Key Points
- A study of 27,382 cardiac MRI patients in the UK and US found AI-derived ejection fraction measurements were slightly but significantly better than experienced clinicians at identifying patients at increased risk of death.
- The AI tool analysed each scan in about 2–3 seconds versus roughly 13 minutes for expert clinician analysis, with an approximately 1% patient-level error rate, supporting a model where AI measures first and clinicians review and interpret.
- The study, published in The Lancet Digital Health, did not show that deploying AI in routine care reduces deaths, so adoption depends on further validation across populations, scanners and clinical settings.
How AI Outperformed Cardiologists at Reading 27,382 Heart Scans
Artificial intelligence can measure a key indicator of heart function from cardiac MRI scans slightly more accurately than experienced clinicians — and dramatically faster — according to a large study published in The Lancet Digital Health. Researchers from City St George's, University of London, and University College London analysed scans from 27,382 patients in the United Kingdom and the United States who were being investigated for known or suspected heart conditions. Over an average follow-up of more than five years, 3,458 deaths were recorded, allowing the team to test how well AI-derived measurements predicted survival compared with clinician readings.
The measurement at the centre of the study is ejection fraction — the proportion of blood pumped out of the heart's main chamber with each contraction. It is one of the most important indicators cardiologists use, informing decisions about heart failure medication, implantable cardiac devices, procedures to restore blood flow, surgical risk and monitoring during chemotherapy. Calculating it manually requires a clinician to trace the heart's changing blood volume across the cardiac cycle in MRI images, a task that takes an expert roughly 13 minutes per patient. The AI tool did it in about 2–3 seconds.
The AI measurements were slightly but significantly better at identifying patients at increased risk of death, and the association held after adjusting for age and sex. Lead author George Joy, an NIHR Clinical Lecturer at City St George's, described the improvement as small but meaningful across thousands of patients, particularly for those at highest risk. The tool was not flawless: it produced errors in roughly 1% of patients, either overestimating or underestimating ejection fraction, which is why the researchers propose that AI performs the routine measurement while clinicians review every result and retain responsibility for interpreting it alongside symptoms, medical history and other tests.
The study did not establish that introducing AI into routine care reduces deaths or improves outcomes — only that its measurements track survival risk more closely. The team's next line of inquiry is whether combining AI-derived cardiac MRI measurements with electrocardiogram data could detect heart muscle abnormalities earlier, before more substantial damage develops.
At a Glance
| Research Institutions | City St George's, University of London and UCL UK universities that conducted the cardiac MRI and AI comparison study. |
| Lead Author | George Joy, MBBS, MRCP NIHR Clinical Lecturer who framed the AI-first, clinician-review workflow proposal. |
| Journal | The Lancet Digital Health Peer-reviewed publication where the study findings appeared. |
| Study Size | 27,382 patients Patients in the UK and US undergoing investigation for known or suspected heart conditions. |
| Follow-Up | Average of more than 5 years Period during which 3,458 deaths were recorded and linked to scan measurements. |
| Key Measurement | Ejection fraction Share of blood pumped from the heart's main chamber per contraction; guides treatment decisions. |
| AI Analysis Speed | Approximately 2–3 seconds per patient Compared with around 13 minutes for expert clinician analysis in prior research by the same team. |
| AI Error Rate | Approximately 1% of patients Errors led to ejection fraction being overestimated or underestimated, requiring clinician review. |
| Clinical Uses of Ejection Fraction | Heart failure medication, implantable devices, surgical risk, chemotherapy monitoring Decisions that depend on accurate measurement of heart function. |
| Future Research Focus | Combining cardiac MRI AI with electrocardiogram (ECG) data Aim is to detect heart muscle abnormalities before substantial damage develops. |
What the 3-Second Scan Means for Hospitals, Radiologists and AI Vendors
Why a 1% Accuracy Edge Matters at Population Scale
The headline number — AI slightly outperforming clinicians — sounds marginal until it is placed against the study's scale. With 27,382 patients and 3,458 deaths, even a small gain in precision shifts how many high-risk patients are correctly flagged for closer monitoring or earlier intervention. Joy's own framing is instructive: the gain is small per patient but meaningful across thousands, especially at the highest-risk end of the distribution. The statistically significant association after adjusting for age and sex strengthens the case that the AI is capturing something real about heart function rather than demographic noise.
The Real Prize Is Throughput, Not Just Accuracy
The 2–3 second analysis time versus roughly 13 minutes for expert manual measurement is the commercially and operationally decisive figure. Cardiac MRI reporting is a bottleneck in many health systems, and demand for diagnostic imaging keeps rising. If AI handles the routine measurement, the clinician's time shifts from tracing images to interpreting results — a higher-value activity that is harder to automate. For hospitals, that is a workforce productivity argument, not just a technology argument. It also points to standardisation: AI measurements are consistent across scans in a way that human measurements, which vary between observers, are not.
Where This Leaves AI Cardiac Imaging Vendors
This study is a validation signal for companies building cardiac MRI analysis tools. Peer-reviewed evidence in The Lancet Digital Health that AI-derived ejection fraction predicts mortality better than clinician measurement is the kind of proof point that procurement committees and regulators look for. But the study also sets the bar vendors must clear: a 1% patient-level error rate is acceptable only inside a workflow where a clinician reviews every result. Any vendor marketing fully autonomous cardiac MRI interpretation is selling ahead of the evidence. The realistic near-term product is a first-line measurement tool with mandatory clinician sign-off.
What the Study Does Not Show
The researchers are explicit that the work examined the relationship between AI-derived measurements and subsequent mortality — it did not test whether deploying AI in clinical practice actually reduces deaths or improves treatment outcomes. That distinction matters for anyone weighing adoption. Faster, more accurate measurement is a necessary input to better care, not proof of it. Questions also remain about performance across different patient populations, clinical settings and MRI scanner systems, all of which the team flags as areas for further research.
The ECG Combination Bet
The research team's stated next step — combining AI-derived cardiac MRI measurements with electrocardiogram data — is strategically interesting. If imaging and electrical data together can flag heart muscle abnormalities before substantial damage develops, the value proposition shifts from efficiency (doing the same work faster) to earlier detection (changing what care is possible). That is a harder claim to prove but a far larger prize, and it explains why the team is pursuing it rather than resting on the mortality-prediction result.
What Health Systems and Clinicians Should Do With This Evidence
The evidence supports AI as a first-line measurement tool for cardiac MRI ejection fraction, with clinicians reviewing every result — not autonomous AI interpretation.
For Hospital and Imaging Department Leaders (0–6 months)
- Model the workflow change explicitly: the study's comparison is roughly 2–3 seconds of AI analysis versus around 13 minutes of expert manual measurement per scan. Calculate what that time releases in your own cardiac MRI reporting backlog.
- Plan for a review step, not a replacement step. The tool erred in approximately 1% of patients, overestimating or underestimating ejection fraction, so clinician sign-off must be built into any pathway.
- Ask vendors for validation data across multiple MRI scanner systems and patient populations, since the study itself flags these as open questions.
For Cardiologists and Clinical Leads (6–24 months)
- Expect the division of labour to shift toward interpretation: the proposed model has AI doing routine measurement while clinicians assess results alongside symptoms, medical history and other tests.
- Watch for the team's follow-up work combining cardiac MRI AI with electrocardiogram data, which targets earlier detection of heart muscle abnormalities rather than just faster measurement.
For Health System and Policy Decision-Makers (2–5 years)
- Treat this study as evidence of measurement accuracy and speed, not of improved patient outcomes. The researchers state the work did not establish that AI in routine practice reduces deaths or improves treatment.
- Fund or require implementation trials that measure outcomes — not just diagnostic concordance — before committing to system-wide rollout.
Metrics Worth Watching
- Publication of the team's ECG-plus-imaging research results.
- Any regulatory clearances for cardiac MRI AI tools positioned as first-line measurement with clinician review.
- Implementation studies reporting patient outcomes, not only accuracy and time savings.
Risk & Opportunity Assessment
| Commercial Risk | Low | The study validates rather than threatens the cardiac imaging market; the main commercial risk sits with AI vendors whose products cannot match the 1% error rate or the multi-scanner validation the research implies is needed. |
| Competitive Risk | Medium | AI cardiac imaging vendors now have a peer-reviewed benchmark in The Lancet Digital Health to compete against; tools without equivalent mortality-prediction evidence will struggle in procurement against validated competitors. |
| Regulatory Risk | Medium | The proposed AI-first, clinician-review workflow requires regulatory acceptance of AI as a first-line measurement tool, and the study's 1% patient-level error rate means regulators are likely to mandate clinician sign-off rather than autonomous use. |
| Reputation Risk | Medium | Hospitals adopting AI measurement without the clinician review step the researchers insist on risk patient-safety criticism if the roughly 1% error rate leads to misclassified ejection fraction and inappropriate treatment decisions. |
| Technology Disruption | High | A 2–3 second AI measurement replacing roughly 13 minutes of expert manual analysis restructures cardiac MRI reporting work, shifting clinician time from measurement to interpretation and standardising a measurement that currently varies between observers. |
| Commercial Opportunity | High | Validation in a major peer-reviewed journal opens procurement conversations for AI cardiac MRI tools, and the team's planned combination of imaging AI with electrocardiogram data points to an earlier-detection market beyond measurement efficiency. |
Comments 0