A Machine Learning Approach to Identifying Delirium from Electronic Health Records

Abstract

ABSTRACTBackgroundDespite the well-known impact of delirium on long-term clinical outcomes, identification of delirium in electronic health records (EHR) remains difficult due to inadequate assessment or documentation of delirium. The purpose of this research is to present a classification model that identifies delirium using retrospective EHR data. The classification model would support the additional identification of delirium cases otherwise undocumented during routine practice.MethodsDelirium was confirmed with the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU). Age, sex, Elixhauser comorbidity index, drug exposures, and diagnoses were used as features to train the logistic regression and multi-layer perceptron models. The clinical notes from the EHR were parsed to supplement the features that were not recorded in the structured data. The model performance was evaluated with a 5-fold cross-validation area under the receiver operating characteristic curve (AUC).ResultsSeventy-six patients (17 cases and 59 controls) with at least one CAM-ICU evaluation result during ICU stay from January 30, 2018 to February 20, 2018 were included in the model. The multi-layer perceptron model achieved the best performance in identifying delirium; mean AUC of 0.967 ± 0.019. The mean positive predictive value (PPV), mean negative predicted value (NPV), mean sensitivity, and mean specificity of the MLP model were 0.9, 0.88, 0.56, and 0.95, respectively.ConclusionA simple classification model showed a mean AUC over 0.95. This model promises to identify delirium cases with EHR data, thereby enable a sustainable infrastructure to build a retrospective cohort of delirium in the ICU. The cohort would be useful for the evaluation of long-term sequelae of delirium in ICU.