AI-derived ejection fraction outperformed clinician measurements in a 27,382-patient cardiac MRI cohort
Cardiac MRI remains one of the most precise ways to quantify left-ventricular function, and ejection fraction still sits near the center of everyday risk stratification. In this multicentre retrospective study, artificial intelligence-based segmentation of left-ventricular ejection fraction performed better than clinician segmentation for predicting all-cause mortality. The signal was not dramatic in absolute terms, but it was consistent across multiple statistical approaches and large enough to matter in a cohort this size.
The study drew on 27,382 consecutive patients who underwent cardiac MRI in the UK and the USA. Most came from Barts Heart Centre in London, with an additional cohort from the University of Pittsburgh Medical Centre. Mortality follow-up came from national and local databases, allowing the investigators to examine long-term outcomes rather than a narrow snapshot of early events. Median follow-up reached 5.6 years, and 3,458 patients died during that period.
A single sentence captures the core result: AI segmented ejection fraction predicted death better than clinician-derived values.
Model fit and discrimination favored AI across the full cohort
The investigators compared prognostic models built from AI-measured versus clinician-measured left-ventricular volume metrics. For model fit, AI ejection fraction produced a lower Akaike Information Criterion than clinician ejection fraction, with an AIC of 57,936 versus 58,097. The difference of 161 points favored AI and indicates a materially better fit to the observed survival data.
Discrimination also improved, although more modestly. The C index was 0.669 for AI ejection fraction compared with 0.637 for clinician ejection fraction, an absolute difference of 0.033 with a 95% confidence interval of 0.026 to 0.041. In prognostic work, that kind of improvement does not suggest a complete rewrite of risk assessment, but it does show that automated segmentation extracted clinically relevant information that human segmentation missed or captured less precisely.
The incremental value of AI became even clearer when both measures were entered into nested models. AI ejection fraction significantly improved model fit when added to clinician ejection fraction, indicating that the automated measurement carried prognostic information beyond the standard workflow. In practical terms, the AI output was not merely substituting for human interpretation; it was adding signal.
Why small measurement differences can change prognostic performance
Ejection fraction is a blunt but durable clinical metric. Its appeal lies in simplicity: clinicians use it to gauge systolic function, guide therapy, and frame risk discussions. The problem is that the number depends on how the ventricular borders are traced, and segmentation variability can shift a patient across a threshold that alters management. Even small differences in contouring can change the final calculated volume and, in turn, the derived ejection fraction.
Deep-learning segmentation is built to reduce that variability. By applying a consistent algorithm to every scan, AI can standardize the measurement process across scanners, sites, and readers. In a dataset spanning two countries and multiple hospitals, that consistency is not just an engineering feature; it is a plausible explanation for stronger mortality prediction. A stable measurement pipeline is especially valuable when the downstream endpoint is hard clinical truth rather than a surrogate marker.
The study design also strengthens the relevance of the findings. Because patients were drawn from routine clinical CMR practice rather than a narrow trial population, the cohort likely captured a broad mix of indications and disease burdens. That breadth makes the mortality signal more credible for real-world imaging workflows, where ejection fraction is ordered for everything from cardiomyopathy to ischemic disease assessment.
What the mortality data suggest about clinical use
The event count was substantial. Among 27,382 patients, 3,458 deaths occurred over a median 5.6 years of follow-up. A dataset with that many outcomes is less vulnerable to random noise than small derivation cohorts, and it gives prognostic modeling room to separate signal from statistical clutter. The advantage seen with AI segmentation therefore rests on a large clinical foundation, not on a handful of endpoints.
Still, the result should be read as a comparison of prognostic performance, not as proof that AI segmentation alone should dictate patient management. Clinicians do not treat a mortality model; they treat the patient in front of them. The more immediate implication is that automated CMR analysis may improve the accuracy of the data feeding risk models, particularly in settings where manual contouring is time-consuming, variable, or impractical at scale.
The study also raises a methodological point that matters for imaging-based prognostics. If an automated algorithm can outperform expert-derived measurements on a core variable such as ejection fraction, then conventional human interpretation may not be the optimal reference standard for future predictive modeling. In a field increasingly shaped by machine learning, the benchmark is shifting from whether AI can match human readers to whether it can produce more useful clinical predictions.
The limitations are embedded in the design
This was a retrospective prognostic study, so the results describe association rather than intervention. The analysis also focused on mortality prediction, which is an important endpoint but not the only one that determines clinical value. Hospitalization, arrhythmia burden, treatment response, and quality of life were not part of the presented findings here, so the utility of AI segmentation for those outcomes remains untested in this dataset.
There is also an important separation between better statistical prediction and better bedside decision-making. A model can gain discrimination without immediately changing who gets treated, imaged, or followed more closely. For now, the strongest claim supported by the data is narrower: AI-derived left-ventricular ejection fraction from cardiac MRI improved prognostic performance for all-cause mortality relative to clinician segmentation in a large multicentre cohort.
In a field where ejection fraction has long been treated as a familiar constant, this study shows that the way the number is produced still matters. The next question is whether that advantage translates into different clinical decisions when AI is embedded directly into routine reporting workflows.