Development and internal evaluation of an interpretable machine learning model based on clinical and radiomics...

Accurate differentiation between lower extremity arterial embolism (AE) and atherothrombosis is essential. This study aimed to develop and validate an interpretable machine learning (ML) model combining clinical and radiomics features for preoperative non-invasive diagnosis.

This retrospective study enrolled patients with lower extremity AE or atherothrombosis admitted to the Department of Vascular Surgery at our institution between January 2018 and January 2025. Clinical data were collected through the electronic medical record system, and radiomics features were extracted from lower extremity computed tomography angiography (CTA) images. Based on these features, three ML models were constructed, comprising a clinical model, a radiomics model, and a combined clinical-radiomics model. The diagnostic performance was evaluated using the area under the curve (AUC), sensitivity, specificity, F1-score, and accuracy. Furthermore, clinical utility was validated using calibration curves and decision curve analysis, while model interpretation was conducted via the SHAP method.

Among the three predictive models, the fusion model demonstrated the optimal differential diagnostic performance. In the testing set, this model yielded an AUC of 0.821 (95% CI 0.718-0.924), with sensitivity, specificity, F1-score, and accuracy of 0.750, 0.714, 0.761, and 0.734, respectively; the positive likelihood ratio (LR+) and negative likelihood ratio (LR-) were 2.62 and 0.35. In the training set, its AUC reached 0.979 (95% CI 0.959-0.999), with sensitivity, specificity, F1-score, and accuracy of 0.940, 0.939, 0.946, and 0.940, respectively; the LR+ and LR- were 15.41 and 0.06.

The clinical-radiomics fusion model demonstrates potential as a non-invasive auxiliary tool to assist in differentiating lower extremity atherothrombosis from AE in a single-center setting, warranting prospective multicenter validation.