Predicting Future Organ Support Needs Using Longitudinal Emergency Department Data: A Proof-of-Concept Study.

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Existing early warning scores (EWS) often rely on single physiological snapshots, potentially limiting their sensitivity in predicting organ support needs.
• Predicting organ support or death (OSD) may offer greater clinical utility than predicting mortality or transfer alone.
• There is a need for EWS that can effectively utilize longitudinal clinical data to capture patient status changes.

Purpose of the Study:

• To develop and compare novel predictive models using longitudinal data to identify patients at risk of OSD within 48 hours of admission.
• To evaluate the performance of these models against existing EWS, such as the National Early Warning Score 2 (NEWS2).

Main Methods:

• Retrospective analysis of adult emergency department encounters at a quaternary academic medical center.
• Development of transformer-based neural networks, XGBoost, and elastic-net regression models using longitudinal clinical data.
• Comparison of model performance using Area Under the Receiver Operating Characteristic Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), sensitivity, and specificity.

Main Results:

• A transformer-based neural network achieved the highest performance (AUROC 0.84, AUPRC 0.20), outperforming NEWS2 in sensitivity (0.78 vs. 0.61) while maintaining adequate specificity.
• XGBoost and elastic-net regression models also demonstrated improved sensitivity compared to NEWS2.
• The composite outcome of OSD occurred in 1.7% of the study cohort within 48 hours.

Conclusions:

• Longitudinal data analysis, rather than cross-sectional snapshots, may better predict critical illness progression and organ support needs.
• Machine learning models, particularly transformer networks, show promise in enhancing the predictive accuracy of early warning scores.
• Organ support is a relevant marker for critical illness, and improved prediction can inform timely interventions.