AI walked novices through heart scans. The images mostly held up — with caveats.

Software that guides an untrained operator through a heart scan — prompting where to place the probe and grading in real time whether the picture is good enough to read — is the technology behind a push to let nurses and health workers, not just sonographers, acquire echocardiograms. Two 2026 studies found that novices with no prior ultrasound experience could, after roughly a day of AI-guided training, produce images that expert reviewers judged diagnostic most of the time [s1][s2]; both also found that "most of the time" hides large gaps by acoustic window, by clinical finding, and by who is holding the probe.

Why acquisition is the bottleneck

Echocardiography is operator-dependent: it requires acquiring and interpreting numerous parameters across different acoustic windows, and intra- and inter-operator variability remains a significant challenge that affects the reliability of the measurements, as the British Society of Echocardiography (BSE) sets out in a 2026 position statement [s3]. That is why the shortage of trained sonographers matters — and why AI that handles the acquisition step, not just the reading, is being tested as a way to move scanning closer to the patient. The guidance systems integrate into handheld or desktop ultrasound and coach the operator toward each standard view, flagging when an image clears a quality bar.

The multicentre secondary analysis

In a multicentre, prospective secondary analysis, nine novice operators performed limited transthoracic echocardiography on 159 patients using a handheld device with AI-based acquisition guidance, after eight hours of standardised training [s1]. Three blinded expert reviewers graded image quality on a 1–5 scale. Of 954 novice-acquired images, 97.7% met the diagnostic threshold of a score of 3 or higher [s1].

The averages hide where the tool is weak. Experts could reliably rule out left ventricular dysfunction (99.4%) and hypertrophy (98.7%) from the novice images, but agreement was lower for wall motion abnormalities (80.7%) and atrial dilation (86.6%) [s1] — and wall motion is exactly the kind of subtle, regionally specific finding that clinical decisions turn on. Body-mass index had a significant effect on image quality (P = 0.0029), though every subgroup still cleared the diagnostic threshold: mean scores ran from 4.44 in the leanest patients (BMI under 18) down to 4.07 in those with a BMI above 30 [s1].

The rural cohort

A separate cohort study put the same idea in rural and remote communities, where 181 participants (mean age 65, 47% female) were scanned with AI-guided point-of-care ultrasound after one day of lab training plus online mentoring [s2]. Here the window-by-window split was stark: diagnostic-quality images were obtained from 72% of parasternal acquisitions but only 55% of apical ones (P < 0.001) [s2]. The apical views — technically harder and often more diagnostically important — were where the guidance helped least.

Who held the probe mattered as much as the software. Physician-acquired scans were far more likely to reach diagnostic quality than those taken by nurses or health workers (odds ratio 3.85, 95% CI 1.92–8.33, P < 0.001) [s2]. Larger body surface area (OR 0.16, 95% CI 0.05–0.50, P = 0.002) and hypertension (OR 0.50, 95% CI 0.27–0.93, P = 0.03) both lowered the odds of a diagnostic apical image [s2]. AI guidance narrows the skill gap; it does not erase it.

What the profession's own body says

The BSE statement, produced by a multidisciplinary writing group of cardiologists, cardiac physiologists, clinical academics, AI researchers and patient representatives and approved by the society's trustees, reads as cautious enthusiasm rather than endorsement [s3]. It maps AI across the whole workflow — acquisition, analysis, reporting and risk stratification — and notes the potential of large language models to draft automated reports, but frames clinical adoption as still limited by practical, technical and governance challenges, and centres its principles on the UK's model of echocardiography delivery [s3]. The message is that safe, equitable integration needs validation and governance, not just a device that produces a good-looking picture.

What it means, and what to watch

For underserved and remote areas, task-shifting acquisition to non-experts is a genuine prospect on this evidence — but the two studies also mark its limits precisely. A diagnostic-quality image is not a diagnosis; the views and findings where novices struggle (apical windows, wall motion, atrial size) are not incidental; and a physician still outperforms a health worker with the same software [s1][s2]. Both studies validated image quality, judged by experts, not downstream patient outcomes — the gap that has dogged AI imaging devices generally, and the reason the BSE asks for local validation before scanning moves [s3].

What to watch is whether AI-acquired scans, read remotely, actually change diagnosis and management in the settings that lack sonographers — and whether the weaker apical and wall-motion performance closes with better guidance. For the interpretation side of the same problem, see our coverage of an AI that reads structural heart disease off an ECG, whether explanations build sonographer trust, and the persistent gap between AI device accuracy and patient outcomes.

Sources

• Artificial intelligence in breaking the learning curve for echocardiography: a secondary analysis of a multicentre trial — European Heart Journal - Digital Health, 2026-04-20
• Task-shifting to nonexperts using artificial intelligence-guided point-of-care ultrasound: a cohort study of patient selection, image quality, and learning curves — European Heart Journal - Imaging Methods and Practice, 2026-07-06
• Artificial intelligence in echocardiography: a position statement from the British Society of Echocardiography — Echo Research & Practice, 2026-07-20