Multimodal Artificial Intelligence for Precision Critical Care: A Scoping Review

Abstract

Background: Intensive care units (ICUs) generate large, heterogeneous data from critically ill patients. Multimodal artificial intelligence (AI), which integrates diverse data modalities within unified models, offers substantial potential for precision medicine in critical care. However, comprehensive reviews of its current progress, methods, and challenges remain scarce. Methods: This scoping review systematically surveys the literature on multimodal AI models in critical care, following the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews guidelines. A search was conducted across the PubMed, EMBASE, Scopus, Web of Science, and IEEE Xplore databases between 2010 and 2025. We included studies that reported integrating at least 2 data modalities using AI to evaluate outcomes or tasks relevant to critical care. Results: We included 86 studies, 85% published within the last 5 years. Five data modalities were reported, most commonly structured data ( n = 80) and text ( n = 54), followed by imaging (28), waveforms (16), and videos (2). Intermediate fusion was the predominant strategy for integrating multiple modalities, used in 54 studies. We identified 63 feature extraction methods before modality fusion, 21 modeling approaches applied to 35 distinct clinical outcome tasks after fusion, and 14 explainability techniques. Thirty-four studies shared code, and 9 open-source multimodal ICU databases were identified. Overall, multimodal AI outperformed unimodal models, achieving a 4.4% relative improvement in the area under the curve for diagnosis or prognosis prediction tasks. Conclusion: Drawing on the latest evidence, this review provides a strategic roadmap for the design, evaluation, and implementation of multimodal AI systems in the ICU, with the ultimate aim of improving patient outcomes.

Ähnliche Arbeiten