Parsimonious and explainable machine learning for predicting mortality in patients post hip fracture surgery

Area of Science:

• Geriatric Surgery
• Machine Learning in Healthcare
• Surgical Outcomes Research

Background:

• Hip fractures in the elderly are associated with high mortality rates, posing a significant clinical challenge.
• Existing risk prediction models may not fully capture the complexity of post-hip fracture mortality.
• Advancements in machine learning offer potential for more accurate risk assessment.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting 30-day mortality risk in elderly patients undergoing hip fracture surgery.
• To compare models using pre-operative data versus those incorporating both pre-operative and post-operative factors.
• To identify key predictors and ensure model interpretability for clinical application.

Main Methods:

• Utilized data from 62,492 patients in the National Surgical Quality Improvement Program (NSQIP 2012-2017).
• Developed and optimized various machine learning algorithms (e.g., AdaBoost, CatBoost) using tenfold cross-validation and hyperparameter tuning.
• Employed feature selection and explainability techniques (SHAP) to derive parsimonious and clinically plausible models.

Main Results:

• The best pre-operative model (AdaBoost) achieved an AUC of 0.792 (29 features), and the best post-operative model (CatBoost) achieved an AUC of 0.885 (45 features).
• Optimized models with fewer features retained high performance: AUC 0.725 (8 features) for pre-operative and AUC 0.8529 (6 features) for post-operative.
• Explainability analysis confirmed that the learned patterns were clinically relevant.

Conclusions:

• Machine learning models can effectively predict 30-day mortality risk following hip fracture surgery in the elderly.
• Models incorporating post-operative data demonstrate superior predictive performance.
• The developed, explainable models with a limited feature set are highly applicable in clinical settings for risk stratification and decision-making.