Performance of artificial intelligence versus clinician left-ventricular segmentation of cardiac MRI for mortality prediction: a retrospective, multicentre, prognostic study.
Authors
Affiliations (14)
Affiliations (14)
- Cardiovascular Clinical Academic Group, City St George's University of London, London, UK; Cardiology Department, St George's Hospital, London, UK.
- Institute of Cardiovascular Science, University College London, London, UK; Barts Heart Centre, St Bartholomew's Hospital, London, UK; Imperial College Healthcare NHS Trust, London, UK.
- Institute of Cardiovascular Science, University College London, London, UK.
- Barts Heart Centre, St Bartholomew's Hospital, London, UK.
- Institute of Cardiovascular Science, University College London, London, UK; Barts Heart Centre, St Bartholomew's Hospital, London, UK; Medical Research Council Unit for Lifelong Health and Ageing, University College London, London, UK.
- Institute of Cardiovascular Science, University College London, London, UK; Barts Heart Centre, St Bartholomew's Hospital, London, UK.
- Cardiology Department, Hospital Prof Doutor Fernando Fonseca, Amadora, Portugal.
- National Amyloidosis Centre, Royal Free Hospital, London, UK.
- National Heart, Lung, and Blood Institute, National Institutes of Health, Bethesda, MA, USA.
- National Heart and Lung Institute, Imperial College London, London, UK.
- Department of Biomedical Engineering, King's College London, London, UK.
- Minneapolis Heart Institute, United Hospital, Minneapolis, MN, USA.
- Medical Research Council Unit for Lifelong Health and Ageing, University College London, London, UK.
- Institute of Cardiovascular Science, University College London, London, UK. Electronic address: [email protected].
Abstract
Ejection fraction, a key measure of cardiac function, is used in health care for clinical decision making and risk stratification. We hypothesised that a deep-learning artificial intelligence (AI) segmentation of ejection fraction obtained from cardiac MRI (CMR) outperforms expert clinician segmentation for the prediction of all-cause mortality. In this retrospective, multicentre, prognostic study, we studied a cohort of 27 382 consecutive patients who underwent cardiac MRI (CMR) in the UK and the USA. Patients were referred for CMR from nine hospitals to Barts Heart Centre (BHC), London. Patients were identified from the reporting archive at BHC as having undergone a CMR between Oct 1, 2014 and Dec 31, 2019. A further cohort of patients was also identified from University of Pittsburgh Medical Centre (UPMC), Pittsburgh, PA, USA, as having undergone a CMR between June 1, 2010 and March 31, 2016. Mortality data were obtained from national and local databases. AI segmentation was done on all CMR studies, and Cox proportional hazards models were fitted to compare the prognostic value of AI-measured versus clinician-measured left-ventricular volume metrics. Model fit was compared using the Akaike Information Criterion (AIC), discrimination using C indices and the incremental value of AI over clinician segmentation using likelihood-ratio tests of nested models. 27 382 patients received a baseline CMR (26 106 patients from BHC and 1276 from UPMC). The median follow-up was 5·6 years (IQR 4·4-6·9), during which 3458 (13%) of 27 382 patients died. AI ejection fraction showed better model fit (AIC 57 936 vs 58 097, ΔAIC 161 favouring AI) and marginally better discrimination (C index 0·669 vs 0·637; ΔC 0·033 [95% CI 0·026-0·041]). AI ejection fraction significantly improved model fit when added to clinician ejection fraction (χ<sup>2</sup>=168, p<0·0001), with only slight improvement when clinician ejection fraction was added to AI (χ<sup>2</sup>=7·5, p=0·023). After adjustment for age and sex, AI-derived ejection fraction showed better model fit (AIC 55 684 vs 55 742; ΔAIC 59) and had marginally better discrimination (C index of 0·781 vs 0·777; ΔC=0·004 [95% CI 0·002-0·006]). AI-derived ejection fraction shows marginally superior prediction for mortality than clinician-derived ejection fraction in patients who had CMR. Findings support the potential utility of AI for first-line ejection-fraction measurement with clinician oversight, although this should be investigated in prospective studies. National Institute for Health and Care Research Clinical Lectureship and the British Heart Foundation Accelerator Award.