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

cs.LGOct 11, 2026

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

Authors: Abdur Rahman, A. A. M. Muzahid, Istiak Ahmed, Dipto Biswas, Hua Han, Yujin Zhang, Ferdous Sohel

Organizations: School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China. · School of Information Technology, Murdoch University, Murdoch, 6150, WA, Australia. · Department of Information and Communication Engineering, Daffodil International University, Dhaka, 1216, Bangladesh.

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.

Figures & tables

Dataset	Modalities	Classes	Usage
EEGET-ALS	EEG + ET	ALS vs Healthy	Primary dataset
BigP3BCI	EEG only	ALS vs Healthy	Generalization

Table 1: Summary of datasets used in this study

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

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

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 $\mathbf{z}_{i}$ and $\mathbf{z}_{j}$ are compared via an $\ell_{1}$ 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)

Block	Layer	Configuration	Input $\rightarrow$ Output	Params
EEG encoder $f_{\text{EEG}}(\mathbf{e}_{i})$	Conv2D-1	$1\times 15$ , 16 filters, padding $(0,7)$	$(1,32,128)\rightarrow(16,32,128)$	256
	BN-1	BatchNorm (16 ch.)	$(16,32,128)\rightarrow(16,32,128)$	32
	DepthwiseConv	$8\times 1$ , 32 filters, groups $=16$	$(16,32,128)\rightarrow(32,25,128)$	288
	BN-2	BatchNorm (32 ch.)	$(32,25,128)\rightarrow(32,25,128)$	64
	AvgPool-1	$1\times 4$	$(32,25,128)\rightarrow(32,25,32)$	–
	Conv2D-2	$1\times 15$ , 64 filters, padding $(0,7)$	$(32,25,32)\rightarrow(64,25,32)$	30,784

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 :

\mathbf{e}_{i}

(EEG window),

\mathbf{g}_{i}

(ET features),

\mathbf{z}_{i}

(fused embedding)

Held-Out Subject	AUROC	Accuracy	$F_{1}$
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

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

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)

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	–

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<0.05

) 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

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

ID	Modality Used	Model Variant	ACC.	AUROC	$F_{1}$	$\Delta$ AUROC vs. B2
B1	EEG only	Siam-ALS – EEG	$0.943\pm 0.0063$	$0.9843\pm 0.0032$	$0.9432\pm 0.0062$	$+0.4911$
B2	ET only	Siam-ALS – ET	$0.4932\pm 0.0044$	$0.4932\pm 0.0050$	$0.1345\pm 0.0708$	Reference
B3	EEG + ET	Siam-ALS fusion	$\mathbf{0.9872\pm 0.0010}$	$\mathbf{0.9985\pm 0.0001}$	$\mathbf{0.9841\pm 0.0010}$	$\mathbf{+0.5053}$

Table 6: Data modality ablation study for Siam-ALS, reported as mean

\pm

standard deviation over 4 subject-level folds.

\Delta

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	Configuration	AUROC	$F_{1}$
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

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 )

Baseline	$\Delta$ AUROC	$t$ ( $p$ )	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	–

Table 8: Paired

t

-test and Wilcoxon signed-rank test comparing Siam-ALS AUROC against seven baselines across

F=4

subject-level folds (

\mathrm{df}=3

for the

t

-test).

\Delta

AUROC represents the mean paired difference (Siam-ALS

-

baseline). Sig. indicates

p<0.05

(uncorrected); no comparison remains significant after Bonferroni correction (

\alpha=0.05/7\approx 0.0071

). For CNN (A4), one fold had zero AUROC difference and was excluded by the Wilcoxon procedure, leaving

n=3

pairs.

ID	Configuration	AUROC	Sens.	Spec.	Acc.	Sens.	Spec.
		Window level				Subject level
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

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)

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 $\approx 0.99$ ), with closely overlapping training and validation curves, indicating stable learning and good generalization without obvious overfitting

* Exact only for A3 ( $d{=}16$ ); infeasible for A1/A2 ( $d{=}160$ – $169$ ). Neither is learned/disease-specific.
Representation	Attr. Dim	Exact Shapley Cost	Top-10 Fidelity
Raw EEG window (A4–A6, A9–A12)	4,096	$\mathcal{O}(2^{4096})$ (infeasible)	N/A
Hand-crafted PSD/IHAR (A1–A3)	16 / 160–169	$\mathcal{O}(2^{16})$ / $\mathcal{O}(2^{160-169})$ *	N/A
Tree-based feature importance (A7, A8)	4,096	$\mathcal{O}(2^{4096})$ (infeasible)	N/A
Siam-ALS fused embedding (A13)	64	$\mathcal{O}(d)$ (exact, linear)	0.99847

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 (

\mathcal{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 )

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:

\pm

1 SD); SHAP attains the highest mean, but all error bars overlap. (b) Rank-ordered fidelity gap relative to SHAP, measured as

\Delta

AUROC@10 (

\times 10^{-5}

), showing all six methods cluster within a narrow margin

Tier 1: Lower AUROC than Siam-ALS (uncorrected $p<0.05$ )
Model	$\Delta$ AUROC	$p$ -value	Sig.	Explainable
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 $p\geq 0.05$ )

Table 12: Explainability scorecard for the seven baselines with paired significance testing against Siam-ALS (Table 8 ).

\Delta

AUROC and

p

-value are taken directly from Table 8 ; Explainable is obtained from Table 5 . Models are grouped by uncorrected significance (

p<0.05

); no comparison survives Bonferroni correction (

\alpha=0.0071

). Across both tiers, explainability rather than accuracy alone is the distinguishing capability of Siam-ALS

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

Figure 10: Physiological and embedding-level attribution on EEGET-ALS. (Left) EEG channel $\times$ 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

Explore similar work

AlignUS: MRI-Guided Ultrasound Representation Learning for ALS Classification from Tongue Images
Sep 14, 2026Kadija Abdel Ghader, Emani Babe, Lorenzo Pettinari +2Cross-Modal Knowledge Distillation Ultrasound Imaging
Unité de Recherche en Systèmes Intelligents Avancés (URSIA), Institut Supérieur du Numérique (SupNum), Nouakchott, Mauritanie · University of Basel, Basel, Switzerland
Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease
Aug 27, 2026Maggie Lin, Chung-Lin Hou, Tzyy-Ping JungAlzheimer's Disease Classification Electroencephalography
Department of Bioengineering University of California, San Diego La Jolla, CA 92037, USA · Research and Development HippoScreen Neurotech Corp. Taipei, Taiwan · Swartz Center for Computational Neuroscience University of California, San Diego La Jolla, CA 92037, USA
DeepTokenEEG Enhancing Mild Cognitive Impairment and Alzheimers Classification via Tokenized EEG Features
May 14, 2026Thinh Nguyen-Quang, Minh Long Ngo, Ngoc-Son Nguyen +7Electroencephalography Alzheimer's Disease Classification
Hanoi University of Science and Technology, Hanoi, Vietnam · VNU University of Engineering and Technology, Hanoi, Vietnam · Central Military Hospital 108, Hanoi, Vietnam +1

Sep 14, 2026J/K move · Enter open · S save

AlignUS: MRI-Guided Ultrasound Representation Learning for ALS Classification from Tongue Images

Kadija Abdel Ghader, Emani Babe, Lorenzo Pettinari, Meya Haroune+1

Unité de Recherche en Systèmes Intelligents Avancés (URSIA), Institut Supérieur du Numérique (SupNum), Nouakchott, Mauritanie · University of Basel, Basel, Switzerland

Cross-Modal Knowledge Distillation Ultrasound Imaging Medical Image Representation Learning

Open paper AskPDF arXiv

Amyotrophic lateral sclerosis (ALS) is a progressive neurodegenerative disease in which early assessment remains challenging, particularly in low-resource settings where MRI is often unavailable. High-resolution ultrasound (HRUS) of the tongue offers a portable and low-cost alternative for evaluating bulbar involvement, but learning reliable diagnostic models is limited by small datasets and the difficulty of extracting robust representations from ultrasound alone. We propose AlignUS, a cross-modal knowledge distillation framework that transfers anatomical knowledge from MRI to a HRUS-based classifier while requiring only HRUS at inference time. The model combines classification loss, supervised contrastive learning, and feature-level distillation to align HRUS representations with MRI embeddings. AlignUS achieves a patient-level balanced accuracy of 0.958, macro-F1 of 0.963, and ROC-AUC of 0.990, aggregated across four patient-level cross-validation folds, with consistent improvements over HRUS baselines and cross-modal alternatives. These results demonstrate that MRI-derived supervision can substantially improve ultrasound-based ALS assessment while preserving low-cost, inference-time independence from MRI.

Abstract

Explore similar work

AlignUS: MRI-Guided Ultrasound Representation Learning for ALS Classification from Tongue Images

Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease

DeepTokenEEG Enhancing Mild Cognitive Impairment and Alzheimers Classification via Tokenized EEG Features