Using Machine Learning to Identify Risk Scores For Delirium Under Different Clinical Settings

Abstract

Delirium is a common and acute neuropsychiatric syndrome, often overlooked in hospitals due to its diverse and fluctuating nature. Early identification of delirium patients is essential for timely interventions to prevent its progression and associated complications. This study aims to develop a machine learning-based predictive algorithm for incident delirium using electronic health record (EHR) data. Identifying important predictors for delirium is challenging, as per the psychiatric opinion. To achieve this objective, we first derived a dataset for intensive care unit (ICU) delirium from the Medical Information Mart for Intensive Care (MIMIC) data collection. Then, we identified the important clinical features from the derived dataset to extract them from the EHR charts of American University of Beirut Medical Center (AUBMC) patients. This approach allowed us to identify clinical features that may serve as markers for delirium and can facilitate early identification of delirium patients. We then trained various machine learning models, including traditional machine learning, and deep learning models, on the derived datasets to predict the delirium risk score for hospitalized and ICU patients. Our results show promising performance for the CatBoost model, which has the highest performance compared to other models. Ultimately, this predictive algorithm has the potential to improve delirium detection rates and streamline efficiency in hospital electronic systems, thereby enabling prompt interventions to prevent delirium progression and associated complications. Additionally, the outcomes of this study involved the recognition of clinical indicators for delirium in both ICU and hospital settings. These indicators could assist healthcare practitioners in pinpointing potential sources of delirium in their patients.