Deep continual multitask out-of-hospital incident severity assessment from changing clinical features

Area of Science:

• Machine Learning
• Artificial Intelligence
• Emergency Medicine

Background:

• Machine learning models in emergency medical triage are vulnerable to performance degradation due to evolving input features over time.
• Covariate shifts, including feature emergence and disappearance, pose significant challenges to model accuracy and reliability.

Purpose of the Study:

• To evaluate novel deep continual learning pipelines for maintaining and improving model performance in dynamic emergency medical triage scenarios.
• To address challenges posed by changing feature domains and parameter updates in time-varying datasets.

Main Methods:

• Analysis of 1,414,575 emergency events from 2009-2019 to identify and quantify covariate shifts.
• Development and assessment of continual learning strategies including static/dynamic domain approaches and parameter updating methods (from-scratch, fine-tuning, cumulative, rehearsal, Elastic Weight Consolidation).

Main Results:

• The optimal combination of dynamic feature domain and Elastic Weight Consolidation (EWC) improved F1-scores by up to 5.9% for life-threatening situations, 18.6% for response delay, and 2.4% for jurisdiction.
• Continual learning approaches outperformed non-continual methods, showing gains up to 5.4% in critical situations and 11% in response delay.

Conclusions:

• Proposed continual learning methods effectively mitigate the negative impact of covariate shifts in emergency medical triage.
• These advancements enhance the reliability and accuracy of machine learning models in real-world, evolving emergency medical systems.