How AI-Powered Systems Are Transforming Delirium Care in Elderly Patients
Machine learning and artificial intelligence models are being examined for their roles in predicting and detecting delirium risk in hospitalized older individuals.
A comprehensive review study published examines the roles of artificial intelligence, machine learning, and large language models in predicting, detecting, and preventing delirium in hospitalized older adults.
General Characteristics of Delirium Syndrome
Delirium is defined as an acute neuropsychiatric syndrome characterized by disturbances in attention, awareness, cognition, and arousal. Affecting approximately one-fifth of hospitalized elderly patients, this condition is observed at much higher rates in intensive care units.
Diagnosis and Diagnostic Challenges
Clinical heterogeneity makes it difficult to distinguish hypoactive delirium, which manifests with withdrawal, low activity, or drowsiness, from hyperactive forms. Research shows that emergency medicine physicians overlook a significant proportion of delirium cases.
Evaluation Methods Used
Conducting systematic evaluations and using validated tools such as CAM, CAM-ICU, and 4AT alongside clinical judgment are of great importance. These validated tools contribute to more accurate detection of patients' conditions.
Review Methodology and Research Scope
A primary search was conducted in PubMed and MEDLINE databases in August 2026 using the terms delirium, artificial intelligence, and older adults. This search was supported by targeted searches for prediction, detection, natural language processing, prevention, frailty, and large language models.
Machine Learning Models and Performance
Machine learning models aim to identify statistical patterns to predict delirium risk prior to clinical onset. A gradient boosting model in a study involving over 18,000 hospital admissions demonstrated higher discriminative power compared to traditional scores.
Challenges Encountered in Clinical Use
Although a prospectively applied random forest algorithm achieved high discriminative power, it encountered poor calibration issues during clinical use. This demonstrates that high discrimination capability does not always imply clinically reliable risk predictions.