Explainable Machine Learning Model to Preoperatively Predict Postoperative Complications in Inpatients With Cancer...

Area of Science:

• Oncology
• Medical Informatics
• Surgical Outcomes

Background:

• Predicting postoperative complications (PCs) in cancer patients undergoing major surgery is complex.
• Existing methods often lack precision and explainability for heterogeneous patient populations.

Purpose of the Study:

• To develop and validate an explainable machine learning (ML) model for predicting 30-day postoperative complications (CD 3+) in cancer inpatients.
• To identify key predictive features using the Shapley additive explanations (SHAP) method.

Main Methods:

• Retrospective review of 988 operations in 827 cancer inpatients (December 2017 - June 2021).
• Development and testing of an ML model using electronic health record (EHR) data.
• Performance evaluation using AUROC, AUPRC, and calibration plots; SHAP for model interpretability.

Main Results:

• The ML model achieved an AUROC of 0.77 (training) and 0.73 (holdout) and AUPRC of 0.56 (training) and 0.52 (holdout) for predicting CD 3+ complications.
• The complication rate was 28.6% (training) and 27.5% (holdout).
• SHAP analysis identified key risk factors and their contributions at individual and cohort levels.

Conclusions:

• An explainable ML model can accurately predict the risk of severe postoperative complications in cancer patients.
• The model leverages patient-specific EHR data for risk stratification.
• Explainability features enhance clinical utility by highlighting critical predictive factors.