Advancing Clinical Trial Design and Precision Medicine in DMD through Machine Learning and MRI Radiomics-Based Trajectory Modeling with Explainable AI

Topic:

Translational Research

Poster Number: 270 T

Author(s):

Hyunjae Jeong, University of Florida, Rebecca Wilcocks, PhD, University of Florida, Glenn Walter, PhD, University of Florida, Robert Mueller, PhD, University of Florida, Fiorella Sanabria, University of Florida, William Triplett, BS, University of Florida, Krista H Elvire Vandenborne, PhD, University of Florida, Sarah Kim, PhD, MS, University of Florida

Objectives: Duchenne muscular dystrophy (DMD) is a rare pediatric condition characterized by progressive muscle degeneration. This study aimed to inform subgroup selection for clinical trial designs and personalized interventions by quantifying the individual-level rate and severity of DMD progression trajectories using machine learning (ML) models. We leveraged previously developed disease progression models and magnetic resonance imaging (MRI) derived radiomic features collected at screening visits.

Methods: This study utilized baseline data from the ImagingNMD study (NCT01484678), which included 68 using Dixon fat fraction MRIs of individuals with DMD. Radiomic features were extracted using PyRadiomics from segmented soleus muscle in 3 axial landmark slices, including shape-based, first-order, and texture-based characteristics. Baseline clinical demographics were also incorporated as inputs. Feature pre-selection was performed using Pearson correlation and LASSO regularization. Multiple ML models were then trained with 5-fold cross-validation to evaluate predictive performances for individual parameters describing the increase in fat fraction in the soleus as DMD progresses: DPT50 (age at which the measure is half of its maximum increase), DPmax (extent), γ (rate), and S0 (extrapolated measure when age is 0). Key radiomic biomarkers were identified through SHAP feature importance and unsupervised clustering.

Results: 80 baseline features were pre-selected to train ML models. The XGBoost regression model achieved the best predictive performance, with a mean absolute error of ±0.56 years for DPT50 (R²=0.88) and ±0.61 for γ (R²=0.79). Feature importance analysis revealed potential radiomic biomarkers, including texture-based nonuniformity, emphasis, correlation, and first-order kurtosis and median. Unsupervised clustering identified distinct subgroups with mean values of DPT50=12.21 and 14.83 years and γ=8.07 and 7.54 in fast (n=28) and slow (n=40) disease progression, respectively.

Conclusions: These findings highlight the potential of ML approaches with baseline MRI-derived radiomic features to accurately predict disease progression in DMD. The explainable AI-assisted feature importance analysis effectively identified potential imaging biomarkers.