Predicting the risk of postoperative avascular necrosis in patients with talar fractures based on an interpretable machine learning model

TL;DR

This study creates a machine learning model to predict avascular necrosis after talar fractures, helping doctors identify high-risk patients and personalize treatment.

Contribution

The novel contribution is an interpretable machine learning model for predicting postoperative avascular necrosis in talar fracture patients using clinical and surgical variables.

Findings

XGBoost model achieved high AUC values with superior specificity and sensitivity in predicting avascular necrosis.

Six independent risk factors were identified, including BMI, fracture classification, and quality of fracture reduction.

Abstract

This study aims to develop and validate an interpretable machine learning model for predicting avascular necrosis (AVN) following talar fracture, thereby aiding in personalized prevention and treatment.

A retrospective cohort study included patients undergoing surgical intervention for talar fractures at Ningbo No.6 Hospital between January 2018 and December 2023. Multidimensional data encompassing demographic characteristics, fracture-related variables, surgery-related parameters, and follow-up information were collected. Patients were randomly allocated to the training and testing sets in a 7:3 ratio. Potential risk factors for postoperative AVN were screened using univariate and multivariate logistic regression analyses. Six machine learning algorithms were employed to construct the prediction models. The performance of the prediction model was evaluated utilizing metrics including area under the receiver operating characteristic curve (AUC), calibration curves, decision curve analysis (DCA), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, recall, and F1 score. The SHapley Additive exPlanations (SHAP) provided global and local explanations for the optimal model.

A total of 207 patients with talar fractures were enrolled in our study, with 45 (21.74%) developed AVN, and 162 (78.26%) did not. Univariate and multivariable logistic regression identified six independent risk factors including body mass index (BMI), fracture classification, concomitant ipsilateral foot and ankle fractures, smoking, quality of fracture reduction, and fracture type. Performance evaluation demonstrated that Extreme Gradient Boosting (XGBoost model) achieved high AUC values with superior specificity and sensitivity in both the training and testing sets. The SHAP was performed to analyze the relative importance of features within the model visually and illustrate the impact of each feature on individual patient outcomes.

This study successfully developed and validated an interpretable machine learning model incorporating key clinical and surgical variables to predict AVN following talar fractures. The prediction model identified high-risk patients and critical modifiable factors, facilitating personalized prevention strategies to mitigate this severe complication.

Figures