Researchers at Barts Heart Centre in London have found that ejection fraction measured by a deep-learning segmentation tool predicted all-cause death slightly better than the same measure taken by expert clinicians, in a retrospective study of 27,382 patients published in The Lancet Digital Health on 2 November 20261. The work drew on cardiac MRI scans from the UK and the USA and was funded by the National Institute for Health and Care Research and the British Heart Foundation1.
What the study found
Ejection fraction is the standard measure of how well the heart pumps and feeds directly into treatment decisions and risk scoring1. The team ran AI segmentation on every scan and compared the prognostic value of AI-derived and clinician-derived left-ventricular volume metrics using Cox proportional hazards models1.
Of the 27,382 patients, 26,106 were scanned at Barts after referral from nine hospitals between 1 October 2014 and 31 December 2019, and 1,276 at University of Pittsburgh Medical Centre between 1 June 2010 and 31 March 20161. Over a median follow-up of 5.6 years, 3,458 patients (13%) died1.
The AI measure fitted the mortality data better, with an Akaike Information Criterion of 57,936 against 58,097 for clinicians, and gave a C index of 0.669 against 0.6371. Adding the AI value to the clinician value improved the model substantially (χ2 = 168, p<0.0001), while adding the clinician value to the AI value gave only a slight gain (χ2 = 7.5, p=0.023)1.
Once age and sex were accounted for, the gap narrowed sharply. The C index was 0.781 for AI against 0.777 for clinicians, a difference of 0.004 (95% CI 0.002 to 0.006), although model fit still favoured AI, with an AIC of 55,684 against 55,7421.
What it means for the market
The authors read the results as support for using AI as the first-line method for measuring ejection fraction, with a clinician overseeing the output1. For NHS trusts and other health systems running cardiac MRI services, that points towards a workflow in which software produces the routine measurement and specialists review it, rather than one in which AI sits alongside manual contouring as a second opinion1.
For suppliers of cardiac imaging AI, the study offers outcome-linked evidence, which is a stronger basis for a business case than agreement with human readers alone1. The size of the advantage matters, though. After adjustment, the improvement in discrimination was marginal, so the case for adoption rests more on whether AI can deliver at least clinician-level prognostic accuracy at scale than on a large clinical gain1. How much reporting time or cost the AI approach saves, and which software would be used, is not yet set out.
Caveats
The study was retrospective, and the great majority of patients came from a single UK centre, with only 1,276 of the 27,382 patients in the US cohort1. The authors state that the approach should be tested in prospective studies before it changes practice1.
Sources
This briefing summarises publicly available research and reporting for information only. It is not medical, investment or legal advice. Follow the sources listed for the primary research and reporting.