Siam-ALS: An Explainable Siamese Deep Learning Framework for Early ALS Detection Using EEG and Eye-Tracking Signals

Early detection potential with high accuracy: Siam-ALS achieved 98.8% accuracy and 99.8% AUROC on the primary dataset, suggesting the framework could help identify ALS cases earlier than current clinical methods that rely on subjective assessment and heterogeneous symptoms.

Explainability built into the design: The framework uses modality-specific encoders for EEG and eye-tracking signals separately before combining them, which allows researchers and clinicians to understand which physiological signals (brain activity vs. eye movement) are driving the model's predictions—addressing a key concern about "black box" AI in healthcare.

Multimodal approach captures complementary information: By jointly analyzing both brain activity (EEG) and eye control (eye-tracking), the model learns from multiple biological systems affected by ALS, rather than relying on a single signal, which may improve reliability and reduce false positives.

Generalization across different datasets: The model maintained strong performance (approximately 98% AUROC) when tested on an independent dataset (BigP3BCI), indicating the approach isn't just memorizing patterns from one study but learning generalizable ALS biomarkers—important for real-world clinical deployment.

Designed for small, diverse datasets: The use of Siamese learning and metric learning helps the model work effectively with limited ALS patient data, which is realistic given ALS's rarity, making this approach more practical for rare disease diagnosis than methods requiring massive datasets.