# Development and evaluation of an artificial intelligence-based electrocardiogram prediction model for emergency chest pain patients

## TL;DR

This paper presents an AI model that can quickly and accurately detect heart attacks from ECGs in emergency patients, but it struggles with other less clear conditions.

## Contribution

A novel AI model using a CNN with channel attention for rapid ECG-based detection of STEMI and NSTEMI in emergency chest pain patients.

## Key findings

- The model achieved high AUC values for STEMI (0.986) and NSTEMI (0.916) detection.
- Inference time was significantly faster than manual interpretation for STEMI diagnosis.
- Performance for unstable angina and aortic dissection was suboptimal with high sensitivity but low precision.

## Abstract

Rapid triage and etiological differentiation are critical for patients with acute chest pain in the emergency department. The 12-lead electrocardiogram (ECG), as a non-invasive, readily available, and cost-effective diagnostic modality, provides immediate information and serves as the first-line tool for clinical evaluation. However, ECG interpretation remains highly dependent on clinician expertise and is subject to inter-observer variability. Artificial intelligence (AI)-based analytical methods can deliver automated, consistent, and real-time assessment, thereby potentially enhancing diagnostic accuracy and facilitating timely clinical decision-making.

This study included 1,188 patients with acute chest pain who visited the emergency department in the Second Xiangya Hospital of Central South University, between March 2024 and March 2025. Standard 12-lead ECGs, clinical information, and final diagnoses were collected. After data preprocessing, a convolutional neural network (CNN) incorporating a channel attention mechanism was developed and trained. Model performance was assessed using accuracy, precision, recall, F1-score, area under the curve (AUC), and confusion matrices. Additionally, a blinded comparative evaluation was conducted against expert cardiologists.

The model demonstrated excellent discriminative capability for ST-elevation myocardial infarction (STEMI) and non-ST-elevation myocardial infarction (NSTEMI), with AUC values of 0.986 and 0.916, respectively. For STEMI, all performance metrics indicated superior diagnostic accuracy, and inference time was significantly shorter than manual interpretation (0.24 ± 0.08 s, p < 0.001). However, detection performance for unstable angina (UA) and aortic dissection (AD) remained suboptimal, characterized by high sensitivity but relatively low precision.

The deep learning model based on 12-lead ECGs enables rapid and reliable detection of STEMI and NSTEMI, highlighting its potential as a valuable clinical decision-support tool in emergency department. Nevertheless, the recognition of UA and AD remains limited due to non-specific or transient electrophysiological features.