South Korean researchers have developed a multitask AI model that is changing how hospitals anticipate complications after surgery. The tool focuses on three outcomes – acute kidney injury (AKI), post-operative respiratory failure and in-hospital mortality. What makes the AI tool stand out is its balance of accuracy and real-world usability. The model is trained on 80,000 patient records, performs in the 91 percent range for predicting postoperative respiratory failure.
A multitask AI model can predict multiple postoperative risks simultaneously, improving efficiency and outperforming separate single-outcome models.
Most surgical prediction tools collapse when moved outside the hospital where they were developed. They are often overloaded with hundreds of inputs and depend on complex workflows. This Korean model goes in the opposite direction. It uses just 16 common clinical variables, all of which are routinely available in electronic health records. Age, albumin, creatinine, white cell counts, BMI and duration of anaesthesia emerged as the most influential predictors.
A model like this fits naturally into routine surgical care because it demands almost nothing extra from clinical teams. Surgeons can see preoperative risk scores instantly. Anaesthesia teams can plan postoperative monitoring with better foresight. Hospital administrators can prioritise ICU beds or early interventions for patients flagged as high risk. In environments where staff and bed availability are tight, this level of early warning can genuinely improve outcomes.
All testing so far has been within Korean hospitals, so its performance in different populations or surgical systems remains uncertain. The model uses only preoperative variables, which keeps workflow simple but leaves out important intraoperative information. And any AI tool deployed in live clinical environments needs continuous monitoring to detect performance drift. Still, this study shows a promising direction for surgical AI –small, interpretable, externally validated models that actually fit the realities of hospital practice. Instead of oversized systems that never survive beyond the lab, this approach looks built for real-world impact.
