AI meets offloading: Explainable machine learning for footwear prescription in Hallux-affected plantar foot ulceration

Plantar foot ulceration (PFU) in individuals with structural deformities, such as hallux rigidus, hallux limitus, and claw or hammer toes, poses significant risks of re-ulceration, infection, and amputation. Although offloading footwear and orthotic interventions are central to prevention, current prescription practices lack standardization and rely heavily on clinician experience, leading to inconsistent outcomes and poor adherence. This study aims to develop and evaluate artificial intelligence (AI) approaches, particularly explainable machine learning (ML) models that support data-driven, personalized offloading footwear prescriptions for patients with hallux-related PFU risks. Of over 2500 patient records, a retrospective dataset of 97 confirmed cases from a specialized foot clinic was curated, including patient demographics, activity, preferences, pathologies, biomechanical risk factors, and treatment histories. Twelve state-of-the-art ML classifiers were benchmarked to predict three key offloading parameters: footwear rocker profile (FRP), footwear rocker apex position (FRAP), and custom-made insole modification (CMIM). Model performance was assessed via 10-fold cross-validation. SHapley Additive exPlanations (SHAP) were used to interpret model predictions at both population and individual levels. : Ensemble-based classifiers consistently achieved the strongest performance across all three prediction tasks. Gradient Boosting achieved the lowest Brier score for the FRP task (balanced accuracy 98.22%), XGBoost performed best for FRAP (balanced accuracy 94.03%), and CatBoost achieved the lowest Brier score for CMIM (balanced accuracy 94.37%). SHAP analysis revealed that features related to fall risk, deformity type, mobility, and foot structure were the most influential in predicting prescription parameters. Interpretability visualizations demonstrated a clinically coherent rationale behind model decisions, supporting their potential integration into clinical workflows. : The proposed explainable ML framework provides a clinically interpretable decision-support approach for predicting key offloading prescription parameters in patients with structurally complex feet. By generating transparent, evidence-informed recommendations, this framework may support clinicians in delivering more consistent and personalized offloading prescriptions. • A multi-target predictive framework using twelve state-of-the-art ML classifiers. • Integrated SHAP-based interpretability to ensure clinical explainability and trust. • Validation of the framework using a curated dataset of clinically confirmed cases. • Clinician review of interpretability confirming alignment with Pedorthic principles.