Assessing different machine learning algorithms to identify traumatic cardiac arrest: proof of concept

Objective
Traumatic cardiac arrest differs from non-traumatic regarding epidemiology. This study evaluated five machine learning classifiers' ability to identify traumatic cardiac arrest in the Danish Cardiac Arrest Registry, which currently relies on manual review.

Methods
This retrospective study employed split sampling to train and test models using medical records of cardiac arrest patients from 2016 to 2021. Data were preprocessed, resampled, and classified using five classification models. Shapley Additive Explanations values were used to explain the models' predictions except for the BERT model. 

Results
30,171 medical records included, with 985 involving traumatic cardiac arrests . The histogram gradient boosting model achieved the highest F1-score of 0.62, with precision (positive predictive value) of 0.61 and recall (sensitivity) of 0.62. The BERT model demonstrated a recall of 0.96, a precision of 0.21, and an F1-score of 0.35. For histogram gradient boosting and random forest, the increasing presence of “car”, “thorax”, and “head” was linked to trauma. Further error analysis revealed that these models exhibited age-related bias by implicitly learning that traumatic cardiac arrest is more common in younger populations and less frequent in older populations. 

Conclusion
This study represents the first steps towards using machine learning algorithms to support manual validation of the Danish Cardiac Arrest Registry. Histogram gradient boosting yielded the best overall results; however, the BERT model demonstrated significantly improved recall in detecting trauma cases. Hence, applying the BERT model to the validation would substantially reduce the present workload.