Forecasting the fall: the role of machine learning in predicting intraoperative hypotension, a scoping review.

Abstract

Introduction: Intraoperative hypotension is associated with increased risk of postoperative mortality, myocardial injury, acute kidney injury and stroke. Early identification with machine learning models allows pre-emptive management to reduce incidence and duration of intraoperative hypotension. This study aims to assess the accuracy of machine learning models in predicting intraoperative hypotension and its impact on clinical outcomes.

Evidence acquisition: This scoping review looked at databases Medline, Embase, PubMed and Cochrane from inception to 25 June 2024. Inclusion criteria were use of machine learning algorithms predicting intraoperative hypotension in adult surgical patients (≥18 years of age). Data extracted were the type and accuracy of machine learning models used, type of surgery, incidence and duration of hypotension, and patient relevant outcomes including length of stay, end organ dysfunction and mortality.

Evidence synthesis: Twenty-six studies were included (N.=48,707 patients). About 92.3% of studies were done in non-cardiac surgeries; 65.4% of studies used a proprietary machine learning algorithm known as the Hypotension Prediction Index (HPI), followed by neural networks (19.2%) and logistic regression (19.2%). HPI was the most accurate in predicting intraoperative hypotension up to 15 minutes prior to the event with the median area under the receiving operator characteristic curve of 0.912 (0.896-0.930). Machine learning resulted in a statistically significant difference in dose and incidence of vasopressor use in three studies and a significant increase in volume of fluids in two studies. Two studies showed a significant reduction in length of stay, postoperative complications and quantity of blood transfusion products.

Conclusions: Despite the ability of machine learning algorithms to predict intraoperative hypotension to a high degree of accuracy, practical implications are not yet fully elucidated. Studies on machine learning predicting intraoperative hypotension are in their early stages with a larger emphasis on algorithm accuracy rather than clinical outcomes.