Machine Learning Integrating Quantitative MRI and Clinical Features Predicts Loss of Ambulation in Duchenne Muscular Dystrophy.

BACKGROUND

Loss of ambulation (LOA) is a key milestone in Duchenne muscular dystrophy (DMD) with highly variable timing, complicating clinical counselling and trial design. Quantitative MRI non-invasively assesses muscle fat infiltration, yet few models have integrated MRI with clinical variables for individualized LOA prediction. We aimed to develop and validate a machine-learning model integrating quantitative MRI and clinical features to predict time to LOA and to derive interpretable, individualized risk estimates and candidate thresholds.

METHODS

This prospective study enrolled 188 boys with DMD (median age 9.10 years). Quantitative MRI parameters (T1, T2, fat fraction [FF]) of 19 hip/thigh muscles and clinical variables were measured. Patients were randomly split into training (80%) and test (20%) sets. Feature selection used LASSO, Boruta and recursive feature elimination based on random survival forest (RSF). Three survival models were compared: RSF and a ridge-penalised Cox model (both integrating clinical and MRI-derived features) and a multivariable Cox model restricted to clinical variables. Performance was assessed by concordance index (C-index), time-dependent area under the curve (AUC), integrated Brier score (IBS) and decision curve analysis at 12, 24, 36 and 48 months.

RESULTS

Over median 43.5 months follow-up, 66 patients (35.1%) experienced LOA. Six predictors were retained: NSAA score, MRC%, age and three FF parameters (rectus femoris, vastus lateralis, obturator internus). In the test set (n = 40, 15 events), RSF achieved a C-index of 0.870 (95% CI, 0.764-0.957), outperforming the clinical Cox (0.804) and ridge Cox (0.853) model. Time-dependent AUCs of the RSF model were 0.984, 0.971, 0.913 and 0.850 at 12, 24, 36 and 48 months. RSF showed the best calibration, with the lowest IBS (0.076, 95% CI: 0.052-0.111), followed closely by the ridge Cox model (0.078) and the clinical Cox model (0.109). RSF stratified patients into low-risk, intermediate-risk and high-risk groups; no LOA events occurred in the low-risk group, with annual event rates were 7.2% and 34.6% in the intermediate- and high-risk groups (log-rank p = 0.0025). SHAP analysis identified exploratory thresholds (NSAA < 14.55, MRC% < 68.36%, rectus femoris FF > 32.30%, vastus lateralis FF > 58.00%, obturator internus FF > 7.43%). An interactive web application enables individualized risk prediction (https://westcndmd-timer-loa-risk-prediction-tool-8cygtrrbykvd9snmfcnhrp.streamlit.app/).

CONCLUSIONS

A machine-learning-based RSF model integrating quantitative MRI FF metrics and clinical variables exhibited favourable predictive performance for time to LOA in DMD across short- and long-term horizons. It provided interpretable, individualized risk estimates and was deployed as an online tool to support clinical decision-making and trial enrichment.