# Clinical performance of machine-learning algorithms to predict intraoperative hypotension: a meta-analysis

## TL;DR

This study reviews how well machine learning can predict low blood pressure during surgery, which could help reduce complications.

## Contribution

A meta-analysis of machine learning algorithms for predicting intraoperative hypotension is presented.

## Key findings

- Machine learning algorithms showed a specificity of 0.73 and sensitivity of 0.75 in predicting hypotension.
- The area under the ROC curve (AUROC) was 0.83, indicating good predictive performance.
- Observational studies included over 42,509 participants from South Korea and Taiwan.

## Abstract

Intraoperative hypotension is one of the most common surgical adverse effects, often associated with postoperative organ damage and complications. The use of non-invasive methods to predict hypotension during surgery could significantly reduce these complications. Recently developed machine learning algorithms show promise for improving the prediction of hypotension and enhancing detection during surgical procedures. Our meta-analysis aimed to evaluate the clinical performance of machine learning algorithms for predicting intraoperative hypotension.

A systematic search of PubMed, Scopus, Web of Science, and Google Scholar was conducted to retrieve articles published up to August 2024. Two independent researchers screened observational accuracy studies and also identified randomized controlled trials (RCTs) of AI-guided intraoperative management. Observational studies reporting diagnostic/prognostic accuracy metrics were included in the quantitative synthesis. After evaluating the quality of the studies, data extraction and analysis were accomplished.

Five observational studies, including a total of 42,509 participants, were included in this study. Four studies were conducted in South Korea, and one in Taiwan. The machine learning algorithms applied in these studies for electrocardiography, basal systolic, and diastolic blood pressure. These studies assessed model performance based on sensitivity, specificity, and the area under the ROC curve (AUROC). The overall effect size, represented in a forest plot, showed specificity of 0.73 (95% CI: 0.67–0.79), sensitivity of 0.75 (95% CI: 0.71–0.79), and an AUROC of 0.83 (95% CI: 0.77–0.89). In addition, we identified RCTs evaluating AI/HPI-guided intraoperative management, due to heterogeneous endpoints.

Machine learning algorithms demonstrate acceptable performance and advantages in predicting intraoperative hypotension during surgical procedures. The application of these algorithms could be beneficial in future surgeries to reduce complications.