JMIR Publications

ABSTRACT

Background:

Predicting short-term mortality in critically ill patients is crucial for timely intervention in intensive care units (ICUs). Traditional scoring systems like SAPS and APACHE lack real-time adaptability and rely on static data. To address these limitations, this study developed and validated a dynamic risk prediction model using longitudinal and irregular electronic medical record (EMR) data from multicenter cohorts.

Objective:

This study aimed to develop a dynamic, real-time risk prediction model for ICU patients, leveraging longitudinal and irregular EMR data. The primary objective was to improve prediction accuracy and interpretability over traditional scoring systems, enabling timely clinical interventions through continuous mortality risk assessments and enhanced generalizability across multicenter datasets.

Methods:

A time-aware bidirectional attention-based LSTM (TBAL) model was developed using EMR data from the MIMIC-IV and eICU-CRD databases, covering 176,344 ICU stays. The model utilized dynamic variables such as vital signs, lab results, and medications, updated hourly, to perform static and dynamic mortality risk assessments. External cross-validation demonstrated its generalizability across datasets, with performance evaluated using AUROC, AUPRC, and subgroup sensitivity analyses. Integrated Gradients (IG) enhanced interpretability by identifying key predictors influencing mortality risk.

Results:

The TBAL model achieved AUROCs of 95.9 (MIMIC-IV) and 93.3 (eICU-CRD) for 12-hour to 1-day mortality prediction, outperforming traditional models. In dynamic tasks, the AUROC reached 93.6 (MIMIC-IV) and 91.9 (eICU-CRD), improving over time with hourly updates. Cross-validation demonstrated robust performance across databases, with AUROCs of 0.813 (MIMIC-IV to eICU-CRD) and 0.761 (eICU-CRD to MIMIC-IV). Variable importance rankings revealed key predictors like lactate and vasopressor use, with consistency across datasets.

Conclusions:

The TBAL model provides dynamic, interpretable, and real-time mortality predictions for ICU patients, outperforming traditional scoring systems. Its adaptability to evolving clinical conditions highlights its potential as a decision-support tool, though further validation in randomized trials is warranted.