Machine Learning Methods for the Prediction of Intraoperative Hypotension with Biosignal Waveforms.

Background and Objectives: Intraoperative hypotension (IOH) is of great importance in preventing diseases such as postoperative myocardial infarction, acute kidney injury, and mortality. This study aimed to develop and validate machine learning and deep learning models that predict IOH using both biosignals and personalized clinical information for each patient. Materials and Methods: In this retrospective observational study, we used the VitalDB open dataset, which included intraoperative biosignals and clinical information from 6388 patients who underwent non-cardiac surgery between June 2016 and August 2017 at Seoul National University Hospital, Seoul, South Korea. The predictive performances of models trained with four waveforms (arterial blood pressure, electrocardiography, photoplethysmography, and capnography) and clinical information were evaluated and compared at time points at 5 min before the hypotensive event. To predict hypotensive events during surgery, we developed two predictive models: machine learning and deep learning. In total, 2611 patients were enrolled in this retrospective study. Machine and deep learning algorithms were developed and validated using raw waveforms and clinical information as inputs. Results: Gradient boosting machine showed predicted IOH with an AUROC and accuracy of 0.94 (0.93-0.95) and 0.88 (0.86-0.89). A hybrid CNN-RNN model also showed similar performance with an AUROC and accuracy of 0.94 (0.93-0.95) and 0.88 (0.87-0.89). Conclusions: This study developed and validated machine and deep learning models to predict IOH using waveform data and covariate values. In the future, we anticipate that the results of our study will contribute to predicting IOH in real time in the operating room and reducing the occurrence of IOH.