Fast, efficient and accurate prediction of postoperative outcomes using a small set of intraoperative time series

In the period immediately following surgery, patients are at high risk of various negative outcomes such as Acute Kidney Injury (AKI) and Myocardial Infarction (MI). Identifying patients at increased risk of developing these complications assists in their prevention and management. During surgery, rich time series data of vital signs and ventilator parameters are collected. This data holds enormous potential for the prediction of postoperative outcomes. There is, however, minimal work exploring this potential. Moreover, existing approaches rely on deep learning, which is computationally expensive, often requiring specialized hardware and significant energy consumption. We demonstrate that it is possible to extract substantial value from intraoperative time series using techniques that are extremely computationally efficient. We used recordings from 66 300 procedures at the Lyell McEwin Hospital (Adelaide, South Australia), occurring in 2013 through 2020. The procedures associated with 80% of the patients were used for model training, with the remainder held out for testing. A combination of techniques including MultiRocket, Multitask and logistic regression were used to predict Rapid Response Team (RRT) calls within 48 hours of surgery and mortality, AKI and elevated troponin levels within 30 days of surgery. This approach achieved an Area Under the Receiver Operating Characteristic curve (AUROC) (95% CI) on the test data of 0.96 (0.95-0.97) for mortality, 0.85 (0.84-0.87) for AKI, 0.89 (0.87-0.91) for elevated troponin levels and 0.80 (0.78-0.83) for RRT calls, outperforming the ASA score and Charlson comorbidity index on the test population for all outcomes. These results show that roughly equivalent accuracy to computationally expensive modelling approaches using diverse sources of clinical data can be achieved using highly computationally efficient techniques and only a small set of automatically recorded intraoperative time series. This implies substantial potential in the analysis of these time series for the improvement of perioperative patient care. We also performed an analysis of the measurement sampling rate required to achieve these results, demonstrating the advantage of high-frequency patient vitals monitoring.