(382) Deep Learning Models for Continuous Mortality Prediction of ICU Patients

Introduction: Continuous ICU mortality predictors can help identify high-risk patients, guide interventions, and aid in resource allocation. Previous predictive models often used multiple clinical variables, which may not be consistently available or accurately measured in different hospital settings. Furthermore, these models primarily relied on 12-24 hours of prior patient data for prediction. To address these challenges, we propose deep learning models to continuously predict mortality risk utilizing minimal inputs from the last hour of patient data.

Methods: The dataset used for this work is a subset of the de-identified eICU dataset, containing data from over 2.1M patients, across over 300 hospitals, collected between 2007 and 2023 . Patients younger than 16 years old or with length of stay < 4 hours are excluded. Mortality rate within the eICU patients is 4.6%. The outcomes considered in this work are mortality within the next 12, 24 and 48 hours. The model inputs are six key signals: average systolic blood pressure, mean arterial pressure (MAP), mean respiratory rate, mean heart rate, age, and max Glasgow Coma Scale (GCS) within the last hour. Three deep learning models corresponding to each outcome were independently trained and tested. The dataset was randomly divided into 1.8 million patients for the training set, 150,000 for the validation set, and 150,000 for an independent test set.

Results: The deep learning models demonstrated strong performance in predicting mortality. The Area Under the Curve (AUC) for the test set was 0.883 for 12-hour mortality prediction, 0.868 for 24-hour mortality prediction, and 0.853 for 48-hour mortality prediction. With sensitivity of 90%, specificity for the 12-hours, 24-hours and 48-hours mortality was 67%, 61% and 59% respectively. Our analysis showed relatively uniform performance across gender, ethnicities, ICU types, and diagnosis groups with cerebrovascular accident/stroke (CVA) with highest AUC (0.91) and congestive heart failure (CHF) among the lowest AUC (0.77).

Conclusions: This study presents the development of novel deep learning models for continuous ICU mortality prediction using six simple signals, showing high AUC scores and potential for broad clinical use.