Systematic review of Artificial Intelligence-based methods for glycemic control and risk prediction in intensive care units.

Abstract

Achieving safe glycemic targets in intensive care remains difficult due to rapidly changing physiology, treatment effects, and measurement noise. We systematically review artificial-intelligence (AI) methods for ICU glycemic management and risk prediction, summarizing datasets, model classes, validation strategies, clinical endpoints, and implementation barriers. Following PRISMA, we searched PubMed, PubMed Central, and Google Scholar for studies published January 2019-March 2025 using terms combining diabetes/glycemia, ICU/critical care, and AI/ML. Inclusion criteria comprised peer-reviewed English-language studies evaluating AI tools for glucose prediction/control or glucose-linked outcomes in ICU cohorts or ICU-grade datasets; non-AI or pre-2019 studies were excluded. Across heterogeneous cohorts and data sources (bedside glucose, continuous glucose monitoring (CGM), EHR physiologic streams), modern ML (particularly tree ensembles and deep sequence models) improves short-horizon glucose prediction and flags impending hypo/hyperglycemia; several algorithms support insulin titration or closed-loop strategies. However, external validation is uncommon, calibration is inconsistently reported, and outcome endpoints vary (e.g., Time-in-Range, hypoglycemia burden, BGRI), limiting meta-analysis. AI shows promise for proactive alerts and individualized insulin dosing in ICU settings, but routine adoption depends on prospective multicenter trials with standardized endpoints, transparent calibration, workflow integration, and safety monitoring.