'Intraoperative Hypotension' Affecting Patient Mortality, Predicted with 90% Accuracy by AI

Seoul Asan Hospital-National Cancer Center Research Team Develops
Learned from 10,000 Patients' Arterial Blood Pressure Data
Predicts Intraoperative Hypotension for Early Intervention Possible

Intraoperative hypotension commonly occurs in patients under general anesthesia but affects postoperative complication rates and mortality. To prevent this, anesthesiology specialists must monitor changes in arterial blood pressure during surgery to identify the risk of hypotension and minimize its occurrence. Above all, if the probability of hypotension can be predicted in advance, medical staff can take preemptive measures, significantly improving surgical safety and patient prognosis.

Recently, a domestic research team has attracted attention by successfully developing an artificial intelligence (AI) system that predicts the probability of hypotension early. Professors Kim Seong-hoon and Park Yong-seok from the Department of Anesthesiology and Pain Medicine at Seoul Asan Medical Center, along with Professor Kim Joon-tae from the Department of Cancer AI Digital Health at the Graduate School of International Cancer Studies, National Cancer Center, developed an AI model trained on arterial blood pressure data from over 10,000 surgical patients. This model predicted patients at risk of intraoperative hypotension with approximately 91% accuracy.

Professors Seonghoon Kim and Yongsuk Park from the Department of Anesthesiology and Pain Medicine at Asan Medical Center, Seoul, and Professor Juntae Kim from the Department of Cancer AI Digital Health at the Graduate School of Cancer Science and Policy, National Cancer Center (from left).

Intraoperative hypotension is defined as a condition where the patient's mean arterial pressure remains below 65 mmHg for at least one minute and can occur due to bleeding or drug side effects. When hypotension occurs during surgery, it can lead to complications such as acute kidney injury and myocardial infarction, affecting mortality rates, making it crucial to minimize hypotensive states. However, previous AI research had limitations due to a lack of interpretability in the prediction process.

To overcome this, the research team trained an AI model using arterial blood pressure data from 10,454 patients who underwent surgery at Seoul Asan Medical Center between 2018 and 2021, extracting arterial blood pressure trends that represent the volume of blood flowing inside blood vessels. They then predicted the probability of hypotension occurring 10 minutes later by comparing the similarity between representative arterial blood pressure trends for different risk levels of hypotension. Validation was conducted through internal validity testing using patient data and external validity testing using patient data registered in an open-source dataset. The results showed prediction accuracies of approximately 91% for internal validation and about 90% for external validation.

Probability of hypotension occurrence predicted by an artificial intelligence model and interpretation data.
[Photo by Seoul Asan Hospital]

Additionally, when evaluating the applicability of the AI model with 17 anesthesiology specialists at Seoul Asan Medical Center, the model scored 24%, 41%, and 26% higher than the widely used SHAP (Shapley Additive Explanations) method in clinical accuracy, clinical usefulness, and intraoperative decision-making willingness, respectively.

Professor Kim Seong-hoon stated, "This study is meaningful in that it provides medical staff not only with the probability of hypotension occurrence but also with real-time reasoning, addressing unmet medical needs." He added, "We will continue research to determine whether the AI model helps improve patient prognosis." Professor Kim Joon-tae explained, "The significance of this study lies in it being the first to evaluate anesthesiology specialists working in actual operating rooms, rather than focusing solely on improving prediction performance."

This study was recently published in the prestigious AI journal 'IEEE Transactions on Neural Networks and Learning Systems (IF=14.255)'.

This content was produced with the assistance of AI translation services.