Development and validation of an explainable machine learning model for predicting postoperative pulmonary...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Thoracic Surgery

Background:

• Postoperative pulmonary complications (PPCs) significantly impact patient outcomes after lung cancer (LC) surgery.
• Existing predictive models for PPCs often lack comprehensive validation and interpretability.
• There is a critical need for reliable and explainable tools to predict PPCs in LC surgery patients.

Purpose of the Study:

• To develop and validate an explainable machine learning (ML) model for predicting PPCs in patients undergoing LC surgery.
• To enhance the interpretability of predictive models for PPCs.
• To improve early identification and management of patients at risk for PPCs.

Main Methods:

• A risk factor pool was established through meta-analysis and Delphi surveys.
• Retrospective (n=883) and prospective (n=308) cohorts of LC surgery patients were utilized.
• Feature selection involved univariate analysis, collinearity analysis, nine ML algorithms, and expert consensus, leading to the development of 12 independent and 26 stacking ensemble models. The Shapley Additive Explanation (SHAP) method was used for interpretability.

Main Results:

• The stacking ensemble model combining Support Vector Machine (SVM) and Decision Tree (DT) achieved the highest performance with an AUROC of 0.860 (95% CI: 0.809-0.911) in the retrospective cohort.
• Prospective validation yielded an AUROC of 0.790 (95% CI: 0.744-0.835).
• The model demonstrated robust predictive performance and favorable clinical utility, as indicated by Decision Curve Analysis (DCA).

Conclusions:

• The developed stacking ensemble model (SVM + DT) shows significant potential for predicting PPCs in LC surgery patients.
• The model's explainability (using SHAP) enhances its clinical utility and trustworthiness.
• Further validation in large, multi-center cohorts is necessary before clinical implementation to optimize patient care and resource allocation.