Multi-scale machine learning model predicts muscle and functional disease progression in FSHD

Area of Science:

• Biomedical Engineering
• Neurology
• Machine Learning

Background:

• Facioscapulohumeral muscular dystrophy (FSHD) is a heterogeneous genetic neuromuscular disorder with variable progression.
• Current clinical metrics lack sensitivity for personalized assessment, hindering clinical trial design.

Purpose of the Study:

• To develop a multi-scale machine learning framework for predicting regional, muscle, joint, and functional progression in FSHD.
• To create a 'digital twin' for individual FSHD patients to enhance clinical trials.

Main Methods:

• Integrated whole-body MRI-derived metrics (fat fraction, lean muscle volume, fat heterogeneity) with clinical and functional data from over 100 patients.
• Developed a three-stage random forest model to predict annualized changes in muscle composition and the Timed Up-and-Go (TUG) test.

Main Results:

• Models demonstrated strong predictive performance on holdout datasets, predicting fat fraction change (RMSE 2.16%) and lean volume change (RMSE 8.1ml).
• Fat heterogeneity within muscle was identified as a predictor of muscle-level progression.
• The final model predicted TUG test changes with an RMSE of 0.6 seconds.

Conclusions:

• Machine learning models integrating muscle and performance data can effectively predict MRI-based disease progression and functional outcomes in FSHD.
• This approach addresses the inherent heterogeneity and nonlinearity of FSHD, with potential broad applicability to other neuromuscular diseases.