AI model built on 240,000 clinical records

38 pre-surgery variables used to assess risk

Professor Yoon Hye-eun (left) of the nephrology department at Catholic University of Korea Seoul St. Mary's Hospital and Professor Min Ji-won of the nephrology department at Bucheon St. Mary's Hospital. [Seoul St. Mary's Hospital]
Professor Yoon Hye-eun (left) of the nephrology department at Catholic University of Korea Seoul St. Mary's Hospital and Professor Min Ji-won of the nephrology department at Bucheon St. Mary's Hospital. [Seoul St. Mary's Hospital]

Catholic University of Korea Seoul St. Mary's Hospital announced Thursday that a joint research team led by Professor Yoon Hye-eun of its nephrology department and Professor Min Ji-won of Bucheon St. Mary's Hospital's nephrology department has developed and patented an AI-based technology to predict the risk of acute kidney injury following surgery.

The registered patent, titled "Method and System for Predicting Post-Operative Acute Kidney Injury," covers a technology that screens high-risk patients in advance using clinical and laboratory data obtainable before surgery.

Acute kidney injury is a condition in which kidney function deteriorates sharply over a short period. It occurs in roughly 5 to 7.5 percent of all hospitalized patients and about 20 percent of intensive care unit patients, with an estimated 30 to 40 percent of hospital-acquired cases linked to surgical procedures.

Predicting the risk of post-operative acute kidney injury before surgery is difficult because multiple factors interact — including baseline kidney function, age, comorbidities, medications, and the type and duration of the procedure. Serum creatinine levels, which are used for diagnosis, rise only after kidney damage has already progressed, making early identification of high-risk patients all the more critical, the research team noted.

The team developed an AI model called CMC-AKIX to predict the risk of post-operative acute kidney injury, drawing on clinical big data from seven university hospitals under the Catholic Medical Center network.

The model was trained on data from 239,267 cases of patients who underwent non-cardiac surgery under general anesthesia between 2009 and 2019. Of those, 7,935 cases, or 3.3 percent, developed acute kidney injury within 30 days of surgery.

The AI model incorporates 38 pre-operative clinical and laboratory variables, including basic information such as age, sex, blood pressure and body mass index; comorbidities including chronic kidney disease, diabetes and hypertension; medication use; laboratory results such as creatinine, glomerular filtration rate and albumin levels; and surgical duration and department.

The team corrected for missing test values using multiple imputation by chained equations, known as MICE, and applied the SMOTE technique to the training data to account for the low incidence rate of acute kidney injury at 3.3 percent. Multiple prediction algorithms were then applied to the same dataset and compared for performance.

A deep neural network model delivered the highest predictive performance. The area under the receiver operating characteristic curve, or AUC, for the model analyzing all variables reached 0.832.

Building on those findings, the team also developed a web-based CMC-AKIX system that calculates a patient's acute kidney injury risk in real time when pre-operative clinical data are entered. The related research was published in the international journal Journal of Medical Internet Research in April last year.

"We can now proactively identify high-risk patients in advance and establish personalized prevention strategies before and after surgery — such as appropriate fluid and blood pressure management and minimizing the use of nephrotoxic drugs," Yoon said. "This will serve as a tool to support rapid and accurate clinical decision-making by medical staff."

Min said the patent covers not only the predictive algorithm itself but also the method of implementing the web-based system that calculates risk scores in real time from pre-operative clinical inputs. "This is significant in that it enhances the prospects for clinical application and commercialization," she added.

Meanwhile, Seoul St. Mary's Hospital was selected this year as the lead institution for the third phase of the government's Medical Data Center Hospital Support Project, and is operating a consortium with Asan Medical Center, the National Cancer Center, Chung-Ang University Hospital and Konkuk University Hospital. The hospital plans to expand the clinical application of medical AI through joint research using medical data and the performance verification and validation of AI models.


woo@heraldcorp.com