Objective: This study aims to develop a data-driven methodology for stratifying Amyotrophic Lateral Sclerosis (ALS) patients based on longitudinal disease progression patterns, using a novel deep learning framework that combines a Recurrent Variational Autoencoder (RVA) with consensus clustering to identify clinically meaningful subgroups. Methods: The RVA integrates Peephole Long Short-Term Memory networks within the Variational Deep Embedding (VaDE) architecture to simultaneously learn latent representations and cluster assignments from multivariate time-series data. The approach incorporates hyperparameter optimization via prediction strength with two-fold cross-validation, consensus clustering, and internal validation metrics (Silhouette Coefficient, Davies–Bouldin index, Calinski–Harabasz index) for optimal cluster selection. The methodology was validated on simulated data and applied to 3076 ALS patients from the PRO-ACT dataset, using ALSFRS-R total scores, domain subscores, and MiToS staging from the first six months of observation. Results: Simulation experiments demonstrated that consensus clustering consistently outperformed single-model predictions across all noise levels. Applied to the PRO-ACT real data, the framework identified five distinct patient subgroups. These clusters exhibited distinct progression patterns and statistically significant differences in baseline clinical features, disease onset characteristics, and survival outcomes, with median survival ranging from 12.8 months to 27.5 months. Conclusion: The proposed deep learning framework effectively captures the heterogeneous nature of ALS progression and identifies clinically relevant patient subgroups using routine clinical assessments. The stratification provides a foundation for personalized prognosis, optimized clinical trial design, and tailored therapeutic strategies, representing a practical tool for improving ALS patient management.
Bajolin et al. (2026) studied this question.