mRUST Estimation of Tibial Fracture Healing After Intramedullary Nailing Using Deep Forest Model with a Genetically Optimized Wearable Sensor Layout.

Abstract

The assessment of tibial shaft fracture healing using the mRUST score is limited by radiation exposure and subjective interpretation. This study aimed to develop a quantitative model to estimate mRUST scores using continuous plantar pressure data from a portable insole system and to identify an optimal, cost-effective sensor layout. 23 Patients with tibial shaft fractures treated with intramedullary nails were enrolled. Plantar pressure data and corresponding mRUST scores were collected across 103 follow-up visits. During each visit, data from 5 gait analysis segments were recorded, yielding a total of 515 gait analysis segments. A Deep Forest Regression (DFR) model was developed to estimate mRUST from continuous gait data. A Genetic Algorithm (GA) optimized the sensor layout using the model's coefficient of determination (R²) as the fitness function. Model interpretability was assessed using Shapley Additive Explanations (SHAP). The optimization process identified an optimal 6-sensor layout, which achieved a Mean Absolute Error of 0.641 and an R² of 0.902-performance comparable to the full 99-sensor array. The model demonstrated high estimation accuracy across early, intermediate, and late healing stages. SHAP analysis validated the model's clinical relevance, revealing that sensor contributions shifted from the heel to the forefoot as healing progressed. A DFR model with a GA-optimized plantar pressure insole provides an accurate, objective assessment of patients following intramedullary nailing of tibial fractures. This portable, data-driven approach presents a viable alternative to traditional radiographic methods, offering potential for timely and convenient clinical monitoring.