Development and Multicenter External Validation of a Real-Time Artificial Intelligence Diagnostic System for Diabetic Peripheral Neuropathy Based on Wearable Devices

INTRODUCTION: Diabetic peripheral neuropathy (DPN) is a frequent diabetes complication, affecting over half of patients and causing pain, falls, ulcers, and amputations if undetected. Traditional methods like electromyography are resource-intensive and inaccessible. This study developed and validated STRIDE, a machine learning model using wearable gait and balance data for real-time DPN classification. METHODS: In this study conducted across multiple centers, 206 participants from Chongqing hospitals were categorized as follows: healthy controls (n = 32), diabetics without DPN (n = 47), asymptomatic DPN (n = 48), and symptomatic DPN (n = 79). A separate group of 42 was used for validation. During a one-minute walk and modified balance test, LEG-Sys/BalanSens sensors produced 68 features. STRIDE used random forest and logistic regression with cross-validation. The metrics used were area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, F1-score, and SHAP for interpretability, with a scoring system allowing for real-time evaluation. RESULTS: Random forest achieved AUROC 0.80 (95% CI: 0.64-0.92); logistic regression 0.79 (95% CI: 0.62-0.95). Key SHAP features: stride length variability (weight = 0.61) and double-support time. External validation: initial accuracy 78.6%, recall 86.5%, precision 88.9%; three of four false positives developed DPN within 1 year, yielding 89.3% adjusted concordance. STRIDE outperformed other parameters in mid-term prediction, unaffected by anemia for up to 4 years. CONCLUSION: STRIDE provides a scalable, interpretable tool for DPN classification, supporting early intervention and improved outcomes. Larger longitudinal studies are needed for clinical integration.