Smartwatch ECG and artificial intelligence in detecting acute coronary syndrome compared to traditional 12-lead ECG

Abstract

Background: Acute coronary syndromes (ACS) require prompt diagnosis through initial electrocardiograms (ECG), but ECG machines are not always accessible. Meanwhile, smartwatches offering ECG functionality have become widespread. This study evaluates the feasibility of an image-based ECG analysis artificial intelligence (AI) system with smartwatch-based multichannel, asynchronous ECG for diagnosing ACS.

Methods: Fifty-six patients with ACS and 15 healthy participants were included, and their standard 12-lead and smartwatch-based 9-lead ECGs were analyzed. The ACS group was categorized into ACS with acute total occlusion (ACS-O(+), culprit stenosis ≥ 99 %, n = 44) and ACS without occlusion (ACS-O(-), culprit stenosis 70 % to < 99 %, n = 12) based on coronary angiography. A deep learning-based AI-ECG tool interpreting 2-dimensional ECG images generated probability scores for ST-elevation myocardial infarction (qSTEMI), ACS (qACS), and myocardial injury (qMI: troponin I > 0.1 ng/mL).

Results: The AI-driven qSTEMI, qACS, and qMI demonstrated correlation coefficients of 0.882, 0.874, and 0.872 between standard and smartwatch ECGs (all P < 0.001). The qACS score effectively distinguished ACS-O(±) from control, with AUROC for both ECGs (0.991 for standard and 0.987 for smartwatch, P = 0.745). The AUROC of qSTEMI in identifying ACS-O(+) from control was 0.989 and 0.982 with 12-lead and smartwatch (P = 0.617). Discriminating ACS-O(+) from ACS-O(-) or control presented a slight challenge, with an AUROC for qSTEMI of 0.855 for 12-lead and 0.880 for smartwatch ECGs (P = 0.352).

Conclusion: AI-ECG scores from standard and smartwatch-based ECGs showed high concordance with comparable diagnostic performance in differentiating ACS-O(+) and ACS-O(-). With increasing accessibility smartwatch accessibility, they may hold promise for aiding ACS diagnosis, regardless of location.

Keywords: Acute Coronary Syndrome; Artificial Intelligence; Diagnostic Performance; Smartwatch ECG.

Fig. 1

Image-based AI-ECG analysis on standard and smartwatch ECGs: a representative case. The application of the QCG™ analyzer, AI-ECG analysis system, on a standard 12-lead ECG (3x4 format) and a 9-channel smartwatch ECG images (9x1 format, asynchronous) of the same patient is demonstrated. The system generated probability scores for STEMI, ACS, and myocardial injury (qSTEMI, qACS, and qMI) on a scale from 0.0 to 1.0. For this case, the AI-ECG scores indicated prominent elevations across both ECG formats. Coronary angiography confirmed a near-total occlusion of the right coronary artery, classifying the patient into the ACS-O(+) group. ACS, acute coronary syndrome; ACS-O(+), ACS with acute coronary artery occlusion; ACS-O(−), ACS without coronary artery occlusion; AI, artificial intelligence; MI, myocardial injury; NSTEMI, Non-ST-elevation myocardial infarction; QCG, quantitative ECG; STEMI, ST-elevation myocardial infarction.

Fig. 2

Correlations between the QCG scores from standard and smartwatch ECGs The scatter plots demonstrate high concordance between standard and smartwatch ECGs for (A) qSTEMI, (B) qACS, and (C) qMI scores. Spearman’s correlation coefficient (ρ) is provided for each QCG score. Other abbreviations are as in Fig. 1.

Fig. 3

Bland-Altman plots of the QCG scores across standard and smartwatch ECGs Bias and the limit of agreement for (A) qSTEMI, (B) qACS, and (C) qMI scores are presented as scatter points and dashed lines. For clarity, the averaged QCG score from standard and smartwatch ECGs is on the x-axis, on a logarithmic scale. The dashed horizontal lines represent the 95% confidence intervals for the differences in QCG scores between standard and smartwatch ECGs. Abbreviations are as in Fig. 1.

Fig. 4

Performance of the QCG scores derived from standard and smartwatch ECGs The diagnostic performances of the QCG scores are illustrated as the area under the receiver operating characteristic curves (AUROC). For comparison, the performance of the cardiologic expert consensus is represented on the AUROC curve, with each result shown as a single dot. (A) The qSTEMI, both from standard and smartwatch ECG images, effectively distinguished ACS-O(+) from the normal control. (B) The qACS successfully differentiated the combined ACS-O(+) and ACS-O(−) group from the control. (C) Discriminating ACS-O(+) from ACS-O(−) or control group presented a slight challenge. Abbreviations are as in Fig. 1.