Multicentre validation of point-of-care screening tool for heart failure: single-lead ECG recorded by smart stethoscope predicts low ejection fraction using artificial intelligence. (14th October 2021)

• Record Type: Journal Article
• Authors: Bachtiger, P; Scott, F; Park, S; Petri, C; Padam, P S; Sahemey, H; Dumea, B; Ribeiro, M; Alquero, R; Bual, N; Cheung, W S; Rana, B; Keene, D; Plymen, C M; Peters, N S
• Is Part Of: European heart journal. Volume 42(2021)Supplement 1
• Publication Date: 2021-10-14
• Subjects: Artificial Intelligence (Machine Learning, Deep Learning); Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/eurheartj/ehab724.3071
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background/Introduction: Artificial intelligence (AI) applied to 12-lead ECG can identify left ventricular ejection fraction (EF) ≤35% with a sensitivity and specificity of 86.3% and 85.7%, respectively. Whether AI algorithms trained on 12-lead can accurately predict EF from single-lead ECGs (recorded by a smart stethoscope) remains unknown. This could facilitate point-of-care screening for low EF during routine clinical examination. Purpose: First independent multicentre real-world UK National Health Service (NHS) prospective validation of 12-lead-ECG-trained AI algorithm applied to single-lead ECG recorded by a smart stethoscope, with AI algorithm tuned to detect EF ≤40%. Methods: Prospective recruitment of unselected patients attending for echocardiography across six urban NHS hospital sites (UK). In addition to transthoracic echocardiogram (routine care), all participants had 15 seconds of supine, single-lead ECG recorded at six different positions (figure), encompassing standard anatomical positions for cardiac auscultation. A convolutional neural network (CNN) previously trained on 35, 970 independent pairings of 12-lead-ECG and echocardiograms was retrained to use the single-lead ECG as input. Accuracy of CNN detection of low EF (binary ≤40%) is reported at a threshold of 0.5 against gold-standard; echo-determined percentage EF. Results: Among 353 patients recruited (mean age 63±17; 58% male, 43.1% non-white), 309 (87.5%) had an EF >40%, and 44 (12.5%) had EF ≤40%. The best single recording position in isolation was position 3 (sensitivity 57.9% [42.2–73.6], specificity 86.3% [82.2–90.3]). Taking any of the six positions performed during the examination as predicting EF ≤40%, this achieved a sensitivity of 81.2% and specificity of 61.5%. Conclusion(s): In this first prospective multicentre validation study the retrained AI algorithm reliably detected low EF from single-lead ECGs acquired using a novel ECG-enabled stethoscope in standard auscultation positions. The ability to identify patients with possible low EF during routine physical examination addresses a significant unmet clinical need in point-of-care ruling in/out of heart failure, and has potential to provide broader population-level screening for asymptomatic cardiovascular disease. Funding Acknowledgement: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): National Institute of Health Research, Accelerated Access Collaborative & NHSX: Artificial Intelligence in Health & Social Care Award