Assessing Different Machine Learning Algorithms to Identify Traumatic Cardiac Arrest

Area of Science:

• Emergency Medicine
• Data Science
• Public Health

Background:

• Traumatic cardiac arrest has distinct epidemiological characteristics compared to non-traumatic cases.
• The Danish Cardiac Arrest Registry currently relies on manual review for case classification.
• Automating the identification of traumatic cardiac arrest can improve registry efficiency.

Purpose of the Study:

• To evaluate the effectiveness of five machine learning classifiers in identifying traumatic cardiac arrest.
• To compare the performance of different models for supporting manual validation of the registry.

Main Methods:

• A retrospective analysis of 30,171 cardiac arrest patient records from 2016-2021.
• Training and testing five classification models, including histogram gradient boosting and BERT.
• Utilizing Shapley Additive Explanations for model interpretability, except for the BERT model.

Main Results:

• Histogram gradient boosting achieved the highest F1-score (0.62), with precision of 0.61 and recall of 0.62.
• The BERT model achieved a recall of 0.96 but a lower F1-score (0.35) and precision (0.21).
• Models identified trauma indicators like 'car', 'thorax', and 'head', but also showed age-related bias.

Conclusions:

• Machine learning shows promise for supporting the manual validation of the Danish Cardiac Arrest Registry.
• Histogram gradient boosting offers the best overall performance, while BERT excels in trauma case recall.
• Implementing the BERT model could significantly decrease the manual workload for registry validation.