Point-of-care screening for heart failure with reduced ejection fraction using artificial intelligence during ECG-enabled stethoscope examination in London, UK: a prospective, observational, multicentre study

Abstract

Background: Most patients who have heart failure with a reduced ejection fraction, when left ventricular ejection fraction (LVEF) is 40% or lower, are diagnosed in hospital. This is despite previous presentations to primary care with symptoms. We aimed to test an artificial intelligence (AI) algorithm applied to a single-lead ECG, recorded during ECG-enabled stethoscope examination, to validate a potential point-of-care screening tool for LVEF of 40% or lower.

Methods: We conducted an observational, prospective, multicentre study of a convolutional neural network (known as AI-ECG) that was previously validated for the detection of reduced LVEF using 12-lead ECG as input. We used AI-ECG retrained to interpret single-lead ECG input alone. Patients (aged ≥18 years) attending for transthoracic echocardiogram in London (UK) were recruited. All participants had 15 s of supine, single-lead ECG recorded at the four standard anatomical positions for cardiac auscultation, plus one handheld position, using an ECG-enabled stethoscope. Transthoracic echocardiogram-derived percentage LVEF was used as ground truth. The primary outcome was performance of AI-ECG at classifying reduced LVEF (LVEF ≤40%), measured using metrics including the area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity, with two-sided 95% CIs. The primary outcome was reported for each position individually and with an optimal combination of AI-ECG outputs (interval range 0-1) from two positions using a rule-based approach and several classification models. This study is registered with ClinicalTrials.gov, NCT04601415.

Findings: Between Feb 6 and May 27, 2021, we recruited 1050 patients (mean age 62 years [SD 17·4], 535 [51%] male, 432 [41%] non-White). 945 (90%) had an ejection fraction of at least 40%, and 105 (10%) had an ejection fraction of 40% or lower. Across all positions, ECGs were most frequently of adequate quality for AI-ECG interpretation at the pulmonary position (979 [93·3%] of 1050). Quality was lowest for the aortic position (846 [80·6%]). AI-ECG performed best at the pulmonary valve position (p=0·02), with an AUROC of 0·85 (95% CI 0·81-0·89), sensitivity of 84·8% (76·2-91·3), and specificity of 69·5% (66·4-72·6). Diagnostic odds ratios did not differ by age, sex, or non-White ethnicity. Taking the optimal combination of two positions (pulmonary and handheld positions), the rule-based approach resulted in an AUROC of 0·85 (0·81-0·89), sensitivity of 82·7% (72·7-90·2), and specificity of 79·9% (77·0-82·6). Using AI-ECG outputs from these two positions, a weighted logistic regression with l2 regularisation resulted in an AUROC of 0·91 (0·88-0·95), sensitivity of 91·9% (78·1-98·3), and specificity of 80·2% (75·5-84·3).

Interpretation: A deep learning system applied to single-lead ECGs acquired during a routine examination with an ECG-enabled stethoscope can detect LVEF of 40% or lower. These findings highlight the potential for inexpensive, non-invasive, workflow-adapted, point-of-care screening, for earlier diagnosis and prognostically beneficial treatment.

Funding: NHS Accelerated Access Collaborative, NHSX, and the National Institute for Health Research.

Conflict of interest statement

Declaration of interests We declare no competing interests.

Copyright © 2022 The Author(s). Published by Elsevier Ltd. This is an Open Access article under the CC BY 4.0 license. Published by Elsevier Ltd.. All rights reserved.

Figures

Figure 1

Schematic of ECG-enabled stethoscope and AI-ECG Illustration of anatomical positions for auscultation and position-specific angulation (vector) of ECG-enabled stethoscope; and flow diagram of raw ECG data to cloud-based CNN for interpretation of AI-ECG, with illustration of how raw outputs are classified according to adjustable (optimised) threshold. Anatomical images adapted from BioRender. AI=artificial intelligence. CNN=convolutional neural network. LVEF=left ventricular ejection fraction.

Figure 2

Study profile TTE=transthoracic echocardiogram. LVEF=left ventricular ejection fraction. NHS=National Health Service. *Three hospitals and four community centres.

Figure 3

Receiver operating characteristic curves detection of reduced LVEF Data are shown for the single-best performing position (pulmonary), rule-based optimal combination of two positions (pulmonary and handheld), and exploratory logistic regression model with l2 regularisation using AI-enabled ECG outputs from optimal combination of two positions. AUROC=area under the receiver operating characteristic curve. LR=logistic regression. LVEF=left ventricular ejection fraction.

Figure 4

Convolutional neural network's sensitivity and specificity to detect LVEF ≤40% Data are tabulated across a range stratified into age bands by sex, and by non-White ethnicity, using results from the pulmonary position at the threshold maximising the sum of sensitivity and specificity. The diagnostic OR and associated 95% CI is shown for each group and for the overall study sample. For sensitivity, data presented in brackets represents the number of patients in each group who had a positive result, with the denominator the number of patients with LVEF ≤40%. For specificity, this is the number of patients with a negative result, with the denominator patients with LVEF>40%. OR=odds ratio. LVEF=left ventricular ejection fraction.