Clinical Report: Predictive Models Utilizing Machine Learning for PPCs
Overview
This study developed and validated machine learning models to predict postoperative pulmonary complications (PPCs) in geriatric patients undergoing abdominal surgery. The XGBoost model achieved an AUC of 0.856 in the independent test set.
Background
Postoperative pulmonary complications are prevalent in older adults following abdominal surgery, leading to increased morbidity and healthcare costs. Traditional risk assessment tools may not be adequately applicable to this demographic, necessitating improved predictive models. Machine learning approaches can enhance risk stratification by utilizing complex clinical data.
Data Highlights
Cohort
PPC Incidence
AUC
Sensitivity
Specificity
PPV
NPV
Development/Internal
17.3%
0.856
87.5%
60.9%
32.0%
95.9%
External Validation
19.4%
0.821
-
-
-
-
Key Findings
PPCs occurred in 17.3% of the development cohort and 19.4% in the external-validation cohort.
XGBoost outperformed other algorithms with an AUC of 0.856 in the independent test set.
At a 20% risk threshold, the model achieved a sensitivity of 87.5% and a specificity of 60.9%.
SHAP analysis identified key predictors including ASA physical status, COPD, and surgical duration.
The model stratified patients into low-, moderate-, and high-risk groups with observed PPC rates of 6.9%, 24.4%, and 37.6%, respectively.
Clinical Implications
Prospective validation is necessary to confirm the utility of the developed machine learning model in routine clinical practice.
Conclusion
The study presents a machine learning model for predicting PPCs in geriatric surgical patients, highlighting the need for further validation.
