Development and Validation of a Machine Learning Model for Early Prediction of Delirium in Intensive Care Units Using Continuous Physiological Data: Retrospective Study

• DOI: 10.2196/59520
• Publication Date: 2025-04-04T10:50:49Z
• Authors: Chanmin Park; Changho Han; Su Kyeong Jang; Hyungjun Kim; Sora Kim; Byung Hee Kang; Kyoungwon Jung; Dukyong Yoon
• Keywords: Medical record; Vital signs

Delirium in intensive care unit (ICU) patients poses a significant challenge, affecting patient outcomes and health efficiency. Developing an accurate, real-time prediction model for delirium represents advancement critical care, addressing needs timely intervention resource optimization ICUs. We aimed to create novel machine learning ICU using only continuous physiological data. developed models integrating routinely available clinical data, such as age, sex, monitoring device outputs, ensure practicality adaptability diverse settings. To confirm the reliability of determination records, we prospectively collected results Confusion Assessment Method (CAM-ICU) evaluations performed by qualified investigators from May 17, 2021, December 23, 2022, determining Cohen κ coefficients. Participants were included study if they aged ≥18 years at admission, had CAM-ICU, data least 4 hours before diagnosis or nondiagnosis. The development cohort Yongin Severance Hospital (March 1, 2020, January 12, 2022) comprised 5478 records: 5129 (93.62%) records 651 training 349 (6.37%) 163 internal validation. For temporal validation, used 4438 same hospital (January 28, 31, reflect potential seasonal variations. External validation was 670 Ajou University 2022 September 2022). evaluated algorithms (random forest [RF], extra-trees classifier, light gradient boosting machine) selected RF final based on its performance. utility, decision curve analysis pattern during stay performed. coefficient between labels generated nurses verified researchers 0.81, indicating reliable CAM-ICU results. Our showed robust performance (area under receiver operating characteristic [AUROC]: 0.82; area precision-recall [AUPRC]: 0.62) maintained accuracy (AUROC: 0.73; AUPRC: 0.85). supported effectiveness 0.84; 0.77). Decision positive net benefit all thresholds, gradual increase scores actual time approached. measured variables, including waveforms. demonstrates predicting delirium, with consistent across various scenarios. uses noninvasive making it applicable wide range patients, minimal additional risk.