Machine learning and artificial intelligence for delirium prediction with Electronic Health Records (EHR): a scoping review

Delirium, an acute and fluctuating neurocognitive disorder prevalent among hospitalized and geriatric surgical patients, remains a pervasive yet underrecognized clinical challenge. Leveraging Electronic Health Records (EHRs), Machine Learning (ML) models have emerged as promising tools for early prediction and intervention. This scoping review synthesizes the existing literature, identifies current research gaps, and outlines future directions to advance delirium prediction modeling. Following the PRISMA Extension for Scoping Reviews (PRISMA-ScR) guidelines, literature from 2020 to 2025 was systematically searched across Google Scholar, EMBASE, PubMed, Scopus, and Web of Science using a comprehensive query strategy. The review highlights a significant reliance on structured preoperative and intraoperative EHR for delirium prediction, despite the existence of abundant and highly informative unstructured clinical narratives. Furthermore, a substantial heterogeneity exists in the utilized delirium identification methodologies (e.g. Nursing Delirium Screening Scale (Nu-DESC), Delirium Observation Screening Scale (DOSS), International Classification of Diseases (ICD) criteria, 4AT delirium detection, Confusion Assessment Method (CAM), Intensive Care Delirium Screening Checklist (ICDSC), Cornell Assessment of Pediatric Delirium (CAPD) Diagnostic and Statistical Manual of Mental Disorders 5th version (DSM-5), natural language processing (NLP) based analysis), alongside a focus on specific surgical subgroups. This limited data utilization and methodological variation pose challenges to ML model generalizability and robustness. The literature also showed a research emphasis on critically ill patients, potentially overlooking subtle delirium in low-severity cases. Future research should focus on early risk stratification and prioritize four key areas: (1) expanded utilization of both tabular EHR and unstructured clinical notes; (2) development of integrated multimodal fusion models adaptable to dynamic patient states; (3) investigation of the temporal dynamics of delirium development using time-series analysis; and (4) application of causal inference methods to elucidate the relationships between risk factors and delirium. Superior prediction performance can be achieved by leveraging cutting-edge architectures (e.g. transformers) and parallel computing efficiencies to move beyond traditional machine learning. To enhance real-world adoption, future work should integrate Explainable AI tools such as Shapley Additive Explanations (SHAP) within EHR-based decision support systems, improving interpretability and mitigating subgroup disparities in localized risk assessment.