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

Q: What is the clinical evidence from this study?

Study design: Observational. Population: Out-of-hospital cardiac arrest (n=30171). Intervention: Machine learning classifiers (Histogram gradient boosting and BERT) vs. Manual review (standard practice). Primary outcome: F1-score for identifying traumatic cardiac arrest (F1-score 0.63, 95% CI 0.57-0.67).

Q: What does this research mean for the field?

Machine learning models can accurately identify traumatic cardiac arrest from medical records, with histogram gradient boosting providing the highest overall F1-score (0.62-0.63) while BERT provides high sensitivity (0.96 recall) suitable for screening. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Signe Amalie Wolthers; Mehdi Parviz; Caroline Kamuk Ostenfeldt; Annette Kjær Ersbøll; Theo Walther Jensen; Mathias Geldermann Holgersen; Lars Bredevang Andersen; Stig Nikolaj Fasmer Blomberg; Søren Mikkelsen; Christian Torp-Pedersen; Helle Collatz Christensen

doi:10.1016/j.annepidem.2026.110245

August 11, 2026Annals of EpidemiologyOpen Access

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

View Full Paper

Ask AI

Bookmark

Share

Key result

Histogram gradient boosting achieves a peak 0.63 F1-score for identifying traumatic cardiac arrest.

F1-score 0.63
95% CI 0.57-0.67
n=30,171

Graphical abstract published by Annals of Epidemiology — Graphical abstract · Published by Annals of Epidemiology · CC BY-NC-ND 4.0

Why the study?

Traumatic cardiac arrest differs epidemiologically from non-traumatic cardiac arrest, and the Danish Cardiac Arrest Registry currently relies on manual review to identify these cases.

Do machine learning algorithms accurately identify traumatic cardiac arrest in prehospital medical records?

Key Points

To evaluate the capability of five machine learning classifiers to identify traumatic cardiac arrest cases within the Danish Cardiac Arrest Registry to support manual validation.
Analyzed 30,171 medical records (including 985 traumatic cardiac arrest cases) from the Danish Cardiac Arrest Registry between 2016 and 2021 using split sampling.
Evaluated five classification models (histogram gradient boosting, random forest, logistic regression, linear support vector machine, and BERT) combined with resampling and class weighting techniques.
Applied Shapley Additive Explanations (SHAP) to assess model interpretability and performed error analysis across age, sex, and geographic region.
Histogram gradient boosting achieved the highest overall balance with an F1-score of 0.62, precision of 0.61, recall of 0.62, and an AUC of 0.94 (95% CI: 0.92–0.95).
At a 0.975 threshold, the BERT model achieved a recall of 0.96 (95% CI: 0.93–0.98), precision of 0.21 (95% CI: 0.19–0.24), and an F1-score of 0.35 (95% CI: 0.31–0.39).
Word features such as "car", "thorax", and "head" strongly predicted trauma, while error analysis revealed age-related bias with higher false-negative rates in older patients.

Study Design

Type

Observational (n=30,171)

Multicenter

Yes

Structured PICO

Do machine learning algorithms accurately identify traumatic cardiac arrest in prehospital medical records?

P

Population

30,171 patients with out-of-hospital cardiac arrest from the Danish Cardiac Arrest Registry, including 985 traumatic cases, used to evaluate machine learning classifiers.

E

Exposure

Machine learning classifiers (Histogram Gradient Boosting, BERT, Logistic Regression, Linear SVM, Random Forest) applied to unstructured text data from prehospital medical records

C

Comparator

Manual review/validation (current standard)

O

Outcome

Identification of traumatic cardiac arrest (measured by F1-score, precision, and recall)

Main Result

Effect estimate: F1-score 0.63 (95% CI 0.57-0.67)

Machine learning models, particularly BERT for high recall and Histogram Gradient Boosting for overall balance, can support but not replace manual validation of traumatic cardiac arrest in registries.

Limitations

SHAP attributions may be misleading for the BERT model due to context-dependent token embeddings.
Models were trained and tested on Danish data and may perform differently in other settings, requiring external validation.
Reliance on contextual language introduces potential for misclassification.
Data collected in an acute setting carries a risk of typographical errors.
Class imbalance was handled differently across the traditional classifiers and BERT, creating a potential asymmetry in performance.
Models exhibited age-related bias by implicitly learning that traumatic cardiac arrest is more common in younger populations
Low precision translates into a higher number of false positives and associated manual review workload
Models cannot replace manual validation based solely on text data

Implication

May streamline traumatic cardiac arrest identification in registries; leaves open need for prospective validation before replacing manual review.

Abstract

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.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Wolthers et al. (2026) conducted an observational in Out-of-hospital cardiac arrest (n=30,171). Machine learning classifiers (Histogram gradient boosting and BERT) vs. Manual review (standard practice) was evaluated on F1-score for identifying traumatic cardiac arrest (F1-score 0.63, 95% CI 0.57-0.67). The histogram gradient boosting model achieved the highest F1-score of 0.63 for identifying traumatic cardiac arrest, while the BERT model demonstrated a recall of 0.96, highlighting a trade-off between overall performance and case detection.

synapsesocial.com/papers/6aa6fa1fcd36d0f46f00a199 https://doi.org/10.1016/j.annepidem.2026.110245

View Full Paper

Ask AI

Bookmark

Share

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

Key Points

Study Design

Structured PICO

Main Result

Limitations

Implication

Abstract

Discussion

Cite This Study

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

Key Points

Study Design

Structured PICO

Main Result

Limitations

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading