Using deep learning with attention mechanism for identification of novel temporal data patterns for prediction of ICU mortality. (2022)

• Record Type: Journal Article
• Authors: Ge, Wendong; Huh, Jin-Won; Park, Yu Rang; Lee, Jae-Ho; Kim, Young-Hak; Zhou, Guohai; Turchin, Alexander
• Is Part Of: Informatics in medicine unlocked. Volume 29(2022)
• Publication Date: 2022
• Subjects: Predictive modeling -- Temporal data -- Deep learning -- Recurrent neural networks -- Attention; Medical informatics -- Periodicals; 610.285
• DOI: 10.1016/j.imu.2022.100875
• Languages: English
• ISSNs: 2352-9148

Abstract

Abstract: Background and objectives: Changes in a patient's condition over time are a backbone of clinical decision making. However, most currently used methods for identification of patients in intensive care units (ICUs) at high risk for death do not make effective use of the temporal dimension of available data. We therefore conducted a study to determine whether longitudinal data analysis using recurrent neural networks (RNN) with attention mechanism can identify novel temporal data patterns predictive of adverse outcomes. Methods: We analyzed data on patients admitted to the Medical Intensive Care Unit (MICU) of Asan Medical Center between 2010 and 2017. Static (demographics, diagnoses, procedures, medications) and longitudinal (vitals, laboratory tests, Glasgow Coma Scale) variables were included in the analysis. We used an RNN model with long short-term memory (RNN-LSTM) with attention mechanism to identify and test novel data patterns predictive of ICU death. We also compared accuracy of prediction of ICU mortality between a logistic regression and RNN-LSTM models with and without attention . Results: Among 4896 patients admitted to the MICU, 548 (11.19%) died. RNN-LSTM model with attention identified several high-risk longitudinal variable patterns that were predictive of ICU mortality in a confirmatory analysis, including sustained low blood oxygen content (OR 2.33; 95% CI 1.16 to 4.70) and high frequency of serum sodium measurements (OR 1.27; 95% CI 1.04 to 1.56). RNN-LSTM models with and without attention achieved numerically, but not statistically significantly higher c-statistics for prediction of ICU mortality compared to logistic regression. Conclusions: RNN-LSTM model with attention identified novel temporal data patterns predictive of ICU mortality. These predictors were both statistically significant and clinically plausible, likely representing progressive respiratory failure (sustained low oxygen saturation) and close monitoring of a clinically deteriorating patient (frequent sodium measurements). Graphical abstract: Image 1 Highlights: Longitudinal data may allow detection of novel risk factors for adverse outcomes. RNN-LSTM with attention identified new data patterns predictive of ICU mortality. One predictive pattern category involved specific changes in values over time. Other predictive patterns involved measurement frequencies irrespective of values.