Optimizing deep learning models to combat amyotrophic lateral sclerosis (ALS) disease progression

Abstract

Objective: Amyotrophic lateral sclerosis (ALS), a devastating neurodegenerative disease, poses a significant challenge for targeted treatment development. Accurate prediction of its progression is crucial for this endeavor. Methods: This study investigated deep learning methods for ALS progression prediction using the publicly available PRO-ACT dataset. Initially, machine learning models (XGBoost, LightGBM) and a deep learning sequential model were evaluated with default parameters, using R-squared (R2) and Root Mean Squared Error (RMSE) as performance metrics. Results: Notably, the deep learning model demonstrated superior predictive performance with default settings (RMSE: 4.565, R2: 0.716), followed by XGBoost (RMSE: 4.625, R2: 0.709) and LightGBM (RMSE: 4.596, R2: 0.716). Subsequently, hyperparameter optimization significantly enhanced the deep learning model's performance, achieving the highest prediction accuracy (RMSE: 4.511, R2: 0.718). Slight improvements were also observed for XGBoost (RMSE: 4.532, R2: 0.715) and LightGBM (RMSE: 4.551, R2: 0.716). Furthermore, the optimized XGBoost model demonstrated exceptional classification performance in distinguishing between bulbar and limb onset ALS, with a sensitivity of 100%, specificity of 97.44%, accuracy of 97.96%, F1-score of 95.96%, Matthews Correlation Coefficient (MCC) of 94.12%, and an Area Under the Curve (AUC) of 0.9550. Feature importance analysis with optimized XGBoost identified ZBTB2P1 as the most influential feature, followed by RNF181, with WASH9P being the least influential among the top eight. Conclusions: These findings convincingly demonstrate the potential of optimized XGBoost and deep learning for ALS progression prediction and classification, particularly with optimized parameters. This approach offers significant potential for early risk stratification, personalized treatment planning, enhanced prognostic communication, diagnostic support, streamlined disease monitoring, and improved clinical decision-making, ultimately contributing to better patient outcomes and potentially reducing ALS-related mortality.