A Retrospective Machine Learning Analysis to Predict 3-Month Nonunion of Unstable Distal Clavicle Fracture Patients Treated with Open Reduction and Internal Fixation

Abstract

Background: This retrospective study aims to predict the risk of 3-month nonunion in patients with unstable distal clavicle fractures (UDCFs) treated with open reduction and internal fixation (ORIF) using machine learning (ML) methods. ML was chosen over traditional statistical approaches because of its superior ability to capture complex nonlinear interactions and to handle imbalanced datasets.

Methods: We collected UDCFs patients at Nanjing Luhe People's Hospital (China) between January 2015 and May 2023. The unfavorable outcome was defined as 3-month nonunion, as represented by disappeared fracture line and continuous callus. Patients meeting inclusion criteria were randomly divided into training (70%) and testing (30%) sets. Five ML models (logistic regression, random forest classifier, extreme gradient boosting, multi-layer perceptron, and category boosting) were developed. Those models were selected based on univariate analysis and refined using the Least Absolute Shrinkage and Selection Operator (LASSO). Model performance was evaluated using AUROC, AUPRC, accuracy, sensitivity, specificity, F1 score, and calibration curves.

Results: A total of 248 patients were finally included into this study, and 76 (30.6%) of them had unfavorable outcomes. While all five models showed similar trends, the CatBoost model achieved the highest performance (AUROC = 0.863, AUPRC = 0.801) with consistent identification of the risk factors mentioned above. The SHAP values identified the CCD as the significant predictor for assessing the risk of 3-month nonunion in patients with UDCFs within the Chinese demographic.

Conclusion: The refined model incorporated four readily accessible variables, wherein the CCD, HDL levels, and blood loss were associated with an elevated risk of nonunion. Conversely, the application of nerve blocks, including postoperative block, was correlated with a reduced risk. Our results suggest that ML, particularly the CatBoost model, can be integrated into clinical workflows to aid surgeons in optimizing intraoperative techniques and postoperative management to reduce nonunion rates.

Keywords: distal clavicle fracture; machine learning; nonunion; prediction.

Figure 1

Receiver operating characteristic (ROC) curve and precision-recall curve (PRC) of the five models for predicting the risk of 3-month nonunion. (a) The ROC curve of the training set. (b) The ROC curve of the testing set. (c) The PRC of the training set. (d) The PRC of the testing set.

Figure 2

Calibration curve of the five models.

Figure 3

Shapley additive explanations (SHAP) plots of the selected variables in the CatBoost model. (a) The mean absolute SHAP value ranks the relative importance of the variables in the model. (b) The direction of effects for each variable in the model.