Neural networks for ischaemia detection: revolution or red herring? A systematic review and meta-analysis

Rob Brisk, RR Bond, D Finlay, James McLaughlin, Alicja Jasinska-Piadlo, Michael Jennings, David McEneaney

Research output: Contribution to conferencePosterpeer-review

Abstract

Background: Artificial neural networks (ANNs) are machine learning (ML) algorithms that have been investigated as a means of automatically detecting acute myocardial ischaemia from electrocardiogram (ECG) signals since the early 1990s. In recent years, there has been renewed interest in ANNs as the basis for deep learning (DL), which is cited as the leading edge in artificial intelligence (AI). The purpose of this review is to ascertain what progress has been made in detecting acute myocardial infarction (AMI) from ECG signals using ANNs and DL to date.

Methods: The titles, abstracts and keywords of full-text articles on Medline, Scopus and Web-of-Science were searched using the following terms: ((myocardial infarction OR ischaemia) AND (neural network OR deep learning) AND (electrocardiogram OR ECG)). The searches were performed in November 2019. Abstracts of all search results were screened. All studies specifically pertaining to the use of ANNs to detect AMI from ECG signals were reviewed in full.
Data was extracted and a quality score was constructed around the QUADAS-2 framework. Studies with a quality score above 4 whose endpoint was relevant to the review question were
included in the meta-analysis. To account for different balances between sensitivity and specificity, the meta-analysis concentrated on the F1 score (the harmonic mean of the sensitivity and the positive predictive value).

Results: The search process generated 196 results; 45 studies were reviewed in full; 27 were excluded from the meta-analysis due to quality concerns; 6 studies were excluded because their
end points did not align with the meta-analysis. The 12 studies included in the meta-analysis were published between 1994 and 2019; 3 were prospective; 9 were retrospective. A total of 8480 test subjects were included. Disease prevalence was 23%. The average F1 score (with
95% confidence intervals) was 0.79 (0.72-0.85). A population weighted average F1 score was calculated at 0.83. Further sub-analyses were undertaken (see attached figures).

Conclusions: AMI detection by ANN analysis of ECG signals is likely to be a promising research avenue but the current high-quality evidence base in this area is sparse. Of the studies reviewed in full, 60% did not meet quality criteria. Linear regression analysis of quality scores
revealed that average quality has decreased over time. Only 11% of studies reviewed were undertaken prospectively. A minority of studies discussed issues regarding transparency of ANNs, which is likely to be important for future applications.
Original languageEnglish
Publication statusPublished (in print/issue) - 28 Apr 2021
Event45th International Society for Computerized Electrocardiology -
Duration: 28 Apr 20211 May 2021
Conference number: 45
https://cdn.ymaws.com/www.isce.org/resource/resmgr/hauck/2021_conference/here/isce_abstract_booklet_5.26.2.pdf

Conference

Conference45th International Society for Computerized Electrocardiology
Abbreviated titleISCE
Period28/04/211/05/21
Internet address

Keywords

  • neural networks
  • deep learning
  • ECGs

Fingerprint

Dive into the research topics of 'Neural networks for ischaemia detection: revolution or red herring? A systematic review and meta-analysis'. Together they form a unique fingerprint.

Cite this