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

Area of Science:

• Neuroscience
• Computational Biology
• Medical Informatics

Background:

• Amyotrophic lateral sclerosis (ALS) presents significant challenges for treatment development due to its complex progression.
• Accurate prediction of ALS progression is critical for advancing targeted therapies and improving patient management.

Purpose of the Study:

• To investigate the efficacy of deep learning and machine learning models for predicting ALS progression.
• To evaluate and compare the performance of XGBoost, LightGBM, and deep learning models using the PRO-ACT dataset.

Main Methods:

• Evaluated XGBoost, LightGBM, and a deep learning sequential model with default parameters on the PRO-ACT dataset.
• Utilized R-squared (R2) and Root Mean Squared Error (RMSE) for performance evaluation.
• Performed hyperparameter optimization to enhance model predictive accuracy and classification capabilities.

Main Results:

• The deep learning model initially showed superior predictive performance (RMSE: 4.565, R2: 0.716).
• Hyperparameter optimization improved the deep learning model (RMSE: 4.511, R2: 0.718) and XGBoost (RMSE: 4.532, R2: 0.715).
• Optimized XGBoost achieved high classification performance for bulbar vs. limb onset ALS (AUC: 0.9550) and identified key predictive features like ZBTB2P1.

Conclusions:

• Optimized deep learning and XGBoost models demonstrate significant potential for ALS progression prediction and classification.
• These predictive models can facilitate early risk stratification, personalized treatment, and improved clinical decision-making for ALS patients.
• The findings suggest a pathway to enhanced prognostic communication and potentially reduced ALS-related mortality.