Treatment Recommendations for Clinical Deterioration on the Wards: Development and Validation of Machine Learning...

Area of Science:

• Clinical informatics
• Artificial intelligence in healthcare
• Patient monitoring

Background:

• Clinical deterioration in general ward patients leads to increased morbidity and mortality.
• Early intervention is crucial for improving outcomes in high-risk patients.
• Machine learning (ML) shows promise in identifying deterioration risk but less in guiding treatment recommendations.

Purpose of the Study:

• Establish ML performance benchmarks for predicting 10 common clinical deterioration interventions.
• Compare various ML models to identify optimal approaches for these prediction tasks.

Main Methods:

• Utilized a multicenter dataset of 2480 general ward patient encounters with clinical deterioration, identified via an early warning score.
• Employed manual chart review to label encounters with gold-standard lifesaving treatment data.
• Trained and validated multiple ML models (elastic net logistic regression, gradient boosted machines, long short-term memory, stacking ensemble) to predict intervention needs, using AUROC as the primary metric.

Main Results:

• Model performance varied by task and approach, with AUROCs generally ranging from 0.7 to 0.9.
• Antiarrhythmics were the most predictable intervention (mean AUROC 0.866), while anticoagulants were the least (mean AUROC 0.660).
• Gradient boosted machines often performed best individually, and stacking ensembles matched or exceeded individual model performance.

Conclusions:

• ML models demonstrate variable but generally high performance in predicting lifesaving treatments for deteriorating patients.
• These models can be integrated with early warning scores to provide clinicians with timely, actionable treatment recommendations.
• A significant portion of evaluated patients did not receive timely treatment, indicating an opportunity for ML to reduce treatment latency.