Duchenne muscular dystrophy is an X-linked recessive neuromuscular disorder caused by mutations in the DMD gene that leave skeletal muscle starved of functional dystrophin. Without this structural protein, muscle fibers undergo chronic damage, progressive fibrosis and fatty infiltration, culminating in motor weakness and irreversible loss of function. Loss of ambulation, defined in the study as the inability to complete an unassisted 10-meter walk or run within 45 seconds or the need for long-term wheelchair use for more than three months, is a pivotal clinical milestone tied to accelerated disease progression and poorer physical, social and emotional outcomes. Yet the timing of this milestone varies enormously between individuals, shaped by glucocorticoid regimens, genotype and day-to-day test performance, which complicates counselling, therapeutic decision-making and clinical trial design.

The study, conducted as part of a prospective longitudinal single-centre cohort registered with the Chinese Clinical Trial Registry, screened 447 boys with genetically or pathologically confirmed DMD enrolled between March 2020 and February 2023. After excluding patients with substantial missing data, no quantitative MRI, disease-modifying therapies such as gene therapy or exon skipping at baseline, or existing loss of ambulation, 188 patients entered the final analysis. Their median age was 9.10 years and the median follow-up lasted 43.50 months. During that period, 66 patients, or 35.1 percent, lost ambulation. The cohort was randomly split into a training set of 148 patients with 51 events and an independent test set of 40 patients with 15 events, balanced across demographic, clinical and imaging parameters.

The imaging side of the model rests on quantitative MRI performed on a 3-tesla scanner, covering the pelvis from the iliac crest to the mid-thigh with T1 mapping, T2 mapping and Dixon water-fat sequences. Two radiologists, blinded to clinical data, measured T1 relaxation time, T2 relaxation time and fat fraction for 19 individual hip and thigh muscles, from the gluteal group to the quadriceps and hamstrings, using regions of interest placed at four standardized cross-sectional levels near the sciatic foramen, the greater trochanter, and the proximal and middle femoral diaphysis. Fat fraction, computed as the fat signal divided by the combined fat and water signal, directly quantifies the irreversible fatty infiltration that is a core pathological hallmark of DMD, whereas T1 and T2 relaxation times largely reflect potentially reversible edema and inflammation.

Feature selection proved strikingly selective. Univariate Cox regression retained three clinical variables, the North Star Ambulatory Assessment score, the Medical Research Council muscle strength percentage and age. For the quantitative MRI parameters, a three-step dimensionality reduction combining LASSO regression, the Boruta algorithm and recursive feature elimination yielded an intersection of exactly three fat fraction measurements: rectus femoris, vastus lateralis and obturator internus. Notably, no T1 or T2 parameter survived the screening. The researchers reasoned that this makes physiological sense, since fat fraction shows the highest responsiveness to disease progression and can detect subclinical worsening even in functionally stable patients, while T1 and T2 excel instead at early diagnosis and differentiating DMD from Becker muscular dystrophy. The final six-feature model thus paired three clinical measures with three imaging biomarkers of irreversible muscle damage.

Three survival models were then built and benchmarked. A conventional clinical Cox model used only the three clinical variables, while a ridge-penalised Cox model and a random survival forest, an ensemble of 1000 decision trees that dispenses with assumptions about survival time distribution and proportional hazards, used all six predictors. The random survival forest dominated. In the training set it achieved a concordance index of 0.932 and the best calibration, with an integrated Brier score of 0.050, compared with 0.897 and 0.070 for the ridge model and 0.895 and 0.074 for the clinical Cox model. Time-dependent areas under the curve in training reached 0.977 at 12 months, 0.974 at 24 months, 0.977 at 36 months and 0.963 at 48 months, outperforming both Cox models at every horizon. Decision curve analysis showed the highest net clinical benefit across the entire relevant threshold range, most pronounced at 48 months.

Performance held up in the independent test set, where the model faced data it had never seen. The concordance index was 0.870, versus 0.853 for the ridge Cox model and 0.804 for the clinical model, and the integrated Brier score remained the best at 0.076. Time-dependent AUCs were 0.984 at 12 months, 0.971 at 24 months, 0.913 at 36 months and 0.850 at 48 months, again beating the clinical Cox model at all four horizons. Accuracy declined gradually from short-term to long-term predictions, but the sustained discrimination across four years of follow-up is what the authors highlight as particularly valuable for long-term care planning, from timing rehabilitation interventions and glucocorticoid adjustments to financial planning for rising care costs.

Perhaps the most clinically tangible result came from risk stratification. Individual hazard scores from the model were divided into low-, intermediate- and high-risk tertiles in the training set, and those pre-defined cutoffs were applied unchanged to the test set. Kaplan-Meier curves separated the groups decisively, with log-rank p-values below 0.0001 in training and 0.0025 in testing. The absolute numbers are stark: in the training set, zero of 50 low-risk patients lost ambulation, compared with 8 of 49 intermediate-risk and 43 of 49 high-risk patients, translating to annual event rates of 8.2 percent and 42.4 percent per 100 person-years above the low-risk baseline. In the test set the pattern repeated, with 0, 3 and 12 events and rate differences of 7.2 percent and 34.6 percent. No low-risk patient in either cohort experienced the endpoint during follow-up.

Interpretability analysis using SHAP values ranked the NSAA score as the most influential predictor, followed by the three fat fraction parameters, MRC percentage and age. Dependence plots with LOWESS smoothing identified exploratory candidate thresholds: an NSAA score of 14.55, below which risk contributions turned positive, closely matching the mean score of about 13 reported by other researchers two years before ambulation loss; fat fraction cutoffs of 32.30 percent for rectus femoris, 58.00 percent for vastus lateralis and 7.43 percent for obturator internus, above which risk rose sharply; and an MRC percentage of 68.36 percent. Time-varying SurvSHAP analysis added a temporal dimension, showing that the importance of muscle fat fractions increased steadily over the follow-up period while NSAA importance peaked at medium-term horizons and then declined, suggesting imaging biomarkers of irreversible damage take over as patients approach the endpoint. The selection of obturator internus, a deep hip external rotator not previously reported in DMD prognosis, may reflect systemic fatty infiltration extending beyond ambulatory muscles, though the authors stress this finding requires independent validation.

The researchers are candid about limitations. The SHAP-derived thresholds are exploratory reference values, not definitive clinical cutoffs, and feature collinearity can shift inflection points if different feature combinations are used. The single-centre design and moderate sample size limit generalisability, the zero events in low-risk groups during a median 43.5-month follow-up likely reflect limited observation time rather than true zero risk, time-varying covariates were not incorporated, and comparisons with other machine learning survival approaches such as gradient boosting or deep survival networks were not performed. Manual segmentation of muscle fat fraction also requires trained radiologists, motivating future work on automated segmentation. Still, the study demonstrates that machine learning integrating quantitative MRI with clinical features can deliver accurate, interpretable and individualized predictions of a hard clinical endpoint in DMD, and by making the model freely accessible as a web application where users can run what-if scenarios, such as the effect of a three-point NSAA improvement, the team has lowered the barrier between survival analysis research and the families who need its answers most.

Subject of Research: Machine learning prediction of loss of ambulation in Duchenne muscular dystrophy using quantitative MRI and clinical features

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

Article References: Liu, C., Song, Y., Wan, J., Xu, T., Zhang, X., Yan, Y., Fu, H., Cai, X., Guo, Y., Wang, Y., & Xu, H. (2026). Machine Learning Integrating Quantitative MRI and Clinical Features Predicts Loss of Ambulation in Duchenne Muscular Dystrophy. Journal of Cachexia, Sarcopenia and Muscle, 17(5), Article e70413. https://doi.org/10.1002/jcsm.70413

Image Credits: AI Generated

DOI: 10.1002/jcsm.70413

Keywords: Duchenne muscular dystrophy, machine learning, quantitative MRI, random survival forest, loss of ambulation, fat fraction, predictive modeling, risk stratification, SHAP, biomarkers, clinical prediction, web application