Real-Time Dynamic Prediction of Mortality and Renal Replacement Therapy After Cardiac Surgery Using a Time-Series...

Predicting postoperative deterioration following cardiac surgery remains challenging. Conventional risk scores rely on static variables and fail to capture evolving physiologic trajectories. We developed a time-series deep learning model (DLM) using serial ICU measurements to dynamically predict mortality and continuous renal replacement therapy (CRRT) after cardiac surgery.

Using the Medical Information Mart for Intensive Care database, we analyzed ICU admissions from patients undergoing CABG, isolated or combined with valve surgery, from 2008-2019. Data were split into 70:20:10 training, validation, and testing. Recurrent neural network models were developed to predict in-hospital mortality and CRRT requirement using fixed and dynamic variables. Fixed variables included demographics, comorbidities, and procedure type. Dynamic variables comprised hourly ICU data (vitals, ventilation, vasopressors, and labs). Risk was updated at each timestep by considering the most recent measurements, and a self-attention layer highlighted influential time points.

Of the 7,402 patients included, 1.3% died in-hospital and 1.8% required CRRT. Utilizing full ICU sequences, the DLMs achieved an AUPRC of 0.877 and 0.906, and F1-scores of 0.808 and 0.757 for mortality and CRRT prediction, respectively. Simulating real-world deployment with hourly, updated predictions, model performance improved as physiologic data accumulated. Furthermore, the self-attention layer highlighted critical timepoints driving predictions, offering valuable clinical interpretability.

This is first time-series DLM to provide continuously updated, hourly predictions of both mortality and CRRT for cardiac surgery ICU patients, demonstrating high performance and interpretability. By reliably identifying high-risk patients several days before overt deterioration, this approach may facilitate earlier, proactive clinical intervention.