Automated ECG Report as a Factor in the Clinical Decision Pathway for Acute Chest Pain in the Emergency Department

TL;DR

This study shows how combining traditional ECG analysis with AI can speed up diagnosis for patients with chest pain in emergency rooms.

Contribution

Integrates GPT-4 with the Glasgow algorithm to classify ECG reports for acute chest pain in emergency departments.

Findings

The model achieved 85.9% overall accuracy in classifying ECG reports.

High F1 scores for normal ECGs (0.93) and STEMI (0.80), but lower for new arrhythmias (0.45).

The model showed strong discrimination between STEMI and other categories (AUC=0.91-0.92).

Abstract

Background

Electrocardiographic analysis algorithms have consistently evolved, becoming essential tools for physicians in diverse settings, particularly in assessing patients with acute chest pain. Moving forward, it is crucial to classify unstructured automated ECG reports into clinically relevant outcomes using advanced large language models. This approach holds significant potential to enhance an accelerated clinical decision pathway in clinical settings.

Objective

This study aims to integrate automated electrocardiogram algorithms with advanced machine learning techniques, enhancing the classification of ECG reports within emergency department settings. Specifically, it investigates how natural language processing can augment traditional methods to accelerate the electrocardiographic-directed management of acute chest pain.

Methods

Employing a retrospective observational dataset from Rashid Hospital, Dubai, spanning from June 2022 to August 2022, we analyzed 860 ECGs from patients presenting with acute chest pain. The ECGs were categorized into four classes, namely, STEMI, NSTEMI, normal ECG, and new arrhythmia using a hybrid model that combines the established Glasgow algorithm with a large language model, GPT-4. The Glasgow algorithm produced structured text inputs, which were then classified by GPT-4 using few-shot prompting (temperature = 0.2, top_p=1.0).

Results

The model demonstrates high predictive accuracy for normal ECGs, achieving an F1 score of 0.93, followed by STEMI with an F1 score of 0.80. New arrhythmias, however, present more challenges, reflected by the lowest F1 score of 0.45. Notably, the model excels in discriminating between STEMI and normal ECGs (AUC=0.92) and between STEMI and new arrhythmias (AUC=0.91). Overall accuracy was 85.9% (95% CI: 0.816-0.895)

Conclusion

The findings suggest that leveraging deep learning alongside traditional algorithms can significantly improve the rapid classification of ECGs, supporting accelerated decision-making pathways in clinical practice.

Figures