A recent study in the Ultrasound in Medicine & Biology highlighted the role of various models in predicting diabetic tibial neuropathy (DTN) based on multimodal ultrasound features. Eight machine learning-based models for multimodal ultrasound were developed to predict DTN. For the quantification of each feature, the Shapley Additive exPlanations (SHAP) framework was used. The study aimed to develop models for predicting the development of DTN.
A prospective analysis was conducted using multimodal ultrasound and clinical data on 255 patients with suspected DTN at the Second Affiliated Hospital of Fujian Medical University. Using Least Absolute Shrinkage and Selection Operator (LASSO) regression, key features were selected. Using Extreme Gradient Boosting (XGB), Support Vector Machines, Decision Tree, Logistic Regression, k-Nearest Neighbors, Neural Network, and Naïve Bayes predictive models were constructed. To refine model interpretability, the SHAP method was used. One hundred thirty-five patients from three other tertiary hospitals were also collected for external testing. 
Results demonstrated that the cross-sectional area (CSA), Superb microvascular imaging (SMI), Echo intensity (EI), History of smoking, and elasticity value (Emean) were identified by LASSO and were found to be key features for DTN prediction. The Area Under the Curve (AUC) of 0.94, 0.83, and 0.79 was achieved by the XGB model in the training, Internal test, and external test sets, respectively. The ranking significance of EI, Emean, CSA, History of smoking, and SMI was highlighted by SHAP analysis. The contribution of each feature to the final prediction and the enhancement of model interpretability were demonstrated by the SHAP values. Decision plots demonstrated the influence of features.