UC Irvine

Patients undergoing high-risk surgeries are often at higher risk of developing hemodynamic instability during surgery resulting in poor postoperative outcomes. This is usually associated with significantly increased postoperative morbidity and mortality, which therefore makes the early identification of these critical events and those patients at risk of postoperative complications crucial. With these motivations in mind, we first created a large deidentified research dataset of surgical case medical records from University of California, Irvine Medical Center (UCIMC) matched with physiological waveforms as well as intermittent vital sign values, lab values, and ventilator settings. To our knowledge, such a dataset does not currently exist for the intraoperative environment. We hope that creating a such a dataset will allow for advances in machine learning for intraoperative care. Using medical data from UCLA, we have developed deep neural network models to classify the risks of postoperative mortality, acute kidney injury, and reintubation utilizing readily available intraoperative information. Our risk scores were compared to currently commonly used risk indices ASA and Surgical Apgar as well as logistic regression. While the deep neural network models performed better than the risk scores and logistic regression, clinicians require additional information to assess what led to increased risk of complications. To address this, we also assessed the use of generalized additive neural networks (GANNs) to create a graphical look at how different features contributed to the risk of in hospital mortality. Finally, we were also interested in predicting critical intraoperative events to allow for time for the clinician to avoid such events. We focused on intraoperative hypotension as it is easier to define and has been shown to lead to increased risk of acute kidney injury, stroke, and myocardial injury. For the hypotension prediction models, we looked at the arterial pressure waveform and EMR data as inputs.

Overall, these aims address a gap in current clinical decision guidance and support to reduce adverse events during surgery as well complications after.