UCL Discovery
UCL home » Library Services » Electronic resources » UCL Discovery

Diagnosis and risk stratification in hypertrophic cardiomyopathy using machine learning wall thickness measurement: a comparison with human test-retest performance

Augusto, JB; Davies, RH; Bhuva, AN; Knott, KD; Seraphim, A; Alfarih, M; Lau, C; ... Moon, JC; + view all (2020) Diagnosis and risk stratification in hypertrophic cardiomyopathy using machine learning wall thickness measurement: a comparison with human test-retest performance. The Lancet Digital Health 10.1016/s2589-7500(20)30267-3. (In press). Green open access

[img]
Preview
Text
Lancet DH - Machine Learning MWT measurement in HCM.pdf - Published version

Download (1MB) | Preview

Abstract

Background: Left ventricular maximum wall thickness (MWT) is central to diagnosis and risk stratification of hypertrophic cardiomyopathy, but human measurement is prone to variability. We developed an automated machine learning algorithm for MWT measurement and compared precision (reproducibility) with that of 11 international experts, using a dataset of patients with hypertrophic cardiomyopathy. Methods: 60 adult patients with hypertrophic cardiomyopathy, including those carrying hypertrophic cardiomyopathy gene mutations, were recruited at three institutes in the UK from August, 2018, to September, 2019: Barts Heart Centre, University College London Hospital (The Heart Hospital), and Leeds Teaching Hospitals NHS Trust. Participants had two cardiovascular magnetic resonance scans (test and retest) on the same day, ensuring no biological variability, using four cardiac MRI scanner models represented across two manufacturers and two field strengths. End-diastolic short-axis MWT was measured in test and retest by 11 international experts (from nine centres in six countries) and an automated machine learning method, which was trained to segment endocardial and epicardial contours on an independent, multicentre, multidisease dataset of 1923 patients. Machine learning MWT measurement was done with a method based on solving Laplace’s equation. To assess test–retest reproducibility, we estimated the absolute test–retest MWT difference (precision), the coefficient of variation (CoV) for duplicate measurements, and the number of patients reclassified between test and retest according to different thresholds (MWT >15 mm and >30 mm). We calculated the sample size required to detect a prespecified MWT change between pairs of scans for machine learning and each expert. Findings: 1440 MWT measurements were analysed, corresponding to two scans from 60 participants by 12 observers (11 experts and machine learning). Experts differed in the MWT they measured, ranging from 14·9 mm (SD 4·2) to 19·0 mm (4·7; p<0·0001 for trend). Machine learning-measured mean MWT was 16·8 mm (4·1). Machine learning precision was superior, with a test–retest difference of 0·7 mm (0·6) compared with experts, who ranged from 1·1 mm (0·9) to 3·7 mm (2·0; p values for machine learning vs expert comparison ranging from <0·0001 to 0·0073) and a significantly lower CoV than for all experts (4·3% [95% CI 3·3–5·1] vs 5·7–12·1% across experts). On average, 38 (64%) patients were designated as having MWT greater than 15 mm by machine learning compared with 27 (45%) to 50 (83%) patients by experts; five (8%) patients were reclassified in test–retest by machine learning compared with four (7%) to 12 (20%) by experts. With a cutoff point of more than 30 mm for implantable cardioverter-defibrillator, three experts would have changed recommendations between tests a total of four times, but machine learning was consistent. Using machine learning, a clinical trial to detect a 2 mm MWT change would need 2·3 times (range 1·6–4·6) fewer patients. Interpretation: In this preliminary study, machine learning MWT measurement in hypertrophic cardiomyopathy is superior to human experts with potential implications for diagnosis, risk stratification, and clinical trials.

Type: Article
Title: Diagnosis and risk stratification in hypertrophic cardiomyopathy using machine learning wall thickness measurement: a comparison with human test-retest performance
Open access status: An open access version is available from UCL Discovery
DOI: 10.1016/s2589-7500(20)30267-3
Publisher version: https://doi.org/10.1016/s2589-7500(20)30267-3
Language: English
Additional information: This is an Open Access article under the CC BY-NC-ND 4.0 license. https://creativecommons.org/licenses/by-nc-nd/4.0/
UCL classification: UCL
UCL > Provost and Vice Provost Offices
UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences
UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Population Health Sciences
UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Population Health Sciences > Institute of Cardiovascular Science
UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Population Health Sciences > Institute of Cardiovascular Science > Clinical Science
UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Population Health Sciences > Institute of Cardiovascular Science > Population Science and Experimental Medicine
UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Population Health Sciences > Institute of Cardiovascular Science > Population Science and Experimental Medicine > MRC Unit for Lifelong Hlth and Ageing
UCL > Provost and Vice Provost Offices > School of Life and Medical Sciences > Faculty of Population Health Sciences > Institute of Health Informatics
URI: https://discovery.ucl.ac.uk/id/eprint/10117011
Downloads since deposit
16Downloads
Download activity - last month
Download activity - last 12 months
Downloads by country - last 12 months

Archive Staff Only

View Item View Item