11 November 2026 · Medows
AI Heart Scans: The 4.3-Point Disagreement
A Lancet study of 27,382 patients found AI read ejection fraction 4.3 points lower than clinicians. Consistency, not insight, was the edge.
A cardiac MRI report lands on the round with an ejection fraction on it. The doctor reading it treats that figure as a property of the heart. A study published on November 2 is a reminder that it is also a property of whoever measured it.
The study
Researchers from City St George's, University of London and UCL published a retrospective, multicentre prognostic study in The Lancet Digital Health. They took cardiac MRI scans from 27,382 patients in the UK and the US being investigated for known or suspected heart conditions. Over a median 5.6 years, 3,458 died (about 13%). The question was narrow and good: whose measurement of left ventricular ejection fraction tracked mortality better, the AI segmentation or the clinician's.
The AI won. Slightly.
The margin is smaller than the headline
On ejection fraction alone, discrimination was 0.669 for the AI versus 0.637 for clinicians. Then add age and sex, two things you know before the patient enters the scanner, and the gap collapses to 0.004 (0.781 versus 0.777). Statistically significant, per the authors. Clinically, close to nothing.
Most coverage ran with "AI better than clinicians." The honest version is narrower: the AI's rendering of this one number carries a little more prognostic signal, and nearly all of that edge vanishes once you account for two variables already on the chart's front page.
The 4.3 points nobody led with
Here is the finding I would actually want on a round. Median ejection fraction was 57.8% measured by AI and 62.1% measured by clinicians. Same hearts, same scans, a consistent offset of 4.3 percentage points.
That is not sloppiness. Consistency is precisely what the authors claim as the tool's strength. Lead author Dr George Joy said the findings "suggest AI can measure this important marker with remarkable consistency and precision." Consistent and offset are not contradictory. This AI reads low, and it does it reliably.
But clinical thinking runs against fixed numbers, and a systematic offset means two readers of one scan can land on opposite sides of a line nobody moved. Medical Daily's write-up of the paper makes the same observation: fixed cutoffs are what turn a consistent difference like this into something that could matter.
So the value on the report is not a number. It is a number plus a reader. The report almost never says which reader.
Where it broke
The error rate was about 1% at the patient level, affecting 0.08% of images (6 segmentation errors in a random sample of 600). Those errors were considerably more frequent in scans with extreme ejection fraction values, particularly very low ones, which is to say the sickest patients. In hypertrophic cardiomyopathy, clinician measurement did slightly better, possibly because the software did not fully capture unusual muscle features.
That shape is familiar: most reliable in the middle of the distribution, least reliable at the edges, and the edges are where you were already worried.
What it does not show
It is retrospective. It did not test whether an AI-first read changed a decision or kept anyone alive. There was no cause-of-death data, and performance across ethnic groups was not evaluated. It was funded by the UK's National Institute for Health and Care Research and the British Heart Foundation, and several authors disclosed financial ties to MycardiumAI, the company associated with the tool.
The researchers are measured about it themselves. Senior author Dr Rhodri Davies: "AI is not replacing clinicians, but it can provide faster, more consistent measurements at low-cost." On speed the case is strong: 2 to 3 seconds per patient, against roughly 13 minutes for expert analysis in the team's earlier work.
Why provenance belongs next to the value
This is the part that shapes how we build Medows. A point tool hands you an answer and disappears. A doctor comparing today's 57.8% against last year's 62.1% needs to know whether the heart changed or the reader did, and that is unanswerable unless the workspace carried provenance forward with the number.
Verifiable AI usually gets discussed as citations under a recommendation. This study points at something more mundane: the instrument behind each figure on the chart, kept visible, so a trend reads as a trend rather than as an artefact of who was on shift. The AI here was not smarter than the clinicians. It was steadier. Steadiness is worth having, and only worth trusting if you can see where each number came from.
Sources
- AI measurements of heart scans more strongly linked to patient survival than clinician measurements (EurekAlert, 2 November 2026)
- Performance of artificial intelligence versus clinician left-ventricular segmentation of cardiac MRI for mortality prediction, The Lancet Digital Health
- AI beat clinicians by a small margin at predicting death risk from heart MRI measurements (Medical Daily, 9 November 2026)
- AI measurements of heart scans more strongly linked to patient survival (Medical Xpress, 2 November 2026)
Medows is a clinical AI workspace for the doctor on rounds. Learn more or write to alapan@medows.ai / alapanx@gmail.com.