Photo: Helen King/Getty Images
Researchers from Seoul National University Hospital have built an AI model capable of simultaneously predicting usual post-surgery complications on par with specialist assessments.
Based on a media release, the model predicts acute kidney injury, respiratory failure, and in-hospital death – postoperative complications experienced commonly by four in 10 patients – which drive hospital stays and costs.
HOW IT WORKS
The machine learning model, called MT-GBM, was trained using preoperative EHR data from approximately 80,000 surgical patients.
From this dataset, the team selected 16 variables closely associated with the three major postoperative complications, including age, sex, BMI, surgery type, ASA class (American Society of Anesthesiologists physical status classification), anaesthesia duration, and routine laboratory markers such as haemoglobin, albumin, serum creatinine, and white blood cell count.
FINDINGS
Citing results published in npj Digital Medicine, researchers said the model could distinguish high-risk patients at about 82% for acute kidney injury, 91% for respiratory failure, and 89% for in-hospital death across validation cohorts from SNUH, Nowon Eulji University Medical Center, and Korea University Guro Hospital.
The model outperformed both the widely used ASA class criteria and a single-prediction model using the same method.
Using the SHAP (Shapley Addition Method) analysis, the team found that longer anaesthesia duration and low serum albumin concentration were among the strongest contributors to all three post-surgery complications.
The study noted that these findings reflect the model’s ability to evaluate interacting risk factors, similar to how specialists assess surgical risks considering multiple indicators.
WHY IT MATTERS
According to SNUH, recent AI tools have aimed to flag high-risk surgical patients, but most are built to predict only a single complication, limiting their usefulness in practice.
The MT-GBM model, it said, addresses this limitation by providing early risk stratification across multiple complications before surgery, supporting patient counselling, identifying high-risk groups for closer perioperative monitoring, and informing ICU resource planning.
They also highlighted that the model’s reliance on readily available preoperative variables could allow for quick clinical use while addressing interpretability issues common to deep learning systems.
SNUH added that its consistent performance across institutions indicates potential for real-world deployment. The team now plans to incorporate the model into an EMR system-integrated tool for personalised preoperative risk prediction.
THE LARGER TREND
Similar AI innovations for perioperative care have been developed across the Asia-Pacific in recent years. Manipal Hospitals partnered with Singapore-based ConnectedLife for a programme that uses wearable devices to remotely monitor patients' conditions after surgeries.
In 2023, Singapore General Hospital (SGH) deployed a predictive AI tool that helps check a patient's fitness for surgery. Central Adelaide Local Health Network in Australia rolled out a similar tool in 2022.
Early this year, Asan Medical Center, one of South Korea's largest hospitals, introduced an AI-based model that it claims can objectively assess pain occurrence during and after surgery.
Meanwhile, SGH recently launched a chatbot that helps junior doctors prepare for surgery.
