AI-Powered Model Enhances Blood Sugar Management Post-Heart Surgery

Researchers at the Icahn School of Medicine at Mount Sinai have developed an advanced machine learning system that aids in managing blood sugar levels in patients recovering from cardiac surgery. This innovative tool aims to address the challenge of maintaining optimal glucose levels in critical care settings, where fluctuations can lead to serious complications. The system, named GLUCOSE, employs reinforcement learning techniques to recommend personalized insulin doses based solely on real-time patient data, matching or surpassing the performance of experienced clinicians.

Managing blood sugar immediately after heart surgery is crucial, yet difficult due to unpredictable patient responses and variations in ICU care. Traditional protocols often struggle to keep blood glucose within safe ranges, risking high or low blood sugar episodes. To improve this, the research team trained GLUCOSE using rigorous reinforcement learning methods, including conservative and distributional approaches, to ensure safe and reliable suggestions.

In testing with real-world ICU data, GLUCOSE demonstrated a remarkable ability to recommend insulin doses that maintained blood sugar within desired levels, even with limited input data, whereas clinicians used full patient histories. This evidence suggests that AI can be a valuable aid in critical care, supporting healthcare providers without replacing them, and helping to reduce complications associated with blood glucose instability.

The team emphasizes that GLUCOSE is designed as a decision support tool—clinicians retain the ultimate decision-making authority. Future development plans include integrating the system into electronic health records for real-time guidance, expanding its use to other hospital environments, and conducting clinical trials to validate its effectiveness further. Although the current model does not incorporate nutrition data, its success in providing accurate insulin recommendations based on limited data showcases its potential to improve safety and operational efficiency in postsurgical care.

As summarized by co-author Girish N. Nadkarni, the goal is to develop AI tools that augment healthcare providers' capabilities and improve patient outcomes. Integrating trustworthy, data-driven technology into clinical workflows promises a future where personalized, precise treatment is more accessible, safer, and more effective.