AI-Driven Model Helps Doctors Navigate Complex Insulin Dosing in ICU

After cardiac surgery, patients are at risk of experiencing both high and low blood sugar levels, which can result in severe complications. Properly managing these fluctuations demands precise insulin dosing, but existing protocols often fall short due to the unpredictable nature of intensive care unit (ICU) environments and patient-specific differences. To tackle this challenge, researchers have developed a machine learning tool designed to assist doctors in managing blood sugar levels in patients recovering from heart surgery—an essential but often complex task in the ICU.

Researchers at the Icahn School of Medicine at Mount Sinai (New York, NY, USA) have created a reinforcement learning model called GLUCOSE, which suggests insulin doses personalized to the needs of each patient. In tests using data from real-world ICU cases, GLUCOSE either matched or outperformed experienced clinicians in maintaining blood sugar levels within a safe range—despite having access only to real-time patient data, while clinicians used complete patient histories. The research team trained GLUCOSE using reinforcement learning, which enabled the system to optimize its decisions through trial and error. The team also applied advanced techniques, including conservative and distributional reinforcement learning, to ensure that the model made reliable and cautious recommendations. The model was thoroughly tested and compared with current clinical practices.

The results, published in NPJ Digital Medicine, show promising potential, but the researchers emphasize that GLUCOSE is not meant to replace doctors. Rather, it functions as a clinical decision support tool, offering suggestions that physicians can evaluate and incorporate into their clinical judgment and broader patient context. Eventually, the model could be integrated into electronic health record systems, providing real-time insulin dosing guidance in the ICU, which could help reduce complications and improve patient outcomes. Future efforts will focus on adapting the tool for use in other hospital settings, conducting clinical trials, and finding ways to incorporate it into routine care. One current limitation is that the model does not yet consider nutrition data, which could affect long-term glucose control. Despite this, GLUCOSE's ability to make accurate recommendations using limited real-time data highlights its potential to improve safety and efficiency in postoperative care.

“Our study shows that artificial intelligence can be thoughtfully and responsibly developed to support, rather than replace, the clinical judgment of health care professionals,” said co-senior corresponding author Ankit Sakhuja, MBBS, MS, Associate Professor of Medicine (Data-Driven and Digital Medicine) and a member of the Institute for Critical Care Medicine at the Icahn School of Medicine at Mount Sinai. “In complex and high-pressure environments like the ICU, tools like GLUCOSE can provide real-time data-driven guidance tailored to individual patients. This kind of decision support can enhance safety, reduce the risk of complications, and ultimately allow clinicians to focus more of their attention on critical aspects of patient care.”