Does including machine learning predictions in ALS clinical trial analysis improve statistical power?

Area of Science:

• Neuroscience
• Biostatistics
• Clinical Trials

Background:

• Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disease.
• Predictive modeling for ALS progression is of growing interest due to large clinical trial datasets.
• The utility of predictive models in ALS clinical trial analysis requires thorough evaluation.

Purpose of the Study:

• To evaluate a predictive modeling approach for ALS disease progression using ALSFRS-R.
• To validate these models in an independent clinical trial dataset.
• To examine how predictive models can enhance statistical power in ALS clinical trial simulations.

Main Methods:

• Utilized the PRO-ACT database for model development and a separate clinical trial dataset for validation.
• Employed machine learning models, including super learner and random forest, alongside linear mixed-effects models.
• Conducted simulation studies to assess the impact of incorporating model predictions into clinical trial analysis.

Main Results:

• Models using imputed data demonstrated superior external validation compared to those with complete observations.
• Super learner (R²=0.71) and random forest (R²=0.70) models showed comparable performance, outperforming linear mixed-effects models (R²=0.69).
• Simulations indicated that including machine learning predictions as a covariate could increase effective sample size by 16% in a 12-month study.

Conclusions:

• Predictive modeling for ALSFRS-R explains moderate longitudinal variability, enhanced by effective missing data handling.
• Incorporating post-baseline ALSFRS-R predictions as a covariate may boost statistical power in clinical trials with moderate treatment effects.