Advancing Clinical Trials and Precision Medicine in Duchenne Muscular Dystrophy Using Machine Learning-Based Trajectory Modeling and MRI Radiomics.
Authors
Affiliations (3)
Affiliations (3)
- Department of Pharmaceutics, Center for Pharmacometrics and Systems Pharmacology, Intelligent Critical Care Center, College of Pharmacy, University of Florida, Orlando, Florida, USA.
- Department of Physical Therapy, University of Florida, Gainesville, Florida, USA.
- Department of Physiology and Aging, University of Florida, Gainesville, Florida, USA.
Abstract
Given heterogeneous progression of Duchenne muscular dystrophy (DMD), accurately characterizing individual disease trajectories is critical for effective clinical trials and therapeutic development. In this study, we applied machine learning (ML) approaches leveraging disease progression models and baseline magnetic resonance (MR)-derived radiomic features to quantify individual-level DMD progression trajectories and support quantitative characterization of disease progression. Radiomic features were extracted from segmented soleus (SOL) and vastus lateralis (VL) fat-fraction maps, generating 810 features per individual, with additional functional and demographic measures incorporated as covariates following feature selection. Multiple ML regressors were trained and validated to estimate MR-spectroscopy fat-fraction trajectory descriptive parameters: DPT<sub>50</sub> (age at half-maximal increase), DP<sub>max</sub> (extent), γ (rate), S<sub>0</sub> (extrapolated measure at age 0). Key radiomic features were identified through feature importance analyses, followed by downstream analyses, including unsupervised clustering. The tree-based XGBoost regression model demonstrated best future longitudinal fat-fraction trajectories (SOL: mean absolute error (MAE) = 0.04, R<sup>2</sup> = 0.94; VL: MAE = 0.06, R<sup>2</sup> = 0.95). Subsequent downstream analyses highlighted potential key biomarkers clinically associated with DMD progression. Clustering analysis revealed clinically meaningful slow- and fast-progression subgroups with interpretable decision criteria (SOL: DPT<sub>50</sub> = 12.21 vs. 14.83; γ = 8.07 vs. 7.54, VL: 10.37 vs. 14.21; γ = 7.72 vs. 7.09) that can inform clinical trial designs and therapeutic effects within homogenous groups. Individual-level feature importance interpretation further shows the possibility of tailoring personalized intervention strategies. These findings highlight the potential of ML-based trajectory modeling approaches leveraging baseline MRI radiomics for advancing AI-assisted clinical trial development and precision medicine, charting a new direction for MRI-guided disease modeling in neuromuscular disorders.