Predicting Amyotrophic Lateral Sclerosis Progression Using Optimized Deep Learning Architectures

Abstract Amyotrophic Lateral Sclerosis (ALS) is a progressive and fatal neurodegenerative disorder characterized by the deterioration of motor neurons, leading to muscle weakness, paralysis, and ultimately respiratory failure. Accurate prediction of ALS progression is critical for timely clinical decision-making, personalized patient management, and optimizing clinical trials. However, the heterogeneity of disease trajectories and limited temporal clinical observations pose significant challenges in building robust predictive models. To address these limitations, this study proposes a deep learning-based framework utilizing both Deep Neural Networks (DNN) and Long Short-Term Memory (LSTM) models to forecast ALS progression using the PRO-ACT dataset. The proposed architecture incorporates a multi-stage optimization framework designed to enhance model accuracy and generalization. Two nature-inspired metaheuristic algorithms—the Snake Optimizer and the Dung Beetle Optimizer—are independently applied to fine-tune key hyperparameters. Then, a novel hybrid optimizer is introduced, which sequentially combines the global search capability of the Snake algorithm with the local refinement strengths of the Dung Beetle algorithm. In parallel, Optuna—a Bayesian optimization framework—is employed as a baseline for automated hyperparameter tuning. The model also leverages Pearson correlation analysis for feature selection and standardizes inputs to ensure consistency and convergence. Experimental results show the LSTM model optimized with the Snake algorithm achieved a minimum validation Mean Squared Error (MSE) of 0.3146. The hybrid optimizer demonstrated robust convergence with an MSE of 0.3146 for LSTM and 0.3157 for DNN. Optuna-based models achieved MSEs of 0.3198 for LSTM and 0.3161 for DNN. These results highlight the effectiveness of AI-driven hybrid optimization in predicting ALS progression and supporting clinical intervention strategies.