Predicting amyotrophic lateral sclerosis (ALS) progression with machine learning.

Area of Science:

• Neurology
• Biostatistics
• Computational Biology

Background:

• Amyotrophic Lateral Sclerosis (ALS) is a progressive neurodegenerative disease.
• Predicting ALS progression is crucial for clinical trial design and patient management.
• Current prediction methods may not fully leverage the potential of machine learning.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting ALS progression.
• To assess the impact of varying observation and prediction window lengths on model performance.
• To identify key predictors of ALS disease progression.

Main Methods:

• Utilized demographic, clinical, and laboratory data from 5030 patients in the PRO-ACT database.
• Modeled ALS progression (fast vs. non-fast) using Extreme Gradient Boosting (XGBoost) and Bayesian Long Short Term Memory (BLSTM).
• Evaluated model performance using AUROC and compared across different observation lengths (single visit to 12 months).

Main Results:

• ML models achieved AUROC of 0.570-0.748, performing comparably to clinician assessments.
• Model performance was consistent across observation lengths but improved with longer prediction windows.
• Identified 21 predictors of progression, including disease onset, ALSFRS-R, forced vital capacity, phosphorus, chloride, and albumin.
• BLSTM models showed higher confidence in predictions for certain samples.
• Patient screening via models could hypothetically reduce clinical trial sizes by 18.3%.

Conclusions:

• Machine learning models offer a viable approach for predicting ALS progression.
• Clinical trial observation periods could potentially be shortened to a single visit, reducing trial sizes.
• Identified predictors may serve as novel biomarkers and therapeutic targets for ALS.
• BLSTM confidence levels enhance the trustworthiness of predictions for clinical decision-making.