Predicting Amyotrophic Lateral Sclerosis Mortality With Machine Learning in Diverse Patient Databases

Area of Science:

• Neurology
• Biostatistics
• Machine Learning

Background:

• Predicting mortality in Amyotrophic Lateral Sclerosis (ALS) is crucial for patient care and clinical trial design.
• Existing models have limitations, including reliance on early data, fixed predictor relationships, and lack of diverse population validation.

Purpose of the Study:

• Develop and validate novel ALS mortality prediction models using routinely available clinical data.
• Address limitations of existing models by incorporating data from any clinical visit and validating on diverse populations.

Main Methods:

• Trained Royston-Parmar and eXtreme Gradient Boosting models on the PRO-ACT database for 6- and 12-month mortality prediction.
• Validated models on independent datasets from North American and Singaporean ALS populations.
• Evaluated feature importance and the impact of predictor reduction.

Main Results:

• Models achieved high predictive performance (AUC 0.768-0.865) using data from any clinical visit.
• Albumin emerged as the top predictor, followed by ALS Functional Rating Scale-Revised slope, limb onset, and other clinical variables.
• Models demonstrated robust performance on independent datasets and when reduced to seven key predictors.

Conclusions:

• Visit-agnostic ALS mortality prediction models were developed and validated across diverse populations.
• Identified key prognostic features, including albumin and functional decline.
• These models offer potential to enhance ALS patient care and optimize clinical trial design.