What if the most important number in cardiology has been hiding in a test we already run on almost everyone?

Left ventricular ejection fraction - LVEF, for those who prefer their medical jargon abbreviated - measures how much blood your heart's main pump actually pushes out per beat. Normal is 55-70%. Below 40%, and you're in heart failure territory. It drives decisions about medications, implantable defibrillators, surgery, and basically whether your cardiologist sleeps well at night. The catch: measuring it requires an echocardiogram. That's a trained sonographer, specialized equipment that costs more than your car, a 30-minute imaging session, and a cardiologist to interpret the whole thing. In much of the world, that infrastructure simply doesn't exist.

Meanwhile, the humble electrocardiogram - ten electrodes, five minutes, available in nearly every clinic on earth - just sits there recording squiggly lines nobody thought contained structural information.

Turns out those squiggly lines have been holding out on us.

Teaching a Neural Network to Read Between the Lines

A team from Massachusetts General Hospital and the Broad Institute at MIT just published a study in NPJ Digital Medicine that trained convolutional and probabilistic neural networks to estimate LVEF directly from ECG data (Thambiraj et al., 2026). The dataset was not small: 191,941 patients, 236,623 paired ECG/echocardiogram recordings. They built two flavors - an ECG-only model and a hybrid that also ingests structured clinical features like demographics and lab values.

The population-level ECG-only model achieved a mean absolute error of 7.71%. Not bad for extracting a number that normally requires an ultrasound machine. But here's where it gets interesting.

The "It Knows You" Upgrade

The real headline is personalization. When the models were adapted to individual patients - leveraging their prior ECG history rather than treating every heart like a stranger - the mean absolute error dropped to 5.98%. That's a meaningful jump. Your cardiologist eyeballing an echo has inter-observer variability in that same ballpark.

The probabilistic neural network component deserves a nod here. Instead of just spitting out a number and walking away, the model provides uncertainty estimates. It basically says, "I think this patient's LVEF is 48%, and I'm pretty confident about that" or "I think it's 42%, but honestly I'm squinting." In a clinical setting, that kind of calibrated honesty is worth its weight in contrast dye.

For detecting outright systolic dysfunction (LVEF at or below 40%), the model hit an AUC of 0.88, sensitivity of 0.92, and a negative predictive value of 0.98. That last number is the clinically juicy one: if the model says you're fine, there's a 98% chance you actually are. As screening tools go, that's a very effective bouncer.

Standing on the Shoulders of Squiggly Lines

This work doesn't exist in a vacuum. Mayo Clinic's landmark EAGLE trial showed that AI-flagged ECGs increased diagnosis of low ejection fraction by 32% in a randomized study of over 22,000 patients (Attia et al., Nature Medicine, 2021). A Taiwanese group built ECG12Net - 82 convolutional layers deep - and achieved an AUC of 0.95 for detecting EF below 35%, with the AI-derived EF actually outperforming standard measurements at predicting major cardiac events (Chen et al., J. Personalized Medicine, 2022).

Even more ambitious: MIT's PULSE-HF model, published in Lancet eClinicalMedicine earlier this year, doesn't just detect current low EF - it predicts whether your ejection fraction will decline below 40% within the next year, using a single ECG lead (Yau et al., 2026). One lead. From a signal you could theoretically capture with a smartwatch.

The field is sprinting from "what's your EF right now?" to "what will your EF be next year?" - and the input is a test that costs roughly the price of a decent sandwich.

The Gap Between Lab and Bedside

Before anyone starts replacing echo labs with laptop carts, some cold water. No AI-ECG model for ejection fraction has received FDA clearance yet. Most studies, including this one, are retrospective - meaning the model never had to make a real-time call that affected a real patient's care. External validation across diverse populations remains patchy. And the personalization advantage, while impressive, requires prior ECG data for a given patient - it won't help much on a first visit.

The 0.98 negative predictive value is excellent for screening, but the 7-8% absolute error in population-level models means individual predictions can occasionally be off by clinically significant margins. The uncertainty quantification helps, but it also means the model will sometimes shrug, which isn't ideal when you need an answer.

Why This Actually Matters

Heart failure affects over 64 million people worldwide. In sub-Saharan Africa, some countries have fewer than one cardiac imaging device per million people. A tool that converts a $20, five-minute ECG into a reasonable LVEF estimate - with built-in confidence intervals - could fundamentally change who gets screened and how early dysfunction gets caught.

The personalization angle from Thambiraj et al. pushes the field toward something genuinely useful: models that improve as they learn more about your heart specifically, not just hearts in general. Pair that with wearable ECG devices and remote monitoring, and you start to see a future where cardiac function tracking is as routine as checking your step count - except, you know, actually medically important.

The ECG has been the workhorse of cardiology for over a century. It just might have a whole second career ahead of it.

References

1. Thambiraj, G., Bollepalli, S.C., Johnson, A., Malhotra, R., Isselbacher, E.M., Singh, J.P., & Armoundas, A.A. (2026). Personalized artificial intelligence based left ventricular ejection fraction and systolic dysfunction assessment. NPJ Digital Medicine. DOI: 10.1038/s41746-026-02462-3

2. Attia, Z.I., Noseworthy, P.A., Lopez-Jimenez, F., et al. (2021). Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial. Nature Medicine. DOI: 10.1038/s41591-021-01335-4

3. Chen, H.Y., Lin, C.S., Fang, W.H., Lou, Y.S., et al. (2022). Artificial Intelligence-Enabled Electrocardiography Predicts Left Ventricular Dysfunction and Future Cardiovascular Outcomes. Journal of Personalized Medicine. PMCID: PMC8950054

4. Yau, T., Bergamaschi, T., et al. (2026). PULSE-HF: Predicting changes in left ventricular systolic function from ECGs of patients who have heart failure. Lancet eClinicalMedicine.

5. Multisite External Validation of an AI-Enabled ECG Algorithm for Detection of Low Ejection Fraction. (2025). JACC: Advances. DOI: 10.1016/j.jacadv.2025.102537

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.