An Artificial Intelligence-Based Alarm Strategy Facilitates Management of Acute Myocardial Infarction

Abstract

(1) Background: While an artificial intelligence (AI)-based, cardiologist-level, deep-learning model for detecting acute myocardial infarction (AMI), based on a 12-lead electrocardiogram (ECG), has been established to have extraordinary capabilities, its real-world performance and clinical applications are currently unknown. (2) Methods and Results: To set up an artificial intelligence-based alarm strategy (AI-S) for detecting AMI, we assembled a strategy development cohort including 25,002 visits from August 2019 to April 2020 and a prospective validation cohort including 14,296 visits from May to August 2020 at an emergency department. The components of AI-S consisted of chest pain symptoms, a 12-lead ECG, and high-sensitivity troponin I. The primary endpoint was to assess the performance of AI-S in the prospective validation cohort by evaluating F-measure, precision, and recall. The secondary endpoint was to evaluate the impact on door-to-balloon (DtoB) time before and after AI-S implementation in STEMI patients treated with primary percutaneous coronary intervention (PPCI). Patients with STEMI were alerted precisely by AI-S (F-measure = 0.932, precision of 93.2%, recall of 93.2%). Strikingly, in comparison with pre-AI-S (N = 57) and post-AI-S (N = 32) implantation in STEMI protocol, the median ECG-to-cardiac catheterization laboratory activation (EtoCCLA) time was significantly reduced from 6.0 (IQR, 5.0-8.0 min) to 4.0 min (IQR, 3.0-5.0 min) (p < 0.01). The median DtoB time was shortened from 69 (IQR, 61.0-82.0 min) to 61 min (IQR, 56.8-73.2 min) (p = 0.037). (3) Conclusions: AI-S offers front-line physicians a timely and reliable diagnostic decision-support system, thereby significantly reducing EtoCCLA and DtoB time, and facilitating the PPCI process. Nevertheless, large-scale, multi-institute, prospective, or randomized control studies are necessary to further confirm its real-world performance.

Keywords: acute myocardial infarction; alarm system; artificial intelligence; deep learning; electrocardiogram.

Figure 1

The workflow and development of the AI-based alarm strategy. All visits were divided into two groups with or without chest pain for STEMI detection. The remaining visits with hsTnI were used for subsequent NSTEMI detection. The AUROC curve and PRROC curve were generated by the highest probability of STEMI/NSTEMI prediction by our AI and hsTnI.

Figure 2

STEMIs for which alarms failed to be raised by the AI-based alarm strategy. (A,B) are both post-resuscitation ECGs with STEMI prediction scores of 62.1% and 70.8%, respectively. (C) Hyperacute T presentation with prompt initiation of PPCI with a STEMI prediction score of 66.9%. (D) Inferior wall STEMI with deep Q in III, aVF and reciprocal ST-T change in I, aVL, V2 with STEMI prediction score of 57.7%.

Figure 3

STEMIs for which alarms were raised by the AI-based alarm strategy but for which diagnosis by front-line physicians were delayed. (A) A 36-year-old tall man awaiting chest X-ray to rule out pneumothorax. (B) An 82-year-old woman presenting with left shoulder pain initially. (C) A 75-year-old woman presenting with acute epigastric pain, nausea, and dizziness awaiting an X-ray to rule out a perforated peptic ulcer. (D) A 54-year-old man with a chronic kidney disease awaiting labs to rule out hyperkalemia.

Figure 4

The DtoB time metrics before and after the AI-based alarm strategy. (A) The comparison of the median DtoE, EtoCCLA, CCLAtoCCLD, CCLDtoB, EtoCCLD, EtoB, and DtoB times before and after AI-S implantation. (B) The relative time ratio between each time component over DtoB time.

Figure 5

Stratified analyses for the performance of AI-based alarm strategy in AMI, STEMI, and NSTEMI in the prospective cohort. The 95% CI of F-measures was calculated based on 10,000 bootstrapping experiments. For the standard analysis, the F-measure of AMI detection was 0.851 (95% CI: 0.798–0.896), the F-measure of STEMI detection was 0.932 (95% CI: 0.878–0.973), and the F-measure of NSTEMI detection was 0.701 (95% CI: 0.594–0.789).