Machine learning model with <i>post-hoc</i> SHAP interpretation based on perioperative clinical and imaging parameters for prediction of 1-year radiographic recurrence after hallux valgus surgery.

Authors

Abstract

Radiographic recurrence after hallux valgus surgery is an important issue affecting long-term imaging outcomes. Previous studies have suggested that preoperative deformity, postoperative residual HVA/IMA, and sesamoid reduction are associated with recurrence, but nonlinear relationships may exist among multiple parameters. The aim of this study was to establish and temporally validate a machine learning prediction model based on conventional perioperative clinical and imaging parameters and to provide <i>post-hoc</i> model interpretation using SHAP. 597 patients undergoing hallux valgus surgery in the Department of Trauma, First Affiliated Hospital of Xinjiang Medical University from 2021 to 2025 were retrospectively enrolled. Radiographic recurrence was defined as HVA >20° on the fixed 12-month postoperative weight-bearing anteroposterior radiograph. Patients operated from 2021 to 2023 formed the development cohort (<i>n</i>=389), and patients operated from 2024 to 2025 formed the independent temporal-validation cohort (<i>n</i>=208). Eighteen candidate predictors were screened in the development cohort using collinearity assessment, LASSO, RFE, and permutation importance. A unified eight-feature set was then locked and used for comparison of LR, RF, XGBoost, LightGBM, and SVM using 5-fold outer/3-fold inner nested cross-validation conditional on the fixed feature set. Models were evaluated using ROC-AUC, PR-AUC, Brier score, calibration, and DCA, with <i>post-hoc</i> SHAP-based interpretation of the final SVM. A total of 109 patients met the 1-year radiographic recurrence criterion (18.3%). Eight variables were retained: age, preoperative HVA, preoperative IMA, preoperative DMAA, 6-week postoperative HVA, 6-week postoperative IMA, 6-week postoperative DMAA, and 6-week postoperative sesamoid grade. The development-cohort ROC-AUCs of LR and SVM were 0.908 and 0.906, respectively, with corresponding PR-AUCs of 0.766 and 0.757. SVM was retained as the final model on the basis of its lower Brier score and overall performance profile, without claiming statistical superiority over LR. In the temporal-validation cohort, SVM ROC-AUC was 0.944 (95% CI 0.908-0.973), PR-AUC 0.805, Brier score 0.071, sensitivity 0.892, and specificity 0.842. SHAP showed the largest model contributions for 6-week IMA, HVA, and sesamoid grade. The SVM model with <i>post-hoc</i> SHAP-based interpretation showed promising performance for prediction of fixed 1-year radiographic recurrence using routine clinical and perioperative imaging parameters. Residual IMA, HVA, and sesamoid grade at 6 weeks were leading predictive signals within the model rather than causal determinants of recurrence. The model may support early risk stratification and enhanced follow-up, but multicenter external validation, reproducibility assessment, and prospective impact studies remain necessary before clinical implementation.

Tags

Topics