Development of AI for 3 simultaneous prediction of surgical complications...Accuracy 82-91%
Nov 10, 2025
|
As a result of the verification, it has better accuracy than a single predictive model and shows consistent performance in external verification, which is expected to be a medical AI model with versatility in various medical environments.
Professor Yoon Hyun-kyu and Professor Lee Hyun-hoon of the Department of Anesthesiology and Pain Medicine at Seoul National University Hospital developed a multi-task machine learning model that simultaneously predicts three surgical complications based on the data of 80,000 surgical patients and announced the results of verifying their performance on the 10th.
40% of surgical patients experience complications such as acute renal injury, respiratory failure, and death during hospitalization, which increases the length of hospitalization and the burden of medical expenses, thereby lowering the quality of life for patients and their families. Recently, AI models for predicting high-risk groups of complications have been developed, but most of them are designed to predict only one type of complication, so there is a problem of poor usefulness.
In response, the research team selected 16 variables that were highly associated with three complications based on preoperative electronic medical records (EHR), and developed a 'multiple predictive machine learning model (MT-GBM)' that predicts acute renal injury, respiratory failure, and death during hospitalization. While existing studies have used tens to thousands of broad variables, the model has selected only at least among items measured by default during preoperative assessments as variables.
As a result of verification in internal and external cohorts (Seoul National University Hospital, Nowon Eulji Medical Center, and Korea University Guro Hospital), the average predictive accuracy (AUROC) was excellent with 0.82 acute renal injury, 0.91 respiratory failure, and 0.89 death during hospitalization. In addition, prediction accuracy was consistent across all cohorts, demonstrating versatility in multiple healthcare settings.
The predictive accuracy of this model was higher for all complications than the ASA physical condition classification criteria widely used for preoperative risk assessment, and consistently outperformed a single predictive model designed by the same method. The research team explained that this is because the multi-model works in a way that comprehensively judges various risk factors, such as the actual thinking process of a specialist.
Furthermore, the research team applied the Shapley Addition Explanation (SHAP) to determine which variables have how significant effects. Results showed that 'long anesthesia time' and 'low blood albumin concentration' were key variables common to the three complications. The longer the anesthesia time, the more complicated the surgery and the greater the physical burden on the patient, and the lower the albumin level, the less nutritious and less able to recover.
The research team explained that the MT-GBM model can accurately predict the risk of complications in patients before surgery, help patients make decisions, select high-risk groups, and streamline the allocation of intensive care unit resources.
Professor Hyun-kyu Yoon (Department of Anesthesiology and Pain Medicine, 1st author) said, "This study is significant in that it has developed a minimum information-based prediction model that can be used immediately in the preoperative treatment stage. In particular, the biggest advantage is that it improves the problem unique to deep learning models that are difficult to interpret the result derivation process and increases the reliability of prediction results."
Professor Lee Hyun-hoon (Department of Convergence Medicine, corresponding author) said that "The consistent performance of the same model in multiple institutions is an important example that artificial intelligence can be used in real-world medical fields. In the future, we plan to develop this model into a patient-specific risk prediction tool before surgery by linking it with an electronic medical record system."
Meanwhile, the results of the study were published in the international academic journal 'npj Digital Medicine (IF; 15.1)'.
|
This article was translated by Naver AI translator.
