siamese Deep Learning Framework for Early ALS Detection Using EEG and Eye-Tracking Signals
Abdur Rahman, A. A. M. Muzahid, Istiak Ahmed, Dipto Biswas, Hua Han, Yujin Zhang, Ferdous Sohel
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Amyotrophic lateral sclerosis (ALS) diagnosis remains challenging due to heterogeneous symptoms and reliance on clinical assessment, underscoring the need for objective, data-driven approaches that leverage patient-generated physiological signals. We propose Siam-ALS, an explainable Siamese deep learning framework that jointly models synchronized electroencephalography (EEG) and eye-tracking (ET) signals. EEG captures ALS-related changes in brain activity, while ET provides information about oculomotor control, allowing the model to learn complementary patterns associated with ALS. The framework employs two modality-specific encoders: a convolutional EEG encoder that captures spatiotemporal neural patterns and an MLP-based ET encoder that models oculomotor dynamics. These modality-specific features are processed through weight-sharing Siamese branches and fused into a compact 64-dimensional joint embedding. A lightweight classifier is used to predict ALS, while pairwise metric learning helps the model learn more effective patterns from the small, diverse ALS datasets. Performance was evaluated on the public EEGET-ALS dataset using strict subject-level cross-validation, where Siam-ALS achieved an accuracy of 0.988 and an AUROC of 0.998, demonstrating strong performance compared with classical machine learning, DL, and non-Siamese fusion approaches. Cross-dataset evaluation on the independent BigP3BCI dataset achieved an AUROC of approximately 0.98, supporting the generalization capability of Siam-ALS across different BCI paradigms. To enhance interpretability, six XAI methods were evaluated in the learned embedding space, with SHAP achieving the highest, though statistically comparable, fidelity in identifying important EEG and ET patterns that contribute to ALS predictions. Overall, Siam-ALS provides an accurate, generalizable, and explainable framework for multimodal ALS detection.
Tags
Links
- Source: https://arxiv.org/abs/2610.15351v1
- Canonical: https://arxiv.org/abs/2610.15351v1
Trouble viewing inline? Open PDF directly →
Full Text
101,289 characters extracted from source content.
Expand or collapse full text
Siam-ALS: An Explainable Siamese Deep Learning Framework for Early ALS Detection Using EEG and Eye-Tracking Signals Abdur Rahman Affiliation: School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China A. A. M. Muzahid Email: muzahid@sues.edu.cn Affiliation: School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China Affiliation: School of Information Technology, Murdoch University, Murdoch, 6150, WA, Australia Istiak Ahmed Affiliation: School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China Dipto Biswas Affiliation: Department of Information and Communication Engineering, Daffodil International University, Dhaka, 1216, Bangladesh Hua Han Affiliation: School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China Yujin Zhang Affiliation: School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China Ferdous Sohel Email: f.sohel@murdoch.edu.au Affiliation: School of Information Technology, Murdoch University, Murdoch, 6150, WA, Australia Abstract Amyotrophic lateral sclerosis (ALS) diagnosis remains challenging due to heterogeneous symptoms and reliance on clinical assessment, underscoring the need for objective, data-driven approaches that leverage patient-generated physiological signals. We propose Siam-ALS, an explainable Siamese deep learning (DL) framework that jointly models synchronized electroencephalography (EEG) and eye-tracking (ET) signals. EEG captures ALS-related changes in brain activity, while ET provides information about oculomotor control, allowing the model to learn complementary patterns associated with ALS. The framework employs two modality-specific encoders: a convolutional EEG encoder that captures spatiotemporal neural patterns and an MLP-based ET encoder that models oculomotor dynamics. These modality-specific features are processed through weight-sharing Siamese branches and fused into a compact 64-dimensional joint embedding. A lightweight classifier is used to predict ALS, while pairwise metric learning helps the model learn more effective patterns from the small, diverse ALS datasets. Performance was evaluated on the public EEGET-ALS dataset using strict subject-level cross-validation, where Siam-ALS achieved an accuracy of 0.988 and an AUROC of 0.998, demonstrating strong performance compared with classical machine learning (ML), DL, and non-Siamese fusion approaches. Cross-dataset evaluation on the independent BigP3BCI dataset achieved an AUROC of approximately 0.98, supporting the generalization capability of Siam-ALS across different BCI paradigms. To enhance interpretability, six XAI methods were evaluated in the learned embedding space, with SHAP achieving the highest, though statistically comparable, fidelity in identifying important EEG and ET patterns that contribute to ALS predictions. Overall, Siam-ALS provides an accurate, generalizable, and explainable framework for multimodal ALS detection. keywords Amyotrophic lateral sclerosis, Explainable AI, Brain-computer interface (BCI), Multimodal fusion, Siamese neural networks, Biomedical signal processing 1 Introduction ALS is a rapidly progressing disease of the nervous system that destroys motor neurons, gradually diminishing a person’s ability to move, speak, and communicate Hardiman et al. (2022). The diagnostic process today is slow: doctors rely on clinical judgment, electromyography, and the exclusion of other look-alike conditions Turner and Al-Chalabi (2022), a process that can take months and delay access to early treatment, clinical trials, and assistive devices Al-Chalabi et al. (2016). Early diagnosis matters because ALS progresses quickly, and every month lost to diagnostic delay is a month of missed access to disease-modifying therapies, clinical trial enrollment, and timely assistive support Grollemund et al. (2019); Bede and Chio (2022). Yet early detection remains difficult: symptoms overlap with other neuromuscular disorders, and non-invasive biomarkers such as EEG and ET signals are still being validated for routine clinical use Dukic et al. (2022); Iyer et al. (2015); Spataro et al. (2014); Hwang et al. (2017). Compounding this, the ML models built on such signals often serve as opaque predictors, offering little insight into the underlying biology that drives their decisions Tjoa and Guan (2021); Loh et al. (2022). These gaps have increased interest in non-invasive, data-driven screening tools that may help identify ALS earlier than traditional diagnostic approaches. Current ALS detection methods have progressed from classical ML on handcrafted clinical or motor-imagery features Liu et al. (2017); Kocar et al. (2021); Chio et al. (2009); Al-Chalabi et al. (2016) to DL directly on raw EEG or EMG using CNNs, RNNs, and attention-based architectures Grollemund et al. (2019); Bede and Chio (2022); Sengur et al. (2017); Altuwaijri and Muhammad (2022). Although these models improve detection accuracy, they provide limited information about the features responsible for their predictions Tjoa and Guan (2021); Band et al. (2023). Separately, ET has been recognized as a non-invasive biomarker for detecting motor and cognitive changes in ALS patients Proudfoot et al. (2015). Previous studies combining ET with EEG have mainly focused on applications such as attention classification using simple early or late fusion approaches Vortmann et al. (2022). These methods combine features from both modalities but do not explicitly model the complex temporal relationship between brain activity and eye movement patterns. Deeper cross-modal architectures such as Siamese and attention-based networks address this coupling problem in other domains You et al. (2023); Lin et al. (2025); Le et al. (2025) but have not been applied to ALS, and explainability tools like SHAP and attention weights remain largely unused in this specific context Founta et al. (2023); Qin et al. (2025); Chowdhury et al. (2026). This left two gaps unaddressed together: no ALS-specific model captures the deeper EEG–ET coupling, and none pairs this with clinically interpretable output. EEG captures ALS-related changes in brain activity Dukic et al. (2022); Iyer et al. (2015), while ET provides information about eye-movement control affected by the nervous system Spataro et al. (2014); Hwang et al. (2017). Thus, these two signals provide complementary information for ALS detection, and their integration into a joint model can yield a more complete view of ALS-related changes than either signal alone. Inspired by this, we propose a multimodal EEG–ET Siamese architecture with embedding-level explainability. The framework is designed to directly learn interactions between neural and oculomotor signals while producing interpretable representations consistent with known ALS biology. Beyond ALS, EEG–ET fusion has been explored in applications such as emotion recognition and attention classification Vortmann et al. (2022). However, many existing approaches rely on shallow feature fusion strategies Fu et al. (2023), where features from different modalities are combined without explicitly learning a shared representation or capturing their underlying interactions. Recent transformer-based EEG–ET models have improved multimodal representation learning for tasks such as depression detection, yet they remain task-specific and have not been extended to neurodegenerative disease screening Zhu et al. (2025); Pillalamarri and Shanmugam (2025); Bin Heyat et al. (2024). Within ALS, EEG and ET have primarily been investigated as separate modalities, despite gaze-controlled BCI spellers used by many ALS patients naturally generating both signals simultaneously Peters et al. (2020); Ngo et al. (2024); however, their integration for early ALS prediction remains largely unexplored. Theoretically, the two (EEG and ET) are not redundant: EEG reflects central motor and cortical degeneration, while ET reflects downstream oculomotor output that remains relatively preserved even in late-stage ALS Proudfoot et al. (2015), so modeling how these two levels of the motor pathway co-vary should reveal information neither signal captures alone. Moreover, existing ALS detection models, typically EEG-only, offer no explanation of their predictions Fernandes et al. (2021); Grollemund et al. (2019); Bede and Chio (2022), no reliability guarantees on small, imbalanced cohorts Grollemund et al. (2019); Fernandes et al. (2021), and no exploration of EEG-ET fusion for this task Ngo et al. (2024). Together, these limitations limit the ability to build accurate, explainable, and robust ALS detection models that support clinical decision-making. To address these limitations, we introduce Siam-ALS, a Siamese DL framework that jointly learns an explainable EEG–ET representation for early ALS detection. The framework uses two weight-sharing branches to encode EEG and ET signals into a shared 64-dimensional embedding space, where sample pairs are compared through metric learning rather than relying solely on individual classification. This approach enables robust and generalizable predictions while maintaining clinical interpretability. The main idea of combining EEG and ET is to capture complementary aspects of ALS-related changes, where EEG reflects underlying neural dysfunction and ET provides information about oculomotor control alterations. Additionally, the pairwise formulation with the Siamese framework is more data-efficient under small cohorts than standard classifiers, which tend to fit patient-specific quirks rather than disease-general patterns You et al. (2023); Iloon et al. (2022). Because both modalities are projected into the same embedding space, the network learns their cross-modal coupling directly, rather than merging two separate feature sets only at the final decision layer as shallow fusion does. Furthermore, to improve model transparency and clinical trust, we apply explainable artificial intelligence (XAI) methods to identify which learned EEG–ET patterns contribute to ALS predictions. We investigate six XAI techniques, including SHAP, LIME, Integrated Gradients, Saliency, Occlusion, and Permutation Importance, to interpret the learned representation and examine whether the identified brain and eye-movement features align with known ALS-related biological mechanisms Lundberg et al. (2020). Therefore, our proposed Siam-ALS provides a unified framework that addresses generalization, multimodal fusion, and explainability through pairwise similarity learning in a shared, low-dimensional embedding space. The main contributions of this work are summarized as follows: i) We introduce Siam-ALS, a novel explainable Siamese DL framework that jointly models EEG and ET signals for early ALS detection. By combining EEG-based neural activity patterns with ET-based oculomotor control information, the framework captures complementary manifestations of ALS-related changes, thereby improving disease characterization. i) We develop a pairwise metric-learning strategy with weight-sharing branches to improve representation learning on the small, heterogeneous datasets commonly found in ALS research. This approach enables robust generalization while reducing reliance on conventional individual sample classification. i) We compare Siam-ALS against a wide range of classical ML and DL models on the EEGET-ALS dataset, where it matches the highest AUROC and achieves the best threshold-dependent metrics while being the only explainable model. iv) We demonstrate cross-dataset generalizability on EEGET-ALS and BigP3BCI datasets using the same encoder and investigate six XAI methods, where SHAP provides meaningful interpretations by identifying key EEG–ET features associated with ALS prediction. 2 Literature Review Early detection of ALS remains difficult because the disease progresses quickly, varies across patients, and lacks a single reliable non-invasive biomarker. Researchers have explored EEG-based ML/DL Alonso-Vázquez et al. (2023), while EEG–eye tracking fusion has separately proven effective in eye-based interaction systems for individuals with disabilities Le et al. (2025). ET has also attracted interest, as eye movements often remain intact even in late-stage, locked-in ALS. More recently, Siamese neural networks (SNNs) and contrastive representation learning have opened new possibilities for models that work well with very little labeled data. This review covers five relevant areas: i) EEG biomarkers and neurophysiology of ALS i) Multimodal EEG–ET fusion in BCI systems i) Siamese network architectures for biosignal learning iv) Explainability in EEG and biomedical DL. v) Deep EEG models and cross-dataset generalization 2.1 EEG biomarkers and neurophysiology of ALS EEG is widely used to study brain dysfunction in ALS, especially damage to upper motor neurons, which is hard to measure clinically Donaghy and Pioro (2024). Studies consistently show that ALS disrupts normal brain rhythms, producing distinct patterns. Dukic et al. analyzed resting-state EEG and found four distinct ALS sub-types based on brainwave and connectivity patterns Dukic et al. (2022), suggesting EEG can capture the disease’s variability and support more personalized diagnosis. Chmiel et al. reviewed multiple studies and found that reduced alpha-wave power is the most consistent EEG marker of ALS Chmiel and Stępień-Słodkowska (2025), widely seen as a sign of brain hyperexcitability and motor neuron damage Calma et al. (2024). Beyond simple brainwave power, researchers have also studied brain connectivity during motor and speech tasks using related electrophysiological modalities. For example, Dash et al. showed that beta-band connectivity changes during speech tasks, measured via magnetoencephalography (MEG), could reliably separate ALS patients from healthy controls Dash et al. (2024), complementing EEG-based spectral findings. These findings motivated ML and DL applications: Pancotti et al. used recurrent neural networks on longitudinal clinical data to predict how a patient’s condition would decline over time Pancotti et al. (2022). However, most of these methods rely on manually designed features or require large labeled datasets, limiting their generalization, which motivates architectures that can learn patient-independent representations without heavy manual feature engineering. 2.2 Multimodal EEG–Eye Tracking Fusion in BCI Systems As gaze-controlled communication devices become common in ALS care, ET has become more than a cursor-control tool: eye movements themselves carry physiological information, and abnormal patterns, such as unstable fixation, appear across several neurodegenerative diseases. Peters et al. showed that ET remains functional even in late-stage ALS Peters et al. (2020), making it practical to capture EEG and ET together using existing spelling devices without additional patient burden. Synchronized multimodal datasets, historically scarce, are now emerging: Ngo et al.’s EEGET-ALS dataset provides synchronized 32-channel EEG and ET from ALS patients and controls during a spelling task Ngo et al. (2024) (the main dataset used here), while Guttmann-Flury et al. released a larger EEG–ET–video dataset across multiple BCI tasks Guttmann-Flury et al. (2025). Outside ALS, EEG–ET fusion consistently improves performance, in attention decoding Vortmann et al. (2022), neuromarketing Usman et al. (2025), emotion recognition Fu et al. (2023), and multimodal biosignal-based mental-state detection, with related work confirming EEG’s sensitivity to visual processing Yue et al. (2022). Yet multimodal fusion for ALS remains largely untested, and most hybrid BCI systems still rely on shallow fusion, concatenating features or averaging outputs, which misses the deeper brain–eye relationship and assumes more data than ALS research can provide. 2.3 Siamese Neural Networks for EEG and Biosignal Learning A core challenge in ALS-based BCI research is the shortage of large labeled datasets: standard CNNs and Transformers need large training sets and tend to overfit on small, heterogeneous patient groups. SNNs offer an alternative: learning whether two samples belong to the same class by comparing their representations rather than classifying each one directly, allowing the network to learn patient-independent features from very little data. Beyond biosignals, hierarchical attention mechanisms have similarly strengthened Siamese architectures for real-time discrimination tasks such as visual tracking, underscoring the generality of the Siamese paradigm for similarity-based learning. More advanced designs have addressed multimodal and cross-subject tasks, dual-attention SNNs for human-robot collaboration Zhou et al. (2025), EEG-signature fusion for identity verification Munia et al. (2021); Rehman et al. (2025), and therapy progress tracking in cerebral palsy Moreno Escobar et al. (2024). SNNs also reduce the cross-subject "calibration bottleneck" in BCI systems. Despite this progress in sleep staging, seizure detection, and motor imagery, no prior work has applied SNNs to multimodal ALS detection, even though their ability to fuse different signal types under limited data makes them well suited for exactly this problem. 2.4 Explainability in EEG and Biomedical Deep Learning Alongside the push for higher accuracy, a separate line of work has focused on making biomedical DL models interpretable, since clinicians are often unwilling to trust a prediction they cannot understand. SHAP has become one of the most widely used tools for this purpose, assigning each input feature a contribution score to the final prediction Lundberg et al. (2020). In EEG research, SHAP and similar attribution methods have identified which channels, frequency bands, or time windows drive a model’s decision in tasks like seizure detection and sleep staging, helping confirm that models rely on physiologically meaningful signals rather than noise, while saliency- and attention-based methods have similarly highlighted important spatial and temporal regions in raw EEG. However, this kind of analysis remains rare in ALS research and virtually absent in multimodal EEG–ET work: most ALS studies reviewed above report accuracy or AUROC without explaining which features or brain regions drove the output. This gap matters especially for ALS, where an incorrect or unexplained prediction carries real clinical consequences, motivating the embedding-level SHAP analysis used in this work. 2.5 Deep EEG Models and Cross-Dataset Generalisation DL has reshaped EEG decoding: CNNs and, more recently, Transformers have largely replaced pipelines built on manually designed features. Altaheri et al. Altaheri et al. (2023) showed temporal-spatial convolutions outperform earlier motor-imagery methods by learning features automatically from raw signals, and this was extended by fusing multi-scale convolutional branches with attention Yang and Liu (2024) and by combining channel attention with continuous-time temporal modeling for emotion recognition from raw multi-channel EEG Wang et al. (2024). The pattern holds elsewhere: deep models beat classical correlation analysis for SSVEP classification, Transformers capture long-range temporal patterns that standard convolutions miss, and hybrid CNN-LSTM Xu et al. (2025) and graph-based models have modeled brain connectivity effectively Liu et al. (2025), with related hybrid frameworks performing well in clinical applications such as neurological disorder detection Ma et al. (2026). Still, a major challenge remains: models trained in one lab often fail in another due to hardware, electrode, or population differences, the generalization gap. Transfer learning can help, but it still usually requires some labeled data from the new setting, limiting its true plug-and-play use, which is why single-dataset validation is no longer sufficient proof of clinical usefulness. Recent work instead learns representations that stay stable across subjects, via Euclidean alignment Junqueira et al. (2024), contrastive learning, and contrastive predictive coding. Motivated by this, the present study goes beyond internal cross-validation and also tests on the independent BigP3BCI dataset Mainsah et al. (2025), using the same architecture and training setup as in our main experiments, to check whether the learned EEG representations truly generalize across recording setups and patient populations, a key requirement for any clinically usable ALS detection system. 3 METHOD ALS detection is formulated as a window-level binary classification problem. Each sample is a short, overlapping temporal window of multichannel EEG aligned with a corresponding ET segment, supporting both unimodal (EEG-only) and multimodal (EEG+ET) learning. The proposed pipeline consists of three main stages: (i) data preprocessing, covering signal cleaning, epoching, ET feature extraction, and class balancing; (i) a DL framework that encodes EEG and ET signals into modality-specific representations; and (i) development of Siam-ALS, a Siamese architecture that fuses these representations into a compact, subject-invariant embedding for explainable ALS classification. 3.1 Data Preparation All EEG and ET recordings undergo a unified preprocessing pipeline tailored to the proposed window-based learning framework and multimodal architecture. Raw EEG signals are band-pass filtered to suppress low-frequency drift and high-frequency noise, re-referenced to a common average, and normalized on a per-channel basis to facilitate stable representation learning. ET signals, when available, are temporally aligned with EEG using event markers and resampled to ensure synchronization across modalities. Both EEG and ET streams are segmented into short, fixed-length overlapping windows, which constitute the fundamental input units of the proposed model. Each window inherits the subject-level diagnostic label (ALS or control), enabling weakly supervised learning at the window level. To prevent subject leakage and ensure reliable generalization, preprocessing, window extraction, and normalization are performed independently within each cross-validation fold using subject-level splits. Table 1 summarizes the two datasets used in this study, described in detail below. Table 1: Summary of datasets used in this study Dataset Modalities Classes Usage EEGET-ALS EEG + ET ALS vs Healthy Primary dataset BigP3BCI EEG only ALS vs Healthy Generalization 3.1.1 Data Sources and Sampling EEGET-ALS Our primary dataset, EEGET-ALS Ngo et al. (2024), contains synchronized 32-channel EEG and ET recordings from ALS patients and controls during a gaze-controlled spelling task. We use only full 32-channel sessions with valid ET streams, balance classes by downsampling, and split subjects at the subject level to avoid leakage. BigP3BCI As a secondary, EEG-only dataset Mainsah et al. (2025), BigP3BCI provides P300-speller EEG recordings from ALS and healthy participants, used for cross-dataset generalization. We infer ALS/Non-ALS labels from file metadata, exclude ambiguous cases, and build class-balanced, subject-level stratified folds as with EEGET-ALS. 3.1.2 Data Processing EEG Preprocessing and Epoch Construction For both corpora, continuous EEG was band-pass filtered (0.5–40 Hz) with notch filtering to attenuate slow drifts, line noise, and high-frequency noise, re-referenced to a common average, and resampled to 128 Hz for EEGET-ALS and 124 Hz for BigP3BCI. In EEGET-ALS, signals were segmented into 32×12832× 128 epochs (channels × time samples) via a 1-s sliding window with 50% overlap, time-locked to spelling-related events; epochs below the target length or flagged as corrupted in the metadata were discarded. In BigP3BCI, all recordings were first restricted to a common channel subset present across the selected EDF files, and 1-s windows (124 samples) were then extracted with 50% overlap, excluding recordings too short to yield a complete window. Each BigP3BCI window was zero-padded to 128 samples to match the encoder input (Section 4.8) and inherited its subject’s ALS/non-ALS label. For both datasets, channel-wise standardization was applied strictly within each cross-validation fold: mean and standard deviation were estimated on the training partition only and subsequently used to normalize the training, validation, and test epochs, preventing information leakage across splits. Fig. 1(a)–(b) illustrates this pipeline on a representative 10-s window, showing raw EEG overlaid with synchronized gaze trajectories, followed by the corresponding band-pass filtered and resampled signal. Eye-Tracking Feature Extraction (EEGET-ALS Only) EEGET-ALS provides ET streams sampled at approximately 30 Hz, comprising gaze coordinates and per-sample validity flags. Each EEG epoch was temporally aligned with its corresponding ET segment and summarized into a nine-dimensional feature vector encoding gaze position, velocity, path length, and signal validity (Table 2). Consistent with the EEG pipeline, ET features were z-score normalized per fold using training-set statistics only, and epochs with insufficient or invalid ET coverage were excluded from analysis. Fig. 1(c) shows the resulting standardized horizontal and vertical gaze trajectories after removal of invalid samples. Class Balancing To mitigate bias arising from class imbalance, ALS and non-ALS windows were balanced by random downsampling of the majority class, applied independently within each fold after preprocessing and preserving the subject-level splits. Training, validation, and test partitions are therefore class-balanced, yielding a window-level ALS prevalence of 50%. Consequently, prevalence-dependent metrics (PPV, NPV) reflect this balanced setting rather than clinical prevalence. Fig. 2 summarizes the complete EEG and ET preprocessing pipeline, including class balancing. Table 2: Engineered ET features extracted per epoch from the EEGET-ALS dataset. Each feature is z-score normalized per fold using training-set statistics only No. Feature Category 1 Mean horizontal gaze position Position 2 Std. horizontal gaze position Position 3 Mean vertical gaze position Position 4 Std. vertical gaze position Position 5 Mean gaze velocity Dynamics 6 Std. gaze velocity Dynamics 7 Total gaze path length Dynamics 8 Proportion of low-velocity (near-fixation) samples Fixation 9 Proportion of valid gaze samples Quality Figure 1: Multimodal synchronization and preprocessing of EEG and ET signals over a 10-second interval. (a) Raw EEG signal overlaid with temporally aligned valid gaze trajectories (horizontal and vertical components). (b) Preprocessed EEG after band-pass filtering (0.5–40 Hz), notch filtering, and resampling, demonstrating improved signal quality and noise reduction. (c) Standardized ET trajectories after removal of invalid samples, capturing normalized gaze dynamics Figure 2: EEG and eye-tracking preprocessing pipeline with window-level class balancing between ALS and control. Blue blocks denote the EEG-only processing path; orange blocks denote the ET-only path; both streams converge at the green class-balancing stage, which produces the balanced epoch sets used by all downstream models Figure 3: The block diagram of our Siam-ALS framework. Two branches process EEG windows and ET features from subject i and subject j with shared EEG and ET encoders. Fused 64-D embeddings iz_i and jz_j are compared via an ℓ1 _1 distance Siamese head, and an embedding-level classifier operates on the learned representation for ALS detection. The detailed configurations of Conv2D, pooling, EEG and ET encoders and MLP are provided in Table 3. (ET and EEG encoders share weights) Table 3: Parameter configuration and architectural design of the proposed multimodal Siam-ALS network. Both weight-sharing branches use identical parameters across all layers, enforcing the Siamese learning constraint. Notation matches Section 3.2: ie_i (EEG window), ig_i (ET features), iz_i (fused embedding) Block Layer Configuration Input → Output Params EEG encoder fEEG(i)f_EEG(e_i) Conv2D-1 1×151× 15, 16 filters, padding (0,7)(0,7) (1,32,128)→(16,32,128)(1,32,128)→(16,32,128) 256 BN-1 BatchNorm (16 ch.) (16,32,128)→(16,32,128)(16,32,128)→(16,32,128) 32 DepthwiseConv 8×18× 1, 32 filters, groups=16=16 (16,32,128)→(32,25,128)(16,32,128)→(32,25,128) 288 BN-2 BatchNorm (32 ch.) (32,25,128)→(32,25,128)(32,25,128)→(32,25,128) 64 AvgPool-1 1×41× 4 (32,25,128)→(32,25,32)(32,25,128)→(32,25,32) – Conv2D-2 1×151× 15, 64 filters, padding (0,7)(0,7) (32,25,32)→(64,25,32)(32,25,32)→(64,25,32) 30,784 BN-3 BatchNorm (64 ch.) (64,25,32)→(64,25,32)(64,25,32)→(64,25,32) 128 AvgPool-2 1×41× 4 (64,25,32)→(64,25,8)(64,25,32)→(64,25,8) – Dropout p=0.5p=0.5 (64,25,8)→(64,25,8)(64,25,8)→(64,25,8) – FC-EEG Linear 12,800→6412,800→ 64 flatten → 64-D EEG embedding 819,264 ET encoder fET(i)f_ET(g_i) FC-1 Linear 9→649→ 64, ReLU 9→649→ 64 640 Dropout p=0.3p=0.3 64→6464→ 64 – FC-2 Linear 64→3264→ 32, ReLU 64→3264→ 32 (ET embedding) 2,080 Fusion encoder ffuse(⋅)f_fuse(·) Concat EEG(64) ⊕ ET(32) 96-D vector – FC-1 Linear 96→12896→ 128, ReLU 96→12896→ 128 12,416 Dropout p=0.4p=0.4 128→128128→ 128 – FC-2 Linear 128→64128→ 64 128→64128→ 64 (iz_i, final embedding) 8,256 Siamese head (Eq. (2)) ℓ1 _1 distance dab=|a−b|d_ab= _a-z_b 64→6464→ 64 – FC-1 Linear 64→6464→ 64, ReLU, Dropout p=0.3p=0.3 64→6464→ 64 4,160 FC-2 + Sigmoid Linear 64→164→ 1 + sigmoid 64→s^ab64→ s_ab 65 3.2 Proposed Siam-ALS Architecture Siam-ALS is a Siamese dual-branch network that learns a joint, subject-invariant embedding of EEG and ET signals through pairwise metric learning, rather than direct per-sample classification. As illustrated in Fig. 3, two weight-sharing branches, indexed by samples i and j, each process an input pair through identical modality-specific encoders and a shared fusion encoder, producing fused embeddings that are compared via a similarity head during training. This design is deliberately chosen for two reasons: (i) pairwise supervision increases the effective number of training instances available from a small cohort, mitigating overfitting under scarce ALS data, and (i) constraining both branches to identical weights enforces a single, subject-invariant mapping ϕ(⋅)φ(·), rather than allowing the network to memorize subject-specific shortcuts. 3.2.1 Modality-Specific Encoders Each branch independently encodes its EEG and ET inputs prior to fusion. The EEG encoder follows an EEGNet-style design, combining standard and depthwise convolutions with batch normalization, average pooling, and dropout to capture spatiotemporal cortical patterns while remaining parameter-efficient. The ET encoder is a lightweight two-layer multilayer perceptron (MLP) with an intermediate dropout layer; it is intentionally shallower than the EEG encoder, reflecting the lower dimensionality and intrinsic simplicity of the nine-dimensional ET feature vector relative to raw multichannel EEG. Full layer-wise configurations for both encoders are provided in Table 3. 3.2.2 Fusion Encoder and Joint Embedding Space The modality-specific embeddings are combined by a fusion encoder ffuse(⋅)f_fuse(·), implemented as a compact MLP, which projects the concatenated EEG and ET representations into a single 64-dimensional joint embedding: i=ϕ(i)=ffuse(fEEG(i),fET(i)),z_i=φ(x_i)=f_fuse\! (f_EEG(e_i),\;f_ET(g_i) ), (1) where ie_i and ig_i are the EEG window and ET feature vector for sample i, and i∈ℝ64z_i ^64 is the resulting fused embedding shared by both the Siamese similarity head and the downstream embedding-level classifier. The 64-dimensional bottleneck is a deliberate design choice: keeping the joint representation compact, rather than high-dimensional, later enables tractable and explainable AI (XAI) methods across the full embedding space (Section 4.9), as opposed to explaining thousands of raw, correlated input features. 3.2.3 Siamese Pairwise Learning Objective Given two samples a and b, their fused embeddings a,bz_a,z_b are compared via an element-wise ℓ1 _1 distance and passed through a head network h(⋅)h(·): s^ab=σ(h(|a−b|)), s_ab=σ (h( _a-z_b ) ), (2) where h is a two-layer MLP (64→6464→ 64 with ReLU and dropout p=0.3p=0.3, then 64→164→ 1; Table 3), σ(⋅)σ(·) is the sigmoid function, and s^ab∈(0,1) s_ab∈(0,1) is the predicted probability that the pair belongs to the same diagnostic class. The network is trained on balanced sets of same-class (ALS–ALS, healthy–healthy) and different-class (ALS–healthy) pairs using the binary cross-entropy loss ℒpair=−1M∑(a,b)[sablogs^ab+(1−sab)log(1−s^ab)],L_pair=- 1M _(a,b) [s_ab s_ab+(1-s_ab) (1- s_ab) ], (3) where sab∈0,1s_ab∈\0,1\ is the ground-truth pair label and M is the number of training pairs. Minimizing ℒpairL_pair drives the network to assign high similarity to same-class pairs and low similarity to different-class pairs, shaping the embedding space around disease-relevant structure rather than subject identity. Training uses early stopping on validation AUROC. 3.2.4 Embedding-Level Classification Once trained, the Siamese encoders are frozen and used purely as feature extractors, producing one fused embedding iz_i per epoch. A lightweight logistic-regression classifier is then trained on top of these frozen embeddings for sample-level ALS versus non-ALS discrimination: p(yi=1∣i)=σ(⊤i+c),p(y_i=1 _i)=σ (u z_i+c ), (4) where u and c are the classifier’s trainable parameters, optimized with the same cross-entropy objective as in (3). Separating representation learning (Siamese encoders) from decision-making (logistic classifier) is itself a design choice motivated by interpretability: a simple, linear final classifier over a compact, frozen embedding makes the subsequent XAI-based attribution both computationally tractable and analytically exact, as detailed in Section 4.9. The same architecture, restricted to the EEG branch, is also evaluated on the EEG-only BigP3BCI dataset to assess cross-dataset generalization (Section 4). 3.3 Explainability Framework Although Siam-ALS provides strong predictive performance, accuracy alone does not reveal why the model makes a particular decision or whether the learned patterns correspond to meaningful ALS-related characteristics. Therefore, we develop an explainability framework for the frozen-trained model to investigate the features contributing to ALS prediction. The framework consists of two stages: first, a systematic comparison of six XAI methods at the fused-embedding level; and second, a signal-level analysis to trace important learned representations back to their corresponding physiological signals. Embedding-Level Explainability. The explainability analysis is performed on the 64-dimensional fused embedding iz_i generated by the Siam-ALS fusion encoder. This compact representation enables efficient interpretation of the learned multimodal features compared with directly analyzing high-dimensional raw EEG and ET inputs. Since the final classifier operates on this embedding space, explanations are generated with respect to the embedding-level prediction, allowing the contribution of each latent dimension to be quantified. Comparison of XAI Methods. To evaluate different explanation strategies, we investigate six widely used XAI methods, including SHAP, LIME, Integrated Gradients, Saliency, Occlusion, and Permutation Importance. Each method is used to rank the 64 embedding dimensions by their contribution to the model’s output. The reliability of each explanation method is evaluated using fidelity, defined as the AUROC retained when a new classifier is trained using only the top-ranked embedding dimensions identified by the corresponding XAI method. This evaluation measures how effectively each method captures the most informative features learned by Siam-ALS. Feature Attribution and Physiological Mapping. After identifying the most influential embedding dimensions, further analysis is performed to investigate their relationship with the original physiological signals. Embedding-level attribution indicates which learned multimodal representations contribute to ALS predictions, while gradient-based saliency analysis traces individual embedding dimensions back to the original EEG input space. This produces channel- and temporal-importance maps that highlight the EEG regions contributing to the learned features. Two-Level Interpretation Strategy. The proposed explainability framework provides interpretation at two complementary levels. The first level identifies important latent features within the learned EEG–ET representation, while the second level connects these features to the underlying physiological signals. This two-level strategy allows the model predictions to be examined beyond accuracy and provides insights into whether Siam-ALS relies on meaningful EEG and ET patterns associated with ALS-related changes rather than dataset-specific artifacts. 3.4 Baseline Model Families. To rigorously contextualize Siam-ALS, we benchmark it against a broad set of baselines spanning classical ML, deep sequence models, tree ensembles, and non-Siamese multimodal fusion, all trained and evaluated under the identical preprocessing, cross-validation, and evaluation protocol described in Section 3.1. 3.4.1 Classical SVM Baselines. We implement three RBF-kernel SVM baselines on hand-crafted spectral features: SVM–PSD (EEG-only power spectral density across standard frequency bands, concatenated over channels), SVM–PSD+ET (PSD features augmented with the nine ET features from Table 2), and SVM–IHAR (EEG-only, using inter-hemispheric amplitude ratio features from left/right channel pairs). All features are z-score normalized, with C and γ selected via grid search on validation AUROC. The SVM decision function and RBF kernel are given by f()=sign(∑j=1NαjyjK(,j)+b),f(x)=sign\! ( _j=1^N _jy_jK(x,x_j)+b ), (5) K(,′)=exp(−γ∥−′∥22),K(x,x )= \! (-γ -x _2^2 ), (6) where αj,b _j,b are the dual-optimized parameters, yj∈−1,+1y_j∈\-1,+1\ the training labels, and γ the kernel width controlling locality. 3.4.2 Deep EEG Baselines All deep models are trained on identical 32×12832× 128 raw EEG windows with Adam and binary cross-entropy, using early stopping on validation AUROC: pθ(yi=1∣i)=σ(fθ(i)),σ(t)=11+e−t,p_θ(y_i=1 _i)=σ (f_θ(x_i) ), σ(t)= 11+e^-t, (7) ℒCE(θ) _CE(θ) =−1N∑i=1N[yilogpθ(yi=1∣i) =- 1N _i=1^N [y_i p_θ(y_i=1 _i) (8) +(1−yi)log(1−pθ(yi=1∣i))]. +(1-y_i) (1-p_θ(y_i=1 _i) ) ]. Convolutional Models (CNN, EEGNet) The CNN baseline uses generic temporal-spatial filters, while EEGNet incorporates EEG-specific depthwise and separable convolutions for parameter-efficient channel-wise and temporal modeling. Recurrent and Hybrid Temporal Models An LSTM treats each epoch i∈ℝC×Tx_i ^C× T as a sequence, with hidden state t=LSTM(t,t−1)h_t=LSTM(x_t,h_t-1) feeding a linear head as in (7). A CNN–LSTM variant first extracts local convolutional features, then aggregates them recurrently, combining spatial and temporal representation learning. Tree-Based Baselines Random Forest and XGBoost operate on flattened EEG windows: p(y=1∣) p(y=1 ) =1M∑m=1Mhm(), = 1M _m=1^Mh_m(x), (9) f() f(x) =∑k=1Kfk(),fk∈ℱ, = _k=1^Kf_k(x), f_k , (10) with Random Forest averaging M decision trees hm\h_m\ and XGBoost forming an additive ensemble over regression trees ℱF, converted to a probability via (7). State-Space Model (Mamba) As a modern alternative to recurrence, Mamba models temporal evolution via a linear state-space formulation, t+1=t+t,t=t,s_t+1=As_t+Bx_t, _t=Cs_t, (11) with learned ,,A,B,C; the pooled sequence representation is mapped to a binary logit and trained with (8). 3.4.3 Non-Siamese Multimodal Fusion Baselines To isolate the benefit of Siamese metric learning from multimodal fusion itself, we implement two EEGNet+ET baselines, trained end-to-end with Adam and early stopping, using identical splits. The late-fusion variant averages independent EEGNet and ET-MLP probabilities; the middle-fusion variant concatenates EEG and ET embeddings before a shared classifier, representing a stronger, embedding-level (but non-metric-learned) fusion baseline against which Siam-ALS is directly compared. 4 Experiment and Result Analysis 4.1 Training Configuration All experiments were implemented in Python using the PyTorch framework and trained on a GPU-enabled environment (RTX 5090 with 32 GB memory). The software stack consisted of PyTorch 2.8, Python 3.12, and CUDA 12.8. A fixed random seed was used to ensure reproducibility across all experiments. The full Siam-ALS model (EEG+ET fusion, embedding-level classifier) is independently re-trained for each experimental table in which it appears: as A13 in the main baseline comparison (Table 5), as B3 in the modality ablation (Table 6), and as M4 in the architectural ablation (Table 7). These three reportings use an architecturally identical model trained under the identical protocol described above, but as independent runs rather than a single shared checkpoint. Consequently, minor run-to-run variation is expected across these three tables, consistent with the fold-level standard deviations reported throughout this section: AUROC and accuracy typically agree to within ≲ 0.1 percentage points across independent runs, since AUROC is threshold-independent and accuracy is evaluated at a stable default threshold, whereas F1F_1, being sensitive to the precise precision–recall balance at the classification threshold, shows somewhat larger run-to-run variation (≲ 0.4 percentage points). 4.2 Dataset Statistics and Splits In this study, we use two complementary public datasets, EEGET-ALS Ngo et al. (2024) and BigP3BCI Mainsah et al. (2025). The EEGET-ALS corpus provides synchronized EEG and ET recordings from ALS and control participants performing a gaze-assisted spelling task. The BigP3BCI corpus contains raw EEG data from ALS and non-ALS subjects who participated in a P300-based spelling paradigm. We exclude subjects with incomplete or unusable recordings. For both datasets, EEG is segmented into 1-s windows with 50% overlap; for EEGET-ALS, windows are extracted within the labelled task intervals (Thinking, Acting, Typing, and Resting). In EEGET-ALS, each EEG window is paired with a 9-dimensional ET feature vector that summarizes gaze position, velocity, path length, and data validity. All experiments use subject-level cross-validation to avoid leakage: in EEGET-ALS, subjects are partitioned into four folds with approximately balanced class proportions, and in each run, two folds are used for training, one for validation, and one for testing. All windows from a given subject appear in exactly one split. For BigP3BCI, we apply the same 4-fold subject-level protocol to a balanced subset of all 48 ALS subjects and 48 non-ALS subjects selected from the 279 available [SELECTION METHOD, e.g., at random with a fixed seed], used exclusively for the independent cross-dataset generalization experiments in Section 4.8. The EEG encoder is trained from scratch on BigP3BCI using the identical architecture validated on EEGET-ALS, with no shared weights, checkpoints, or pretraining transfer between the two datasets. Table 4.2 summarizes the number of ALS and non-ALS subjects in each dataset. tableorg [t] Subject-Level Statistics for EEGET-ALS and BigP3BCI Dataset ALS Subjects Non-ALS Subjects Total Subjects EEGET-ALS 6 170 176 BigP3BCI 48 279 327 tableorg [t] Window-level sample counts per class after preprocessing and class balancing. Each window is a 1-s segment for both datasets Dataset ALS Windows Non-ALS Windows Total Windows EEGET-ALS 15,280 15,280 30,560 BigP3BCI 244,550 244,550 489,100 Table 4.2 reports the resulting window-level sample counts per class after preprocessing and class balancing. 4.3 Leave-One-Subject-Out Validation To address the limited number of ALS subjects (n=6) in EEGET-ALS, we additionally report leave-one-subject-out (LOSO) cross-validation, where each of the six ALS subjects is held out as the sole test-set ALS subject in turn, with all other ALS subjects and a class-balanced set of control subjects used for training. This yields six independent test folds, each anchored to a single, distinct patient, directly exposing between-subject variance rather than aggregating it into a single 4-fold summary as in Section 4.2. Table 4 reports per-subject AUROC, accuracy, and F1 for the full Siam-ALS model under this protocol. Performance remains high for five of the six ALS subjects (AUROC 0.981–0.996), with a mean AUROC of 0.983±0.0150.983± 0.015 across all six. This is lower than the 4-fold estimate in Table 5 (AUROC 0.998), even though LOSO trains on more ALS subjects (five, versus roughly three under the 4-fold protocol, in which two of four folds are used for training). The difference therefore does not stem from reduced training data. It more plausibly reflects how the test set is formed: each LOSO fold is evaluated on a single ALS patient, so subject-specific recording characteristics cannot be averaged out by other ALS subjects in the same test fold, and the per-fold estimate directly reflects individual patient variability. The two protocols thus answer different questions and are not directly comparable. The 4-fold result estimates pooled window-level performance, whereas LOSO shows how performance varies from patient to patient. Subject S4 shows the lowest AUROC (0.954), which we attribute to a shorter, noisier ET calibration session for this participant. Excluding S4, the mean AUROC across the remaining five subjects is 0.988±0.0060.988± 0.006, which narrows but does not close the gap to the 4-fold estimate. Because the drop is concentrated in a single subject rather than spread across all folds, the 4-fold results do not appear to depend on one favorable train/test partition. The S4 result does, however, indicate sensitivity to session quality, which we identify as a limitation for future data collection. Table 4: Leave-one-subject-out (LOSO) validation results for the full Siam-ALS model on EEGET-ALS. Each row reports performance when the listed ALS subject is held out as the sole test-set ALS subject, trained on the remaining five ALS subjects plus a class-balanced control set Held-Out Subject AUROC Accuracy F1F_1 S1 0.991 0.972 0.968 S2 0.996 0.981 0.979 S3 0.985 0.963 0.958 S4 0.954 0.921 0.907 S5 0.981 0.958 0.951 S6 0.988 0.969 0.964 Mean ± SD 0.983 ± 0.015 0.961 ± 0.021 0.955 ± 0.024 Figure 4: Confusion matrices of Siam-ALS on EEGET-ALS and BigP3BCI at window level (mean per fold) and subject level (majority vote over windows) 4.4 Metrics and Aggregation On each test fold, we evaluate binary classification performance for ALS versus non-ALS using multiple complementary metrics. Let yi∈0,1y_i∈\0,1\ denote the ground-truth label and p^i∈[0,1] p_i∈[0,1] the predicted probability for sample i. The primary metric is the area under the receiver operating characteristic curve (AUROC), defined via the true positive rate (TPR) and false positive rate (FPR): (τ) (τ) =(τ)(τ)+(τ), = TP(τ)TP(τ)+FN(τ), (12) (τ) (τ) =(τ)(τ)+(τ). = FP(τ)FP(τ)+TN(τ). (13) where τ denotes the decision threshold. We additionally report average precision (AP), summarizing the precision–recall curve: =+,=+, = TPTP+FP, = TPTP+FN, (14) =∑n(Rn−Rn−1)Pn. = _n(R_n-R_n-1)P_n. (15) where PnP_n and RnR_n denote precision and recall at the n-th operating point. Threshold-dependent performance is reported using accuracy and F1-score: =++++, = TP+TNTP+TN+FP+FN, (16) F1 F1 =2⋅+. = 2·Precision·RecallPrecision+Recall. (17) All metrics are computed independently on each test fold, and final results are reported as the mean and standard deviation across folds: μ=1F∑f=1Fmf,σ=1F−1∑f=1F(mf−μ)2,μ= 1F _f=1^Fm_f, σ= 1F-1 _f=1^F(m_f-μ)^2, (18) where mfm_f denotes the metric value on fold f and F is the total number of folds. As shown in Section 4.9, however, accuracy and AUROC alone are insufficient to differentiate leading ALS detection methods, since several baselines converge to statistically indistinguishable performance; this motivates evaluating models along a second, equally critical axis: explainability. 4.5 Baseline Comparison on EEGET-ALS We evaluate Siam-ALS against twelve baseline approaches (A1–A12), spanning classical ML, DL, and multimodal fusion strategies, under the identical subject-independent 4-fold cross-validation protocol (Table 5). All baselines reach high AUROC (0.988–0.998), so differences between them are small. The classical SVM baselines (A1–A3; AUROC 0.992–0.993) remain competitive with most deep EEG models: CNN (A4) reaches the highest baseline AUROC (0.998), whereas LSTM (A5; 0.991) and CNN-LSTM (A6; 0.992) do not exceed the SVMs. Deep and tree-based models (A4–A10) do, however, achieve somewhat higher accuracy (0.963–0.974 vs. 0.959–0.962). Among the multimodal baselines, middle fusion (A12), which concatenates EEG and ET embeddings before a shared classifier, clearly outperforms late fusion (A11), which averages unimodal probabilities (AUROC 0.998 vs. 0.988; accuracy 0.983 vs. 0.971). This indicates that fusing at the representation level is more effective than combining decisions. Siam-ALS (A13) matches the highest AUROC (0.998, tied with A4 and A12) and achieves the best value in every threshold-dependent metric (accuracy 0.988, MCC 0.976). Fig. 4(a)–(b) shows the corresponding confusion matrices: at window level, errors are small and balanced between false positives and false negatives, and at subject level, majority voting correctly identifies all six ALS patients with a single misclassified control (169/170). Fig. 6 visualizes this qualitatively via t-SNE projections of the learned embedding spaces: Siam-ALS’s metric-learned fused embedding shows markedly tighter, better-separated ALS/control clusters than either the late-fusion or middle-fusion embeddings, consistent with the pairwise contrastive objective (Eq. (3)) shaping a more discriminative representation than direct end-to-end classification training alone. This result is further examined via leave-one-subject-out validation (Section 4.3, Table 4), which shows high, though lower and more variable, per-patient performance for five of the six ALS subjects, suggesting the 4-fold estimate is not driven by a single favorable train/test partition. Beyond AUROC and accuracy, Table 5 reports sensitivity, specificity, PPV, NPV, MCC, and each model’s explainability status. Siam-ALS achieves the best or joint-best value in every metric column and is the only model marked explainable. Fig. 5 visualizes this across all seven metrics: the Siam-ALS polygon matches the best baseline on AUROC and exceeds it on the remaining six axes. Table 5: Clinical performance comparison on EEGET-ALS across all twelve baselines (A1–A12) and the proposed Siam-ALS (A13), reported as mean over 4 subject-level folds. Feature indicates the representation type each model operates on. PPV and NPV are computed on class-balanced test partitions (50% ALS prevalence) and therefore do not reflect clinical prevalence. Sig. reports uncorrected (p<0.05p<0.05) significance of AUROC difference from Siam-ALS (Table 8) for the seven directly-tested baselines: ✓ = significantly lower AUROC than Siam-ALS; – = not significant or not tested; Reference = comparison baseline. Explainable indicates whether the model supports attribution over a learned, disease-specific representation ID Model Name Feature Explainable AUROC ACC. Sens. Spec. PPV NPV MCC Sig. A1 SVM–PSD Hand-crafted EEG ✗ 0.992 0.959 0.951 0.966 0.965 0.952 0.918 ✓ A2 SVM–PSD + ET Hand-crafted EEG ✗ 0.993 0.962 0.958 0.967 0.967 0.958 0.925 ✓ A3 SVM–Ihar Hand-crafted EEG ✗ 0.992 0.959 0.949 0.968 0.967 0.950 0.917 ✓ A4 CNN Raw EEG ✗ 0.998 0.974 0.971 0.977 0.977 0.971 0.947 – A5 LSTM Raw EEG ✗ 0.991 0.963 0.955 0.970 0.970 0.956 0.926 – A6 CNN-LSTM Raw EEG ✗ 0.992 0.968 0.963 0.972 0.972 0.963 0.936 – A7 Random Forest Raw EEG Partial 0.994 0.971 0.966 0.976 0.976 0.966 0.942 – A8 XGBoost Raw EEG Partial 0.993 0.970 0.964 0.975 0.975 0.964 0.939 – A9 Mamba (SSM) Raw EEG ✗ 0.993 0.973 0.968 0.977 0.977 0.968 0.945 – A10 EEGNet Raw EEG ✗ 0.994 0.969 0.963 0.974 0.974 0.963 0.937 ✓ A11 EEGNet+ET Late Fusion Raw EEG + ET ✗ 0.988 0.971 0.965 0.976 0.976 0.965 0.941 – A12 EEGNet+ET Middle Fusion Raw EEG + ET ✗ 0.998 0.983 0.980 0.985 0.985 0.980 0.965 – A13 Siam-ALS (Proposed) Learned EEG + ET ✓ 0.998 0.988 0.986 0.990 0.990 0.986 0.976 Reference ∗Siam-ALS (A13) is re-trained independently for each comparison table in this paper and appears again as B3 in the modality ablation (Table 6) and as M4 in the architecture ablation (Table 7); minor numerical differences across these three reportings reflect run-to-run training variation rather than a different model. Figure 5: Clinical performance envelope of Siam-ALS against the full range of twelve baselines (A1–A12) on EEGET-ALS, across seven clinically relevant metrics (Table 5). The dotted inner polygon shows the weakest baseline per metric, the dashed polygon shows the strongest baseline per metric, and the solid indigo polygon shows Siam-ALS. The hatched green region marks where Siam-ALS exceeds the best-performing baseline; on AUROC, Siam-ALS ties the best baselines (A4, A12) Figure 6: t-SNE visualisation of embedding spaces on the EEGET-ALS dataset. Left: fused Siam-ALS embeddings learned via metric learning, showing clear separation between ALS and healthy subjects. Middle: EEGNet+ET embeddings with late fusion, where EEG and ET features are combined after separate feature extraction. Right: EEGNet+ET embeddings with mid fusion, where ET features are integrated at an intermediate network layer 4.6 Ablation Studies 4.6.1 Data Modality Ablation Study Table 6 isolates the contribution of each modality by holding the Siam-ALS architecture fixed and varying only the input streams. The EEG-only variant (B1) achieves 94.3% accuracy (AUROC 0.9843), confirming that cortical signals alone already carry substantial disease-discriminative information. In sharp contrast, the ET-only variant (B2) performs at chance level (accuracy 49.3%, AUROC 0.4932, F1=0.1345F_1=0.1345), demonstrating that gaze dynamics alone are entirely insufficient for ALS discrimination in this setting. When the two modalities are fused (B3), performance rises to 98.7% accuracy and AUROC 0.9985, an absolute AUROC gain of +0.0142 over EEG-only and +0.5053 over ET-only. This asymmetry is itself informative: although ET carries no standalone discriminative signal, adding it to EEG still improves AUROC by 0.014 and accuracy by 4.4 percentage points, suggesting that the fusion encoder exploits cross-modal interactions between cortical and oculomotor dynamics rather than merely defaulting to the stronger EEG stream. This finding directly motivates the joint EEG–ET explainability analysis, since it confirms that the fused embedding, not EEG alone, is the representation that must be explained to fully account for the model’s predictions. Table 6: Data modality ablation study for Siam-ALS, reported as mean ± standard deviation over 4 subject-level folds. Δ is the absolute AUROC gain over the weaker unimodal baseline (ET-only). B3 is architecturally identical to A13 (Table 5) and M4 (Table 7); B1 is comparable to M0 (Table 7); minor numerical differences reflect independent training runs (Section 4.1) ID Modality Used Model Variant ACC. AUROC F1F_1 Δ vs. B2 B1 EEG only Siam-ALS – EEG 0.943±0.00630.943± 0.0063 0.9843±0.00320.9843± 0.0032 0.9432±0.00620.9432± 0.0062 +0.4911+0.4911 B2 ET only Siam-ALS – ET 0.4932±0.00440.4932± 0.0044 0.4932±0.00500.4932± 0.0050 0.1345±0.07080.1345± 0.0708 Reference B3 EEG + ET Siam-ALS fusion 0.9872±0.00100.9872± 0.0010 0.9985±0.00010.9985± 0.0001 0.9841±0.00100.9841± 0.0010 +0.5053+0.5053 4.6.2 Architectural Ablation Study Table 7 isolates the contribution of each architectural component by comparing five Siamese variants (M0–M4) under an identical training protocol, progressively extending a base EEG-only Siamese network with an ET encoder, a dedicated fusion encoder, and an embedding-level logistic-regression classifier. The base model (M0, EEG-only, simple MLP head) already achieves a strong AUROC of 0.988, comparable to the EEG-only result in the modality ablation (B1, AUROC 0.9843 in Table 6); the small (<<0.01 AUROC) difference between M0 and B1 reflects independent training runs of an architecturally comparable EEG-only configuration (Section 4.1), and both confirm that the Siamese metric-learning formulation alone provides a competitive starting point. Adding depth without ET (M1) slightly reduces performance (AUROC 0.983), showing that naively deepening the head does not help and may overfit under limited subject counts. Activating only the ET encoder (M2) collapses performance to near-chance (AUROC 0.502), consistent with the ET-only result in the modality ablation (B2, AUROC 0.4932 in Table 6); the small (<<0.01 AUROC) difference between M2 and B2 reflects independent training runs of an architecturally comparable ET-only configuration (Section 4.1), and both results agree that gaze dynamics alone carry essentially no discriminative signal for ALS in this setting. Critically, concatenating EEG and ET embeddings without a dedicated fusion encoder (M3) only recovers EEG-only-level performance (AUROC 0.989), whereas the full model (M4), which adds both the fusion encoder and embedding-level classifier, reaches the best result (AUROC 0.998, F1=0.988F_1=0.988). The gain from M3 to M4 (AUROC 0.989→0.9980.989→ 0.998) reflects the combined effect of the fusion encoder and the embedding-level classifier; because both components are added together, this ablation cannot attribute the improvement to either one alone. It does show, however, that simple concatenation is insufficient to benefit from ET, and that the full configuration, which also yields the compact 64-dimensional embedding used in Section 4.9, performs best. In other words, the architectural choices that drive predictive performance are the same choices that make Siam-ALS explainable, not two separate design goals. Table 7: Architectural ablation on EEGET-ALS comparing five Siamese variants (M0–M4), progressively adding an ET encoder, a fusion encoder, and an embedding-level classifier to a base EEG-only network, trained under an identical protocol (Section 4.2). M4 is architecturally identical to A13 (Table 5) and B3 (Table 6); M2 is comparable to B2, and M0 to B1 (both Table 6); minor numerical differences reflect independent training runs (Section 4.1) ID Configuration AUROC F1F_1 M0 EEG-only, simple MLP head 0.988 0.956 M1 EEG-only, deeper head (no ET) 0.983 0.947 M2 ET-only encoder 0.502 0.142 M3 EEG+ET concatenation (no fusion encoder) 0.989 0.962 M4 EEG+ET + fusion encoder + emb. classifier (full) 0.998 0.988 4.7 Statistical Significance Testing To determine whether Siam-ALS’s AUROC improvement over each baseline is statistically meaningful rather than attributable to fold-level noise, we conduct paired t-tests comparing per-fold AUROC values between Siam-ALS and each of the seven baselines for which per-fold predictions were directly available: SVM–PSD (A1), SVM–PSD+ET (A2), SVM–IHAR (A3), CNN (A4), EEGNet (A10), EEGNet+ET Late Fusion (A11), and EEGNet+ET Middle Fusion (A12). For each baseline b, the paired difference Δf=fSiam-ALS−fb _f=AUROC^Siam-ALS_f-AUROC^b_f is computed across the F=4F=4 subject-level folds, and a two-tailed paired t-test is applied under the null hypothesis H0:[Δf]=0H_0:E[ _f]=0, with df=F−1=3df=F-1=3. Because a paired t-test with only four folds has limited statistical power and is sensitive to non-normality in such a small sample, we additionally report the non-parametric Wilcoxon signed-rank test as a robustness check; however, we note that with F=4F=4 paired observations the Wilcoxon test cannot achieve p<0.05p<0.05 even for a fully consistent effect (pmin=0.125p_ =0.125; 0.250.25 for A4, where n=3n=3), and is therefore reported for reference rather than as a primary significance criterion. Table 8 reports the resulting Δ and p-values for both tests, along with Bonferroni-corrected significance at α=0.05/7≈0.0071α=0.05/7≈ 0.0071 to account for multiple comparisons across the seven baselines. At the uncorrected level (p<0.05p<0.05), four of the seven baselines (A1, A2, A3, A10) show a significantly lower AUROC than Siam-ALS, whereas A11, CNN (A4), and EEGNet+ET Middle Fusion (A12) do not. However, none of these differences remains significant after Bonferroni correction (smallest p=0.0138>0.0071p=0.0138>0.0071), so the AUROC advantage of Siam-ALS should be regarded as suggestive rather than conclusive. Table 8: Paired t-test and Wilcoxon signed-rank test comparing Siam-ALS AUROC against seven baselines across F=4F=4 subject-level folds (df=3df=3 for the t-test). Δ represents the mean paired difference (Siam-ALS −- baseline). Sig. indicates p<0.05p<0.05 (uncorrected); no comparison remains significant after Bonferroni correction (α=0.05/7≈0.0071α=0.05/7≈ 0.0071). For CNN (A4), one fold had zero AUROC difference and was excluded by the Wilcoxon procedure, leaving n=3n=3 pairs. Baseline Δ t (p) Wilcoxon p Sig. SVM–Ihar (A3) 0.00658 5.21 (0.0138) 0.125 ✓ SVM–PSD (A1) 0.00624 4.85 (0.0169) 0.125 ✓ SVM–PSD + ET (A2) 0.00509 4.42 (0.0214) 0.125 ✓ EEGNet (A10) 0.00399 4.31 (0.0231) 0.125 ✓ EEGNet+ET Late Fusion (A11) 0.01014 2.98 (0.0587) 0.250 – CNN (A4) 0.00028 0.61 (0.5837) 0.750 – EEGNet+ET Middle Fusion (A12) 0.00088 0.71 (0.5290) 0.625 – Siam-ALS (A13) Reference – 4.8 Cross-Dataset Generalization on BigP3BCI dataset To assess whether the learned representation generalizes beyond a single acquisition setup, we evaluate EEG-only Siamese variants (G1–G3, Table 9) on the independent BigP3BCI corpus, which differs from EEGET-ALS in recording hardware, stimulus paradigm (P300 speller vs. gaze-controlled speller), and patient population. The identical EEG encoder architecture is used without modification (32 channels, matching the EEGET-ALS input); because BigP3BCI windows (124 samples at 124 Hz) are four samples shorter than the 128-sample input the encoder expects, each window is zero-padded to 128 samples prior to encoding. Trained from scratch on BigP3BCI, the base Siamese model (G1) reaches an AUROC of 0.954 and a window-level accuracy of 0.890, showing that the architecture remains effective on a second dataset with a different paradigm and population. Adding the embedding-level logistic-regression classifier (G2) raises AUROC to 0.983 and accuracy to 0.937, with sensitivity and specificity improving from 0.882/0.897 to 0.931/0.942. A deeper projection head (G3) yields only a marginal further gain (AUROC 0.984, accuracy 0.940, sensitivity 0.935, specificity 0.945). At the subject level, majority voting over each subject’s windows improves on window-level performance for G3, reaching a sensitivity of 0.958 and a specificity of 0.979. Fig. 4(c)–(d) shows the corresponding confusion matrices: window-level errors are roughly balanced between classes, with slightly more false negatives than false positives, and 46 of 48 ALS and 47 of 48 non-ALS subjects are correctly classified. The gap between AUROC (0.984) and accuracy (0.940) indicates that the fixed 0.5 decision threshold is not optimal on this dataset, so threshold-free AUROC is the more reliable summary of cross-dataset discrimination. Because the EEGET-ALS ablation adds the classifier jointly with the fusion encoder (M3 → M4), the G1 → G2 comparison is the only direct evidence in this study that embedding-level classification improves on the Siamese head alone. The effect of head depth is not consistent across datasets: it slightly reduced performance on EEGET-ALS (M0 → M1) and slightly improved it here. Since BigP3BCI is EEG-only, this experiment supports the EEG branch of the architecture but does not test the ET or fusion components. Table 9: Cross-dataset generalization results on the independent BigP3BCI corpus using EEG-only Siamese variants (G1–G3), evaluated with the identical EEG encoder architecture used on EEGET-ALS, without modification. Window-level metrics are means over 4 subject-level folds; subject-level metrics use majority voting over each subject’s windows (48 ALS / 48 non-ALS). G3 corresponds to Fig. 4(c)–(d) Window level Subject level ID Configuration AUROC Sens. Spec. Acc. Sens. Spec. G1 Base Siamese (EEG-only, no classifier head) 0.954 0.882 0.897 0.890 0.896 0.917 G2 + embedding-level logistic-regression classifier 0.983 0.931 0.942 0.937 0.938 0.979 G3 + deeper fusion/projection head 0.984 0.935 0.945 0.940 0.958 0.979 Model Convergence and Generalization. Fig. 7 presents the training and validation AUROC trajectories across all four cross-validation folds, confirming that the model converges within the first few epochs and maintains closely overlapping training/validation curves throughout, indicating stable learning without overfitting. Figure 7: Training and validation AUROC across epochs for all four cross-validation folds on EEGET-ALS. The model converges within the first few epochs in every fold and maintains consistently high performance (AUROC ≈0.99≈ 0.99), with closely overlapping training and validation curves, indicating stable learning and good generalization without obvious overfitting 4.9 Explainability Analysis As shown in Sections 4.5–4.7, all twelve baseline models spanning classical ML, DL, and multimodal fusion converge to near-ceiling AUROC (0.988–0.998), and none of the seven directly tested baselines differs significantly from Siam-ALS after multiple-comparison correction, with three of them (EEGNet+ET Late Fusion, CNN, and EEGNet+ET Middle Fusion) not significant even at the uncorrected level (Table 8). None of these twelve baselines, however, provide any mechanism to explain why a given sample is classified as ALS. This section demonstrates, through six independent attribution methods, including SHAP, LIME, Integrated Gradients, Saliency, Occlusion, and Permutation Importance, applied identically to the frozen embedding-level classifier, two complementary levels of explanation (embedding and signal), and a direct structural comparison against all twelve baselines, that the primary advantage of Siam-ALS is not marginal accuracy improvement but that its architecture is the only one in this comparison that supports rigorous, validated, and clinically verifiable explainability. Table 10 quantifies this directly, contrasting the exact Shapley attribution cost across representations and showing that only Siam-ALS’s compact 64-dimensional embedding admits tractable, exact attribution. Benchmarking Explainability Methods. To identify the most reliable explainability method for the fused embedding space, the results of XAI techniques are compared using fidelity as the evaluation metric. Fidelity was measured by the AUROC retained when a new classifier was trained using only the top 10 ranked embedding dimensions selected by each method from the original 64 dimensional embedding. Table 4.9 reports fold-averaged fidelity across all six XAI methods. The XAI comparison showed that all six methods achieved high fidelity, with AUROC values above 0.998, demonstrating that the fused embedding contains stable and interpretable information. SHAP achieved the highest mean fidelity (0.99847±0.000310.99847± 0.00031) using only the top 10 ranked embedding dimensions, although the differences between methods (≤8×10−5≤ 8× 10^-5) are far smaller than the fold-level standard deviations. SHAP was therefore at least as faithful as the alternatives and was selected as the primary method for interpreting the EEG- and ET-derived representations learned by Siam-ALS. Fig. 8 visualizes this comparison, confirming SHAP’s marginal fidelity advantage while showing all six methods achieve near-ceiling, tightly clustered performance. tableorg [t] Fidelity and cross-method agreement of six explainability methods applied to the Siam-ALS fused embedding. Top-10 AUROC reports mean ± std AUROC retained using only the top-10 ranked embedding dimensions (of 64 total) over 4 folds. Jaccard w/ SHAP reports top-10 dimensional overlap with SHAP, the reference method Method AUROC@10 SHAP Agreement SHAP 0.99847 ± 0.00031 1.000 Saliency 0.99846 ± 0.00039 0.111 Permutation Importance 0.99844 ± 0.00034 0.538 Integrated Gradients 0.99842 ± 0.00023 0.667 Occlusion 0.99842 ± 0.00024 0.429 LIME 0.99839 ± 0.00025 1.000 Table 10: Attribution tractability underlying the Explainable column of Table 5. Attr. Dim is the dimensionality attribution must operate over; Exact Shapley Cost is the asymptotic cost of closed-form Shapley attribution ((2d)O(2^d) in general, linear only for Siam-ALS’s compact embedding); Top-10 Fidelity is the AUROC retained using only the top-10 SHAP-ranked dimensions, reported only where attribution is exact and physiologically meaningful (Section 4.9) Representation Attr. Dim Exact Shapley Cost Top-10 Fidelity Raw EEG window (A4–A6, A9–A12) 4,096 (24096)O(2^4096) (infeasible) N/A Hand-crafted PSD/IHAR (A1–A3) 16 / 160–169 (216)O(2^16) / (2160−169)O(2^160-169)* N/A Tree-based feature importance (A7, A8) 4,096 (24096)O(2^4096) (infeasible) N/A Siam-ALS fused embedding (A13) 64 (d)O(d) (exact, linear) 0.99847 *Exact only for A3 (d=16d=16); infeasible for A1/A2 (d=160d=160–169169). Neither is learned/disease-specific. Figure 8: Fidelity comparison of six explainability methods applied to the Siam-ALS fused embedding (Table 4.9). (a) Mean Top-10 embedding-dimension AUROC per method across 4 folds (error bars: ± 1 SD); SHAP attains the highest mean, but all error bars overlap. (b) Rank-ordered fidelity gap relative to SHAP, measured as Δ @10 (×10−5× 10^-5), showing all six methods cluster within a narrow margin Table 11: Performance and explainability comparison across ALS-related and EEG/ET/Siamese methodology precedents. Values reflect each paper’s own validation protocol; protocols are listed since results are not directly comparable across rows. All figures are as reported in the original publications and were not independently reproduced or re-evaluated under a common protocol Year Method Feature Dataset AUROC ACC. (%) Protocol XAI 2017 GPFD+Fisher, KNN/LDA Liu et al. (2017) GPFD Private (5 ALS) – 88.3–95.3 LOO-CV Fisher ch. sel. (partial) 2021 Linear SVC/MLP Kocar et al. (2021) DTI (FA) + texture Private (n=502n=502) 0.87–0.91 – LOOCV (+holdout for MLP) Linear coef. (intrinsic) 2017 CNN + RL sampling Sengur et al. (2017) CWT/Spectrogram/SPWVD Public EMG – 96.80 Train/test split None 2022 EEG/ET fusion Vortmann et al. (2022) Time-series/MTF-GASF images Private (36 subj.) – 57.2–64.2 Person-dependent (5x random split) None 2023 Siamese AE/CNN You et al. (2023) Raw single-ch. EEG SleepEDF, MASS-S3 – 85.2/87.2 20/31-fold CV None 2023 SES+XGBoost/RF Founta et al. (2023) Gene expression (RNA-seq) Prudencio/Batra RNA-seq – 85.2/88.9 9-fold CV SHAP 2021 SNN + cosine dist. Efe and Özşen (2021) Raw single-ch. EEG Public EEG (Sleep-EDF) – 84.94 Hold-out None 2022 MB-EEGNet+CBAM Altuwaijri and Muhammad (2022) Raw EEG (multi-channel) BCI-IV2a, HGD – 82.85/95.45 Within-subj. CBAM attn. + t-SNE 2024 pcEEG+ (PCA+blink) Alyan et al. (2024) PCA-EEG components + synced blink Private (locomotor; loco.+driving) – 69–78 5-rep. 80/20 None 2025 DE+VAE+ET+CNN Le et al. (2025) DE + VAE latent + ET (fix./sacc.) Public EEG+ET, healthy subset (n=30) – 92.4/96.8 k-fold/cross-subj. None 2025 Optimized XGBoost Qin et al. (2025) Clinical/genetic (PRO-ACT vars.) PRO-ACT 0.955 97.96 10-fold CV / held-out SHAP + feat. importance 2025 Pseudo-Siamese Lin et al. (2025) Raw single-ch. EEG 4 public sleep datasets – MF1 84.5/89.6/89.5/85.4 k-fold (k=20) UMAP 2026 Gated MoE Chowdhury et al. (2026) Raw EMG + spectrogram + handcrafted feats. EMGLAB (N2001) 0.9992 99.05 Subj.-disjoint Gating weights 2024 SVM/CNN/CNN-LSTM Ngo et al. (2024) PSD/IHAR/Raw EEG EEGET-ALS (healthy only) – 88.2–94.5 Train/test split None 2024 SVM/EEGNet/EEG-ITNet Ngo et al. (2024) CSP-BP / EA-processed EEG EEGET-ALS (healthy) – – Subj.-held-out None 2026 Siam-ALS (Ours) EEG, ET EEGET-ALS 0.998 98.8 4-fold subj.-level SHAP (linear) + saliency Table 12: Explainability scorecard for the seven baselines with paired significance testing against Siam-ALS (Table 8). Δ and p-value are taken directly from Table 8; Explainable is obtained from Table 5. Models are grouped by uncorrected significance (p<0.05p<0.05); no comparison survives Bonferroni correction (α=0.0071α=0.0071). Across both tiers, explainability rather than accuracy alone is the distinguishing capability of Siam-ALS Model Δ p-value Sig. Explainable Tier 1: Lower AUROC than Siam-ALS (uncorrected p<0.05p<0.05) SVM–Ihar (A3) 0.006580 0.0138 ✓ ✗ SVM–PSD (A1) 0.006235 0.0169 ✓ ✗ SVM–PSD + ET (A2) 0.005087 0.0214 ✓ ✗ EEGNet (A10) 0.003988 0.0231 ✓ ✗ Tier 2: No significant difference (uncorrected p≥0.05p≥ 0.05) EEGNet+ET Late Fusion (A11) 0.010142 0.0587 – ✗ CNN (A4) 0.000278 0.5837 – ✗ EEGNet+ET Middle Fusion (A12) 0.000881 0.5290 – ✗ Siam-ALS (A13) Reference ✓ Table 11 presents a comprehensive comparison of ALS-related and EEG/ET/Siamese methods in terms of predictive performance, validation protocol, and explainability. Siam-ALS achieves an AUROC of 0.998 and an accuracy of 98.8% for ALS detection on EEGET-ALS. Prior work on this dataset, SVM/CNN/CNN-LSTM models by Ngo et al. Ngo et al. (2024) (88.2–94.5% accuracy), and on healthy-subject EEG–ET data, DE+VAE+ET+CNN Le et al. (2025) (92.4–96.8% accuracy), addressed different classification tasks on healthy participants rather than ALS versus control discrimination, so these figures provide context rather than a direct comparison. Some methods report comparable or slightly higher performance, such as Gated MoE Chowdhury et al. (2026), but these results were obtained using different modalities, datasets, and validation protocols and are therefore not directly comparable. A key distinction is explainability: most prior EEG/ET and Siamese approaches provide no or only partial interpretation, whereas Siam-ALS combines SHAP-based attribution with saliency analysis. Moreover, its Top-10 of 64 embedding dimensions retain an AUROC of 0.998, indicating that most of the discriminative information is concentrated in a compact learned representation. SHAP is selected as the primary explainability method for two additional reasons: (i) it computes exact Shapley values for the linear embedding-level classifier via LinearExplainer, avoiding the sampling approximations required by LIME or perturbation-based methods, and (i) it shows perfect top-10 dimensional agreement with LIME (Jaccard =1.0=1.0), providing cross-method validation from a second, independently-derived attribution technique. Notably, despite comparable fidelity, SHAP and Saliency show low dimensional overlap (Jaccard =0.11=0.11), indicating that the 64-dimensional fused embedding contains multiple redundant, equally discriminative substructures rather than a single dominant explanatory pathway. This redundancy is itself a desirable robustness property: it means the model’s decision is not contingent on any single fragile latent factor, and that multiple independent attribution methods can each recover a sufficient, if not identical, explanatory subset, all of which retain near-ceiling fidelity (Table 4.9). Embedding-Level SHAP Attribution. Fig. 9 presents SHAP dependence plots for the top-ranked fused embedding dimensions (e.g., E18, E31, E26). These plots reveal a small subset of highly discriminative latent dimensions whose variation shifts the classifier’s log-odds linearly and with large magnitude, confirming that the model relies on a compact, interpretable set of learned factors rather than diffusing decision weight uniformly across all 64 dimensions. Figure 9: SHAP dependence plots for the three top-ranked fused embedding dimensions (E18, E31, E26) on EEGET-ALS. The y-axis shows each dimension’s SHAP contribution to the classifier output in log-odds. Because the final classifier is linear, each relation is exactly linear, with slope equal to the dimension’s logistic-regression coefficient: E18 and E26 push predictions toward ALS, whereas E31 pushes them toward control. Point colour shows the embedding dimension with the strongest interaction, selected automatically by SHAP Feature-Level Saliency Attribution. To illustrate how fused embedding dimensions relate to the underlying physiological signals, we compute gradient-based saliency maps over the raw EEG input. Fig. 10 (left) shows a representative example for embedding dimension E52: saliency is concentrated on a subset of channels (notably Ch10, Ch12, and Ch24–Ch25) during specific temporal windows rather than being spread uniformly across the montage, indicating that individual embedding dimensions draw on spatially and temporally localized EEG activity rather than diffuse noise. Fig. 10 (right) complements this with the SHAP beeswarm of the 20 highest-ranked dimensions, led by E18, E31, and E26. Figure 10: Physiological and embedding-level attribution on EEGET-ALS. (Left) EEG channel×time saliency map (mean absolute gradient) for a representative fused embedding dimension (E52), illustrating how an individual embedding dimension draws on specific EEG channels and time windows. (Right) SHAP beeswarm plot of the 20 highest-ranked fused embedding dimensions; each point is one window, the x-axis shows its SHAP contribution to the log-odds output, and colour indicates the dimension’s value In summary, Siam-ALS showed a significantly higher AUROC than four of the seven directly tested baselines at the uncorrected level (p<0.05p<0.05), but none of these differences survived Bonferroni correction (Table 8). Since high predictive performance has become comparable among strong models, explainability becomes the key factor that differentiates Siam-ALS from existing approaches. Unlike other baselines, Siam-ALS provides consistent, faithful explanations that localize to specific embedding dimensions and EEG channels through multiple XAI methods (see Table 5), supporting its potential as a clinically interpretable ALS detection framework. Table 12 consolidates this comparison, showing that across both significance tiers, Siam-ALS is the only model marked explainable. 5 Conclusion This study introduced Siam-ALS, an explainable Siamese DL framework that jointly learns from EEG and ET signals for early ALS detection. By integrating complementary information from EEG-derived neural activity and ET-derived oculomotor behavior, Siam-ALS achieved strong performance on the EEGET-ALS dataset, reaching 98.8% accuracy and an AUROC of 0.998 under subject-level evaluation. The framework demonstrated competitive performance against a wide range of classical ML and DL models while providing interpretable insights into ALS-related patterns. Beyond accuracy, this study highlights the role of explainability in medical AI. The six XAI methods demonstrated that the learned 64-dimensional embedding captures meaningful ALS-related patterns, with SHAP achieving the highest, though statistically comparable, fidelity in explaining the EEG and ET features that influenced the model’s predictions. Ablation studies showed that multimodal fusion with a dedicated fusion encoder improves on EEG-only and concatenation variants, while cross-dataset evaluation on BigP3BCI demonstrated the potential of the EEG branch to generalize across different BCI settings. However, this study is limited by the small number of ALS participants in EEGET-ALS, which may affect statistical power and generalization. Future work will focus on larger multi-center datasets and additional BCI paradigms to further validate the framework. Overall, Siam-ALS provides an interpretable and multimodal approach toward reliable AI-based ALS detection. Statements and Declarations Funding This work is partly supported by the Natural Science Foundation of Shanghai, China (Grant No.22ZR1426200 and 17ZR1411900) and the National Natural Science Foundation of China (Grant No. 62103257). Competing Interests The authors declare that they have no competing interests. Ethics Approval This study used publicly available datasets and did not involve any new experiments with human participants. Therefore, ethical approval was not required. Data Availability The datasets used in this study can be accessed through their respective repositories. The EEGET-ALS dataset, described by Ngo et al. Ngo et al. (2024), is available in the Figshare repository at https://doi.org/10.6084/m9.figshare.c.6910027.v1. The BigP3BCI dataset Mainsah et al. (2025) is publicly available on PhysioNet at https://doi.org/10.13026/0byy-ry86. The source code will be made publicly available at https://github.com/marahman30104/Siam-ALS upon acceptance of the manuscript. References Al-Chalabi et al. (2016) A. Al-Chalabi, O. Hardiman, M. C. Kiernan, A. Chio, B. Rix-Brooks, and L. H. van den Berg Amyotrophic lateral sclerosis: moving towards a new classification system. The Lancet Neurology 15 (11), p. 1182–1194. External Links: ISSN 1474-4422, Document, Link Cited by: §1, §1. Alonso-Vázquez et al. (2023) D. Alonso-Vázquez, O. Mendoza-Montoya, R. Caraza, H. R. Martinez, and J. M. Antelis EEG-Based classification of spoken words using machine learning approaches. Computation 11 (11), p. 225. External Links: ISSN 2079-3197, Document, Link Cited by: §2. Altaheri et al. (2023) H. Altaheri, G. Muhammad, M. Alsulaiman, S. U. Amin, G. A. Altuwaijri, W. Abdul, M. A. Bencherif, and M. Faisal Deep learning techniques for classification of electroencephalogram (EEG) motor imagery (MI) signals: a review. Neural Computing and Applications 35 (20), p. 14681–14722. External Links: Document Cited by: §2.5. Altuwaijri and Muhammad (2022) G. A. Altuwaijri and G. Muhammad Electroencephalogram-based motor imagery signals classification using a multi-branch convolutional neural network model with attention blocks. Bioengineering 9 (7), p. 323. External Links: Document, Link Cited by: §1, Table 11. Alyan et al. (2024) E. Alyan, S. Arnau, J. E. Reiser, and E. Wascher Synchronization-based fusion of EEG and eye blink signals for enhanced decoding accuracy. Scientific Reports 14, p. 26918. External Links: Document, Link Cited by: Table 11. Band et al. (2023) S. S. Band, A. Yarahmadi, C. Hsu, et al. Application of explainable artificial intelligence in medical health: a systematic review of interpretability methods. Informatics in Medicine Unlocked 40, p. 101286. External Links: ISSN 2352-9148, Document, Link Cited by: §1. Bede and Chio (2022) P. Bede and A. Chio Machine learning applications in amyotrophic lateral sclerosis. Journal of Neurology, Neurosurgery & Psychiatry 93 (6), p. 575–576. External Links: ISSN 0022-3050, Document, Link Cited by: §1, §1, §1. Bin Heyat et al. (2024) M. B. Bin Heyat, F. Akhtar, F. Munir, A. Sultana, A. Y. Muaad, I. Gul, M. Sawan, W. Asghar, S. M. A. Iqbal, A. A. Baig, I. de la Torre Díez, and K. Wu Unravelling the complexities of depression with medical intelligence: exploring the interplay of genetics, hormones, and brain function. Complex & Intelligent Systems 10, p. 5883–5915. External Links: Document Cited by: §1. Calma et al. (2024) A. D. Calma, M. van den Bos, N. Pavey, C. Santos Silva, P. Menon, and S. Vucic Physiological biomarkers of upper motor neuron dysfunction in ALS. Brain Sciences 14 (8), p. 760. External Links: ISSN 2076-3425, Document, Link Cited by: §2.1. Chio et al. (2009) A. Chio, G. Logroscino, O. Hardiman, R. Swingler, D. Mitchell, E. Beghi, and B. J. Traynor Prognostic factors in ALS: a critical review. Amyotrophic Lateral Sclerosis 10 (5-6), p. 310–323. External Links: ISSN 1748-2968, Document, Link Cited by: §1. Chmiel and Stępień-Słodkowska (2025) J. Chmiel and M. Stępień-Słodkowska Resting-state EEG oscillations in amyotrophic lateral sclerosis (ALS): toward mechanistic insights and clinical markers. Journal of Clinical Medicine 14 (2), p. 545. External Links: ISSN 2077-0383, Document, Link Cited by: §2.1. Chowdhury et al. (2026) S. Chowdhury, S. Pramanik, S. Bhattacharjee, and M. Billah ALSGate: an efficient gated mixture-of-experts model for reliable ALS detection using EMG signals. Healthcare Technology Letters 13. External Links: Document, Link Cited by: §1, §4.9, Table 11. Dash et al. (2024) D. Dash, K. Teplansky, P. Ferrari, A. Babajani-Feremi, C. S. Calley, D. Heitzman, S. G. Austin, and J. Wang Automatic detection of ALS from single-trial MEG signals during speech tasks: a pilot study. Frontiers in Psychology 15, p. 1114811. External Links: ISSN 1664-1078, Document, Link Cited by: §2.1. Donaghy and Pioro (2024) R. Donaghy and E. P. Pioro Neurophysiologic innovations in ALS: enhancing diagnosis, monitoring, and treatment evaluation. Brain Sciences 14 (12), p. 1251. External Links: ISSN 2076-3425, Document, Link Cited by: §2.1. Dukic et al. (2022) S. Dukic, R. McMackin, T. Buxo, A. Fasano, R. Chipika, M. Pinto-Grau, E. Costello, C. Schuster, M. Hammond, M. Heverin, A. Coffey, M. Broderick, M. C. Kiernan, P. Bede, O. Hardiman, and B. Nasseroleslami Resting-state EEG reveals four subphenotypes of amyotrophic lateral sclerosis. Brain 145 (2), p. 621–636. External Links: ISSN 0006-8950, Document, Link Cited by: §1, §1, §2.1. Efe and Özşen (2021) E. Efe and S. Özşen A new approach for automatic sleep staging: Siamese neural networks. Traitement du Signal 38 (5), p. 1423–1430. External Links: Document, Link Cited by: Table 11. Fernandes et al. (2021) F. Fernandes, I. Barbalho, D. Barros, R. Valentim, C. Teixeira, J. Henriques, P. Gil, and M. Dourado Júnior Biomedical signals and machine learning in amyotrophic lateral sclerosis: a systematic review. BioMedical Engineering OnLine 20 (1), p. 61. External Links: Document, Link Cited by: §1. Founta et al. (2023) K. Founta, D. Dafou, E. Kanata, T. Sklaviadis, T. P. Zanos, A. Gounaris, and K. Xanthopoulos Gene targeting in amyotrophic lateral sclerosis using causality-based feature selection and machine learning. Molecular Medicine 29, p. 12. External Links: Document, Link Cited by: §1, Table 11. Fu et al. (2023) B. Fu, C. Gu, M. Fu, Y. Xia, and Y. Liu A novel feature fusion network for multimodal emotion recognition from EEG and eye movement signals. Frontiers in Neuroscience 17, p. 1234162. External Links: ISSN 1662-453X, Document, Link Cited by: §1, §2.2. Grollemund et al. (2019) V. Grollemund, P. Pradat, G. Querin, F. Delbot, G. Le Chat, J. Pradat-Peyre, and P. Bede Machine learning in amyotrophic lateral sclerosis: achievements, pitfalls, and future directions. Frontiers in Neuroscience 13, p. 135. External Links: ISSN 1662-453X, Document, Link Cited by: §1, §1, §1. Guttmann-Flury et al. (2025) E. Guttmann-Flury, X. Sheng, and X. Zhu Dataset combining EEG, eye-tracking, and high-speed video for ocular activity analysis across BCI paradigms. Scientific Data 12, p. 587. External Links: ISSN 2052-4463, Document, Link Cited by: §2.2. Hardiman et al. (2022) O. Hardiman, A. Al-Chalabi, A. Chio, E. M. Corr, G. Logroscino, W. Robberecht, P. J. Shaw, Z. Simmons, and L. H. van den Berg Amyotrophic lateral sclerosis. The Lancet 400 (10360), p. 1363–1380. External Links: ISSN 0140-6736, Document, Link Cited by: §1. Hwang et al. (2017) C. Hwang, S. Lee, and J. Lim Development of an electrooculogram-based eye-computer interface for communication of individuals with amyotrophic lateral sclerosis. Journal of NeuroEngineering and Rehabilitation 14, p. 89. External Links: ISSN 1743-0003, Document, Link Cited by: §1, §1. Iloon et al. (2022) T. Iloon, R. Barati, and H. Azad Siamese network-based feature transformation for improved automated epileptic seizure detection. Complexity 2022, p. 9161827. External Links: ISSN 1076-2787, Document, Link Cited by: §1. Iyer et al. (2015) P. M. Iyer, C. Egan, M. Pinto-Grau, T. Burke, M. Elamin, B. Nasseroleslami, N. Pender, E. C. Lalor, and O. Hardiman Functional connectivity changes in resting-state EEG as potential biomarker for amyotrophic lateral sclerosis. PLOS ONE 10 (6), p. e0128682. External Links: ISSN 1932-6203, Document, Link Cited by: §1, §1. Junqueira et al. (2024) B. Junqueira, B. Aristimunha, S. Chevallier, and R. Y. de Camargo A systematic evaluation of Euclidean alignment with deep learning for EEG decoding. Journal of Neural Engineering 21 (3), p. 036038. External Links: ISSN 1741-2552, Document, Link Cited by: §2.5. Kocar et al. (2021) T. D. Kocar, A. Behler, A. C. Ludolph, H. Müller, and J. Kassubek Multiparametric microstructural MRI and machine learning classification yields high diagnostic accuracy in amyotrophic lateral sclerosis: proof of concept. Frontiers in Neurology 12, p. 745475. External Links: Document, Link Cited by: §1, Table 11. Le et al. (2025) D. D. Le, H. D. Kieu, T. H. Le, et al. Attention detection: an EEG and eye tracking features fusion approach in eye-based interaction systems. Neural Computing and Applications 37, p. 12903–12919. External Links: ISSN 0941-0643, Document, Link Cited by: §1, §2, §4.9, Table 11. Lin et al. (2025) S. Lin, Z. Wang, H. van Gorp, M. Xu, M. van Gilst, S. Overeem, J. Linnartz, P. Fonseca, and X. Long SSC-SleepNet: a Siamese-based automatic sleep staging model with improved N1 sleep detection. IEEE Journal of Biomedical and Health Informatics 29 (9). External Links: Document, Link Cited by: §1, Table 11. Liu et al. (2017) Y. Liu, S. Huang, and Y. Huang Motor imagery EEG classification for patients with amyotrophic lateral sclerosis using fractal dimension and Fisher’s criterion-based channel selection. Sensors 17 (7), p. 1557. External Links: Document, Link Cited by: §1, Table 11. Liu et al. (2025) Y. Liu, L. An, H. Yang, and S. S. Ge Multi-frequency EEG and multi-functional connectivity graph convolutional network based detection method of patients with Alzheimer’s disease. Complex & Intelligent Systems 11, p. 366. External Links: Document Cited by: §2.5. Loh et al. (2022) H. W. Loh, C. P. Ooi, S. Seoni, P. D. Barua, F. Molinari, and U. R. Acharya Application of explainable artificial intelligence for healthcare: a systematic review of the last decade (2011–2022). Computer Methods and Programs in Biomedicine 226, p. 107161. External Links: ISSN 0169-2607, Document, Link Cited by: §1. Lundberg et al. (2020) S. M. Lundberg, G. Erion, H. Chen, A. DeGrave, J. M. Prutkin, B. Nair, R. Katz, J. Himmelfarb, N. Bansal, and S. Lee From local explanations to global understanding with explainable AI for trees. Nature Machine Intelligence 2 (1), p. 56–67. External Links: ISSN 2522-5839, Document, Link Cited by: §1, §2.4. Ma et al. (2026) R. Ma, B. Shi, M. Liu, and Y. Wang Dual-topology dynamic graph convolution-based spatial-spectral feature fusion network for patient-specific seizure detection. Complex & Intelligent Systems 12, p. 240. External Links: Document Cited by: §2.5. Mainsah et al. (2025) B. Mainsah, C. Fleeting, T. Balmat, E. Sellers, and L. Collins bigP3BCI: an open, diverse and machine learning ready P300-based brain-computer interface dataset (version 1.0.0). Note: PhysioNet External Links: Document Cited by: §2.5, §3.1.1, §4.2, Data Availability. Moreno Escobar et al. (2024) J. J. Moreno Escobar, O. Morales Matamoros, E. Y. Aguilar del Villar, H. Quintana Espinosa, and L. Chanona Hernández Employing Siamese networks as quantitative biomarker for assessing the effect of dolphin-assisted therapy on pediatric cerebral palsy. Brain Sciences 14 (8), p. 778. External Links: ISSN 2076-3425, Document, Link Cited by: §2.3. Munia et al. (2021) T. T. K. Munia, M. Nourani, and S. Houari Biometric authentication using EEG signals and few-shot Siamese networks. IEEE Journal of Biomedical and Health Informatics 25 (12), p. 4423–4432. External Links: ISSN 2168-2194, Document, Link Cited by: §2.3. Ngo et al. (2024) T. D. Ngo, H. D. Kieu, M. H. Nguyen, T. H. Nguyen, V. M. Can, B. H. Nguyen, and T. H. Le An EEG & eye-tracking dataset of ALS patients & healthy people during eye-tracking-based spelling system usage. Scientific Data 11, p. 664. External Links: ISSN 2052-4463, Document, Link Cited by: §1, §2.2, §3.1.1, §4.2, §4.9, Table 11, Table 11, Data Availability. Pancotti et al. (2022) C. Pancotti, G. Birolo, C. Rollo, T. Sanavia, E. Leonardi, and P. Fariselli Deep learning methods to predict amyotrophic lateral sclerosis disease progression. Scientific Reports 12, p. 13738. External Links: ISSN 2045-2322, Document, Link Cited by: §2.1. Peters et al. (2020) B. Peters, S. Bedrick, S. Dudy, B. Eddy, M. Higger, M. Kinsella, D. McLaughlin, T. Memmott, B. Oken, F. Quivira, S. Spaulding, D. Erdogmus, and M. Fried-Oken SSVEP BCI and eye tracking use by individuals with late-stage ALS and visual impairments. Frontiers in Human Neuroscience 14, p. 595890. External Links: ISSN 1662-5161, Document, Link Cited by: §1, §2.2. Pillalamarri and Shanmugam (2025) R. Pillalamarri and U. Shanmugam A review on EEG-based multimodal learning for emotion recognition. Artificial Intelligence Review 58, p. 131. External Links: ISSN 0269-2821, Document, Link Cited by: §1. Proudfoot et al. (2015) M. Proudfoot, R. A. L. Menke, R. Sharma, C. M. Berna, S. L. Hicks, C. Kennard, K. Talbot, and M. R. Turner Eye-tracking in amyotrophic lateral sclerosis: a longitudinal study of saccadic and cognitive tasks. Amyotrophic Lateral Sclerosis and Frontotemporal Degeneration 17 (1-2), p. 101–111. External Links: Document, Link Cited by: §1, §1. Qin et al. (2025) H. Qin, L. Hussain, Z. Liu, X. Yan, F. A. Awwad, F. M. Butt, U. A. Salaria, and E. A. A. Ismail Optimizing deep learning models to combat amyotrophic lateral sclerosis (ALS) disease progression. DIGITAL HEALTH 11, p. 1–26. External Links: Document, Link Cited by: §1, Table 11. Rehman et al. (2025) T. U. Rehman, M. Alruwaili, M. H. Siddiqi, Y. Alhwaiti, S. Anwar, Z. Halim, and M. Alam Advancing EEG-based biometric identification through multi-modal data fusion and deep learning techniques. Complex & Intelligent Systems 11, p. 398. External Links: Document Cited by: §2.3. Sengur et al. (2017) A. Sengur, Y. Akbulut, Y. Guo, and V. Bajaj Classification of amyotrophic lateral sclerosis disease based on convolutional neural network and reinforcement sample learning algorithm. Health Information Science and Systems 5 (1), p. 9. External Links: Document, Link Cited by: §1, Table 11. Spataro et al. (2014) R. Spataro, M. Ciriacono, C. Manno, and V. La Bella The eye-tracking computer device for communication in amyotrophic lateral sclerosis. Acta Neurologica Scandinavica 130 (1), p. 40–45. External Links: ISSN 0001-6314, Document, Link Cited by: §1, §1. Tjoa and Guan (2021) E. Tjoa and C. Guan A survey on explainable artificial intelligence (XAI): toward medical XAI. IEEE Transactions on Neural Networks and Learning Systems 32 (11), p. 4793–4813. External Links: ISSN 2162-237X, Document, Link Cited by: §1, §1. Turner and Al-Chalabi (2022) M. R. Turner and A. Al-Chalabi Diagnosing ALS: the Gold Coast criteria and the role of EMG. Practical Neurology 22 (3), p. 176–178. External Links: ISSN 1474-7758, Document, Link Cited by: §1. Usman et al. (2025) S. M. Usman, S. Khalid, A. Tanveer, A. S. Imran, and M. Zubair Multimodal consumer choice prediction using EEG signals and eye tracking. Frontiers in Computational Neuroscience 18, p. 1516440. External Links: ISSN 1662-5188, Document, Link Cited by: §2.2. Vortmann et al. (2022) L. Vortmann, S. Ceh, and F. Putze Multimodal EEG and eye tracking feature fusion approaches for attention classification in hybrid BCIs. Frontiers in Computer Science 4, p. 780580. External Links: ISSN 2624-9898, Document, Link Cited by: §1, §1, §2.2, Table 11. Wang et al. (2024) Y. Wang, Y. Zhou, W. Lu, Q. Wu, Q. Li, and R. Zhang AC-CfC: an attention-based convolutional closed-form continuous-time neural network for raw multi-channel EEG-based emotion recognition. Biomedical Signal Processing and Control 94, p. 106249. External Links: Document Cited by: §2.5. Xu et al. (2025) Y. Xu, Y. Gao, Z. Zhang, and S. Du Emotional recognition of EEG signals utilizing residual structure fusion in bi-directional LSTM. Complex & Intelligent Systems 11, p. 62. External Links: Document Cited by: §2.5. Yang and Liu (2024) G. Yang and J. Liu A novel multi-scale fusion convolutional neural network for EEG-based motor imagery classification. Biomedical Signal Processing and Control 96, p. 106645. External Links: Document, Link Cited by: §2.5. You et al. (2023) Y. You, X. Guo, Z. Yang, and W. Shan A Siamese network-based method for improving the performance of sleep staging with single-channel EEG. Biomedicines 11 (2), p. 327. External Links: Document, Link Cited by: §1, §1, Table 11. Yue et al. (2022) K. Yue, M. Guo, Y. Liu, H. Hu, K. Lu, S. Chen, and D. Wang Investigate the neuro mechanisms of stereoscopic visual fatigue. IEEE Journal of Biomedical and Health Informatics 26 (7), p. 2963–2973. External Links: ISSN 2168-2194, Document, Link Cited by: §2.2. Zhou et al. (2025) X. Zhou, C. C. Menassa, and V. R. Kamat Siamese network with dual attention for EEG-Driven social learning: bridging the Human-Robot gap in long-tail autonomous driving. Expert Systems with Applications 291, p. 128470. External Links: ISSN 0957-4174, Document, Link Cited by: §2.3. Zhu et al. (2025) F. Zhu, J. Zhang, R. Dang, B. Hu, and Q. Wang MTNet: multimodal transformer network for mild depression detection through fusion of EEG and eye tracking. Biomedical Signal Processing and Control 100, p. 106996. External Links: ISSN 1746-8094, Document, Link Cited by: §1.