Explainable machine learning for differential diagnosis of diabetic foot infection and osteomyelitis: a two-center...

BACKGROUND: Diabetic foot complications, including infections and osteomyelitis, pose significant health risks, with high prevalence and amputation rates. Differentiating diabetic foot infection (DFI) from osteomyelitis (OM) is challenging due to overlapping symptoms and limitations of current diagnostic methods. This study aimed to develop and validate an explainable machine learning (ML) model using routine blood biomarkers to improve differential diagnosis and provide a clinically accessible tool. METHODS: This retrospective, two-center study included 3,612 patients diagnosed with either DFI (n = 1,699) or OM (n = 1,913). Data from Center 1 (n = 3271) were used for model development (75% training, 25% internal validation), and data from Center 2 (n = 341) served as an independent external validation cohort. A robust feature selection pipeline identified the most predictive routine biomarkers. Multiple machine learning classifiers were trained and evaluated, with the top-performing model selected based on the area under the receiver operating characteristic curve (AUC), Brier score, and other key metrics. Explainable AI (XAI) techniques (SHAP, LIME) were used to ensure model transparency. A web-based calculator was developed for clinical translation. RESULTS: A LightGBM model using only six biomarkers—Age, HbA1c, Creatinine, Albumin, ESR, and Sodium—was selected as the final model. It achieved an AUC of 0.879 (95% CI 0.854–0.902) in internal validation and demonstrated excellent, generalizable performance in the external cohort with an AUC of 0.942 (95% CI 0.936–0.950). The model was well-calibrated and showed significant clinical utility in decision curve analysis. SHAP analysis quantified the specific contribution of each biomarker to individual predictions, enhancing interpretability. The final model was deployed as a user-friendly, publicly accessible web calculator. CONCLUSIONS: An externally validated machine learning model based on six routine blood biomarkers can accurately and reliably differentiate DFI from OM. The model demonstrated high discriminative performance and clinical utility. Deployed as a transparent web calculator with integrated explainable AI, this low-cost tool has the potential to aid clinicians in diagnostic decision-making, particularly in resource-limited settings. CLINICAL TRIAL NUMBER: Not applicable.