Machine Learning Outperforms Existing Clinical Scoring Tools in the Prediction of Postoperative Atrial Fibrillation During Intensive Care Unit Admission After Cardiac Surgery

Objective(s) Using the Medical Information Mart for Intensive Care III (MIMIC-III) database, we compared the performance of machine learning (ML) to the to the established gold standard scoring tool (POAF Score) in predicting postoperative atrial fibrillation (POAF) during intensive care unit (ICU) admission after cardiac surgery. Methods Random forest classifier (RF), decision tree classifier (DT), logistic regression (LR), K neighbours classifier (KNN), support vector machine (SVM), and gradient boosted machine (GBM) were compared to the POAF Score. Cross-validation was used to assess the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity of ML models. POAF Score performance confidence intervals were generated using 1,000 bootstraps. Risk profiles for GBM were generated using Shapley additive values. Results A total of 6,349 ICU admissions encompassing 6,040 patients were included. POAF occurred in 1,364 of the 6,349 admissions (21.5%). For predicting POAF during ICU admission after cardiac surgery, GBM, LR, RF, KNN, SVM and DT achieved an AUC of 0.74 (0.71–0.77), 0.73 (0.71–0.75), 0.72 (0.69–0.75), 0.68 (0.67–0.69), 0.67 (0.66–0.68) and 0.59 (0.55–0.63) respectively. The POAF Score AUC was 0.63 (0.62–0.64). Shapley additive values analysis of GBM generated patient level explanations for each prediction. Conclusion Machine learning models based on readily available preoperative data can outperform clinical scoring tools for predicting POAF during ICU admission after cardiac surgery. Explanatory models are shown to have potential in personalising POAF risk profiles for patients by illustrating probabilistic input variable contributions. Future research is required to evaluate the clinical utility and safety of implementing ML-driven tools for POAF prediction.