Importance: Postoperative complications can significantly impact perioperative care management and planning.
Objectives: To assess machine learning (ML) models for predicting postoperative complications using independent and combined preoperative and intraoperative data and their clinically meaningful model-agnostic interpretations.
Design, setting, and participants: This retrospective cohort study assessed 111 888 operations performed on adults at a single academic medical center from June 1, 2012, to August 31, 2016, with a mean duration of follow-up based on the length of postoperative hospital stay less than 7 days. Data analysis was performed from February 1 to September 31, 2020.
Main outcomes and measures: Outcomes included 5 postoperative complications: acute kidney injury (AKI), delirium, deep vein thrombosis (DVT), pulmonary embolism (PE), and pneumonia. Patient and clinical characteristics available preoperatively, intraoperatively, and a combination of both were used as inputs for 5 candidate ML models: logistic regression, support vector machine, random forest, gradient boosting tree (GBT), and deep neural network (DNN). Model performance was compared using the area under the receiver operating characteristic curve (AUROC). Model interpretations were generated using Shapley Additive Explanations by transforming model features into clinical variables and representing them as patient-specific visualizations.
Results: A total of 111 888 patients (mean [SD] age, 54.4 [16.8] years; 56 915 [50.9%] female; 82 533 [73.8%] White) were included in this study. The best-performing model for each complication combined the preoperative and intraoperative data with the following AUROCs: pneumonia (GBT), 0.905 (95% CI, 0.903-0.907); AKI (GBT), 0.848 (95% CI, 0.846-0.851); DVT (GBT), 0.881 (95% CI, 0.878-0.884); PE (DNN), 0.831 (95% CI, 0.824-0.839); and delirium (GBT), 0.762 (95% CI, 0.759-0.765). Performance of models that used only preoperative data or only intraoperative data was marginally lower than that of models that used combined data. When adding variables with missing data as input, AUROCs increased from 0.588 to 0.905 for pneumonia, 0.579 to 0.848 for AKI, 0.574 to 0.881 for DVT, 0.5 to 0.831 for PE, and 0.6 to 0.762 for delirium. The Shapley Additive Explanations analysis generated model-agnostic interpretation that illustrated significant clinical contributors associated with risks of postoperative complications.
Conclusions and relevance: The ML models for predicting postoperative complications with model-agnostic interpretation offer opportunities for integrating risk predictions for clinical decision support. Such real-time clinical decision support can mitigate patient risks and help in anticipatory management for perioperative contingency planning.