Artificial Intelligence in Intensive Care: An Overview of Systematic Reviews with Clinical Maturity and Readiness Mapping

Abstract

Background: ICU care is time critical and data dense, making it a promising but high-risk setting for AI decision support when tools are weakly validated. ICU AI evidence is heterogeneous, with limited external validation, inconsistent clinically actionable reporting, and scarce real-world impact data, yielding fragmented review conclusions. We mapped five prespecified ICU domains and assessed clinical and implementation maturity to identify key translational gaps. Methods: We performed a PRIOR-aligned overview of systematic reviews with prespecified maturity constructs. PubMed, Embase, and Web of Science were searched (title and abstract) on 13 December 2025, supplemented by backward citation searching. Two reviewers screened and extracted data with arbitration, assessed the review-level risk of bias using ROBIS, and synthesized findings without meta-analysis using a SWiM-guided narrative prioritizing AUROC ranges. Results: We included 34 systematic reviews (2017–2025) across five ICU domains, dominated by prognostic and early warning applications, mostly in adult populations and commonly using EHR and multimodal inputs. Reporting focused on discrimination, with AUROC ranges roughly 0.54–0.99 for prognostic tasks and 0.64–0.99 for diagnostic tasks, while calibration and clinical utility were rarely addressed and overlap suggested partial dependence. Maturity signals clustered at low-to-intermediate levels, with no evidence for routine, and regulated CDS deployment at the review level. Conclusions: Review-level evidence indicates a translational gap between retrospective performance and clinically mature, safely deployable ICU AI, supporting priorities for external validation, prospective impact studies, standardized reporting including calibration, and governance-focused implementation.