- AI models using multimodal health records and sensors enable improved risk stratification and continuous assessment, but evidence is limited and few models are implemented.
- AI must augment evidence-based delirium care with defined use cases, external validation, action pathways, equitable performance, privacy, consent, clinician oversight, and monitoring.
- Prospective comparative studies should assess patient outcomes, resource use, and balancing measures because improved detection alone does not guarantee better outcomes.
J Am Geriatr Soc. 2026 Nov 5. doi: 10.1111/jgs.70823. Online ahead of print.
ABSTRACT
BACKGROUND: Delirium affects up to one-third of hospitalized older adults and is associated with adverse outcomes. Despite its clinical importance and the availability of validated screening tools, current care gaps extend beyond under-recognition to inconsistent risk stratification, workforce constraints, documentation burden, fragmented intervention delivery, and practical challenges with monitoring of response. We review how artificial intelligence (AI) could bridge these gaps and distinguish model development and validation, real-world clinical implementation, and emerging concepts in delirium care.
RESULTS: Prediction models leveraging multimodal health record data, including structured variables and unstructured clinical notes, report good discrimination for delirium risk stratification, but only a few models have been integrated into clinical practice. A live multimodal machine learning model increased targeted delirium detection, yet a pre-post comparative study found a longer length of stay; improved recognition alone therefore does not equate with improved outcomes. Natural language processing, large language models, computer vision, ambient sensing, and wearable technologies enable more continuous, objective, and less labor-intensive assessment compared with traditional screening methods, but evidence remains limited with concerns about false positives and generalizability. AI-supported delirium management through reinforcement learning and digital twins is currently largely conceptual. Responsible AI adoption requires a clearly defined intended use, external validation, action pathways, equitable performance, privacy and security, consent and transparency, clinician oversight, patient preferences, and post-deployment monitoring.
CONCLUSION: AI is most defensible as an adjunct that connects evidence-based delirium care to reliable execution rather than as a replacement for clinical assessment. Prospective comparative studies should evaluate patient outcomes, resource use, and balancing measures.
PMID:42988738 | DOI:10.1111/jgs.70823
Share Evidence Blueprint
Save to Google Notes
Search Google Scholar
Save as PDF
⭐ My Revision List
