The authors review how artificial intelligence–based predictive models are being developed to identify patients at risk of delirium using electronic health record (EHR) data. Delirium is a frequent and serious complication among hospitalized patients—particularly older adults and those undergoing major surgery—and is associated with increased mortality, prolonged hospitalization, long-term cognitive decline, and higher healthcare costs. The review synthesizes 63 studies that applied machine-learning techniques—including logistic regression, random forests, gradient boosting algorithms, support vector machines, and neural networks—to detect patterns associated with delirium risk. Most models were trained on structured EHR variables such as demographics, comorbidities, medications, laboratory values, and perioperative factors, while a smaller number incorporated unstructured data such as clinical notes using natural language processing or multimodal approaches combining multiple data sources. Across the studies reviewed, machine-learning models demonstrated promising predictive performance, often capturing complex interactions among risk factors that are difficult to detect using traditional statistical models or bedside risk scores.
From a clinical and healthcare-system perspective, these technologies could significantly improve the prevention and management of delirium. Because delirium is frequently underrecognized and can fluctuate throughout hospitalization, automated predictive systems capable of continuously analyzing EHR data may help identify high-risk patients earlier than routine clinical assessments. Earlier detection could allow clinicians to implement targeted preventive strategies—such as medication optimization, environmental modifications, or closer monitoring—before delirium develops. At a population-health level, AI-based delirium prediction models could also improve surveillance and resource allocation in high-risk settings such as intensive care units and postoperative wards. However, the authors emphasize that several challenges remain before these models can be widely implemented, including variability in delirium definitions, limited integration of longitudinal and unstructured clinical data, and concerns regarding model generalizability and interpretability. Despite these limitations, the review suggests that AI-driven predictive systems have the potential to become valuable clinical decision-support tools, helping reduce the burden of delirium and improve outcomes in vulnerable hospitalized populations.
