Predicting gait patterns from actionable impairments in Duchenne muscular dystrophy: A Machine Learning and Explainable Artificial Intelligence study

Background: Prolonging ambulation is an important treatment goal in children with Duchenne muscular dystrophy (DMD). Clinical management targets 'actionable' (i.e., modifiable) impairments, such as progressive muscle weakness and contractures, that underlie gait pathology. Gait classification may improve clinical decision-making, but the utility of gait classification in clinical practice depends on understanding how underlying, actionable impairments contribute to distinct gait patterns, which remains insufficiently understood. The research questions were: (1) Can DMD gait patterns be accurately classified from actionable impairments? and (2) Can the model's predictions be explained, and do these explanations provide clinical utility and increase trust in the model? Methods: A retrospective dataset of 274 lower-limb observations from 137 assessments in 30 boys with DMD was analyzed, including 3D gait analysis, instrumented strength assessment, and clinical examination (manual muscle testing, goniometry and clinical stiffness scale). Observations were classified into the mildly affected, tiptoeing, or flexion gait pattern. Ten predictors representing actionable impairments were included: nine predictors related to muscle weakness and contractures, and body mass index (BMI). A balanced random forest classifier was evaluated with leave-one-group-out cross-validation. Model interpretability was explored using SHapley Additive exPlanations to generate global and local explanations. An interview with a clinical expert assessed the utility of the explanations as the primary outcome, with trust in and expectations of both the model and the explanations as secondary outcomes. Results: The model achieved an accuracy of 74.5%. Global explanations identified hip and knee weakness, gastrocnemius-soleus contractures, and BMI as the most important predictors across gait patterns. Local explanations illustrated how patient-specific impairments informed individual predictions. The user study demonstrated the clinical utility of the explanations, as they were perceived as interpretable, provided useful insights, and these insights were actionable. The explanations largely aligned with the expectations and increased self-reported trust in the model. Conclusions: Gait patterns in DMD can be predicted from clinically actionable impairments, and explainable artificial intelligence can translate model outputs into meaningful clinical insights. This approach is promising for supporting both general and personalized rehabilitation and orthopedic strategies aimed at prolonging ambulation in DMD. Further validation in larger, multi-center cohorts is needed.