Abstract
Background:
Diabetes Hand Syndrome (DHS) is a little-known musculoskeletal consequence of diabetes mellitus that impairs hand function and quality of life. The late clinical manifestations with the lack of standardized early screening tools often delay the diagnosis. Despite the broad use of machine learning (ML) to the complications of diabetes, predictive modeling of the DHS has not been studied extensively.
Objective:
To test, in a simulated proof of concept model, whether interpretable machine learning models are possible to identify early DHS based on combined clinical, demographic, and biomechanical features.
Methods:
An artificial dataset of 8,000 patient records was created to represent clinically realistic distributions of diabetes related variables, such as HbA1c, body mass index, duration of diabetes, neuropathy status, age, gender, and biomechanical variables such as grip strength, wrist mobility, and range of motion of the fingers. The preprocessing involved imputation, scaling and encoding, and the analysis was done based on nested five-fold cross-validation with train-only preprocessing. Discrimination measures, confusion matrices with counts, precisionrecall analysis, calibration plots and decision curve analysis were used to evaluate random forest, XGBoost, and multilayer perceptron models. SHAP and LIME were used to analyze model interpretability.
Results:
The ensemble models showed better results on the neural network in the simulated environment, and HbA1c, grip strength, and the duration of diabetes were found to be significant predictors. The XGBoost showed more consistent discrimination, calibration and behavior at the decision level in comparison to the other models.
Conclusions:
The presented simulated evidence-of-concept shows that interpretable ML is methodologically feasible when it comes to early risk modeling in the DHS. The results are not a sign of the clinical readiness and must be confirmed by real-world clinical data, external cohorts, and prospective studies before the issue of deployment can be considered.

Key words: Diabetes Hand Syndrome; Machine Learning; Proof-of-Concept; XGBoost; Biomechanical Features; Model Interpretability