A novel machine learning based framework for developing composite digital biomarkers of disease progression.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Current methods for tracking neurodegenerative disease progression, like Parkinson's disease (PD), rely on subjective clinical scales.
• These scales lack the sensitivity and timeliness needed for early detection of progression.
• Digital health technologies (DHTs) offer objective, precise, and sensitive measures but present challenges in data complexity and feature selection.

Purpose of the Study:

• To develop and validate a machine learning framework for constructing composite digital biomarkers from DHT data.
• To improve the accuracy and sensitivity of tracking disease progression in neurodegenerative disorders.

Main Methods:

• A machine learning framework involving univariate digital feature screening, association testing, and feature selection.
• Construction of composite digital biomarkers using Penalized Generalized Estimating Equations (PGEE).
• Application to longitudinal data from a Parkinson's disease study, including sensor-based movement and clinical scores.

Main Results:

• Out of 235 digital features, 77 passed screening, and 11 were selected by PGEE for the composite measure.
• The composite digital measure showed a smoother, more significant trend over time with less variability than the traditional MDS-UPDRS Part III score.
• The composite measure effectively differentiated between early Parkinson's disease and healthy controls.

Conclusions:

• DHT-derived measures can enhance the tracking of neurodegenerative disease progression with greater sensitivity and reduced variability.
• The presented framework provides a novel methodology for creating composite digital biomarkers from high-dimensional DHT datasets.
• This approach may accelerate the development and application of digital biomarkers in clinical trials and drug development.