A risk prediction model for poor joint function recovery after ankle fracture surgery based on interpretable machine...

Area of Science:

• Orthopedics
• Medical Informatics
• Machine Learning

Background:

• Individualized prediction of joint function recovery after ankle fracture surgery is currently lacking.
• Early identification of patients at high risk for poor outcomes is crucial for optimizing treatment strategies.

Purpose of the Study:

• To develop and validate a machine learning-based prediction model for poor joint function recovery following ankle fracture surgery.
• To facilitate early identification of high-risk patients through an individualized prediction tool.

Main Methods:

• A cohort of 750 patients undergoing ankle fracture surgery was analyzed.
• The Boruta algorithm was used for feature selection, followed by the development of five machine learning models: logistic regression, random forest, extreme gradient boosting, support vector machine, and lasso-stacking.
• Model performance was evaluated using AUC and accuracy on training and test sets, with further analysis using SHAP and LIME.

Main Results:

• The lasso-stacking model demonstrated the best performance, with an AUC of 0.877 (training) and 0.791 (test), and accuracy of 0.796 (training) and 0.762 (test).
• Key predictors identified by SHAP analysis included functional exercise compliance, combined ligament injury, and open fracture.
• The lasso-stacking model outperformed other evaluated machine learning algorithms in predicting joint function recovery.

Conclusions:

• The developed lasso-stacking model is a promising tool for predicting joint function recovery after ankle fracture surgery.
• This model can aid clinicians in identifying high-risk patients for targeted interventions and improved surgical outcomes.
• Further external validation and clinical implementation of this predictive model are recommended.