An AI stethoscope raised detection when used, but not across a trial of 205 practices

An AI-enabled stethoscope records a few seconds of heart sounds and a single-lead electrocardiogram, then runs them through algorithms trained to flag three conditions a clinician can easily miss: heart failure, atrial fibrillation, and valvular heart disease. In the largest real-world test of the device yet — a cluster-randomised trial across 205 UK general practices — handing it to primary care did not increase how many heart-failure cases were diagnosed over a year, even though the tool, when it was actually used, was linked to finding more disease [s1].

That split is the whole story, and it is a familiar one: a device can be accurate in the hand of a clinician who reaches for it and still change nothing at the level of a health system, because most of the time it stays in the drawer.

What the device does

During a cardiac examination the AI stethoscope captures 15 seconds of single-lead ECG and phonocardiogram signals and feeds them to three algorithms that return yes-or-no predictions for reduced left ventricular ejection fraction (defined as 40% or below), atrial fibrillation, and valvular heart disease — all functions with regulatory approval [s1]. The promise is a 15-second screen that turns a routine stethoscope into a detector for conditions that usually surface late, often only after a hospital admission.

What TRICORDER tested

TRICORDER was a pragmatic, cluster-randomised controlled implementation trial [s1]. Rather than measure whether the algorithms are accurate — that had been established beforehand — it asked the harder question of whether putting the technology into ordinary practice changes what gets diagnosed. UK primary care practices were randomised 1:1 to the intervention (training and implementation of the AI stethoscope in routine care) or to routine care alone [s1].

Between 30 October 2023 and 22 May 2024, 205 practices were assigned — 96 to the intervention arm, covering 701,933 registered patients, and 109 to control, covering 851,242 [s1]. The primary endpoint was the incidence of any newly coded heart-failure diagnosis, expressed per 1000 patient-years and compared as an incidence rate ratio, drawn from an NHS secure data environment [s1].

The result

Across the intervention practices, clinicians recorded 12,725 examinations with the AI stethoscope, spread across 972 users [s1]. But in the intention-to-treat analysis — the comparison that counts every randomised practice regardless of how much the device was used — heart-failure detection did not differ between the groups: an incidence rate ratio of 0.94 (95% CI 0.86–1.02), with no difference in whether the diagnosis was made in the community or via hospital admission [s1].

The device's defenders point to a second finding in the same paper: actually using the AI stethoscope was independently associated with significantly higher detection of heart failure, and of atrial fibrillation and valvular heart disease [s1]. Both readings are true, and the tension between them is the point. The tool works when it is used; the trial's failure was one of adoption, not algorithm. Over 12 months, the authors conclude, implementation in routine primary care did not significantly increase heart-failure detection or shift it earlier into the community [s1].

The accuracy was never the weak link

A separate prospective study underlines that the underlying detection is real. Among 357 patients aged 50 or older at risk of heart disease, an AI-augmented digital stethoscope was compared against standard analogue auscultation by primary care providers, with echocardiography as the reference [s2]. The AI system reached a sensitivity of 92.3% for moderate-to-severe valvular heart disease versus 46.2% for the clinician's ear alone (p=0.01), though at a cost in specificity, 86.9% versus 95.6% (p<0.001) [s2]. It picked up 12 previously undiagnosed cases of significant valve disease, against 6 for routine auscultation [s2].

So the instrument hears what an ordinary stethoscope cannot. TRICORDER's contribution is to show that this is necessary but not sufficient: a more sensitive tool used in 12,725 examinations across nearly a million patients still did not move the population number, because the bottleneck sits in clinician time, workflow and habit rather than in the sensor.

Why it matters

The AI-medicine field is full of devices cleared on accuracy metrics whose effect on patients was never tested — a gap that runs through our coverage of the FDA's AI radiology testing gap and the wider outcomes gap in AI medical devices. TRICORDER is the rarer thing: a randomised trial that measured the outcome and reported a null, honestly. It sits alongside the evidence on AI ECG for atrial-fibrillation screening and ECG-AI for structural heart disease as a reminder that a tool's ceiling is set by whether clinicians reach for it. The trial was funded by the National Institute for Health and Care Research, the British Heart Foundation, and Imperial Health Charity [s1].

What to watch

Whether a redesign that lowers the friction of use — embedding the prompt in the consultation, or narrowing which patients it is offered to — can convert the per-use detection gain into a population one. The authors frame TRICORDER itself as a template: a randomised design that generates the real-world data needed to understand why healthcare innovations stall between a working device and a changed statistic [s1]. Until that adoption problem is solved, an accurate stethoscope that mostly goes unused is not the screening tool it is sold as.

This article is informational and is not medical advice.

Sources

• [s1] Kelshiker M, Bächtiger P, Petri C, et al. "Triple cardiovascular disease detection with an artificial intelligence-enabled stethoscope (TRICORDER) in the UK: a cluster-randomised controlled implementation trial." The Lancet, published online 28 January 2026. https://doi.org/10.1016/S0140-6736(25)02156-7
• [s2] "Artificial-intelligence-enabled digital stethoscope improves point-of-care screening for moderate-to-severe valvular heart disease." European Heart Journal - Digital Health, published online 5 February 2026. https://doi.org/10.1093/ehjdh/ztag003