A real-time machine learning framework for improved intraoperative risk predictions in cardiac surgery patients

Cardiac surgeries are associated with a range of serious postoperative complications, and current predictive models often rely on static summaries of preoperative and intraoperative data compiled after the case. Such development limits their utility in real-time clinical settings. To address this challenge, we developed TE-Transformer, an end-to-end deep learning framework that continuously ingests heterogeneous electronic health record (EHR) data-including preoperative variables, high-frequency vital signs, medication administration records, and brief surgical notes texts-to provide real-time, multitask predictions of multiple postoperative complications.

A retrospective cohort of 6416 patients undergoing cardiac surgery was used to develop and evaluate TE-Transformer. The architecture reduced reliance on manually engineered temporal summary features by encoding irregular intraoperative observations as event-level representations. Continuous risk estimates were updated at 30-minute intervals, enabling a dynamic perspective on patient status. Performance was benchmarked against traditional machine learning models (eg, XGBoost, Random Forest) and deep learning comparators (LSTM, GRU, CNN) frequently used in sequential modelling.

We analyzed 6416 cardiac surgeries (61% male; median [IQR] age 62 [53-70]). Across classification tasks, TE-Transformer achieved AUROC (area under the receiver operating characteristic curve)/AUPRC (area under the precision-recall curve) of 0.8461/0.3994 (mortality), 0.8837/0.8033 (delirium), and 0.7956/0.6666 (acute kidney injury). For continuous endpoints, mean absolute error (MAE) was 0.0633 (low SBP [systolic blood pressure] time), 0.0920 (N glucose high), 0.0020 (N glucose low), and 0.9781 (postoperative ventilation duration). Ablation study showed the full multimodal model improved performance over the preoperative-only baseline, most notably for 30-day mortality AUROC (+0.049), postoperative delirium AUPRC (+0.059), and postoperative ventilation MAE (-0.082), with intraoperative flowsheet events contributing the largest single increment.

TE-Transformer used routinely collected perioperative data to generate continuously updated risk estimates for multiple postoperative complications and achieved better predictive performance than conventional models. These findings support further evaluation of real-time, multimodal prediction frameworks for intraoperative decision support. This approach highlights the potential of transformer-based, multitask prediction as a foundation for future intraoperative decision support in cardiac surgery.