A Stethoscope That Heard What Routine Care Missed

Shortness of breath, swelling and exhaustion can accompany pregnancy. They can also accompany a weakening heart. That overlap can make pregnancy-related cardiomyopathy hard to recognize in routine care.

A pragmatic randomized trial in Nigeria asked whether artificial intelligence could help clinicians detect more cases among pregnant and postpartum patients. The answer was a qualified yes: a point-of-care digital-stethoscope pathway increased detection of left-ventricular systolic dysfunction, but the trial did not show that screening improved treatment, hospitalization or survival.

Research references: [1]

A screening pathway, not an AI diagnosis

Researchers randomized 1,232 participants at six Nigerian teaching hospitals between August 2022 and September 2023. The modified intention-to-treat analysis included 587 people in the intervention arm and 608 in usual care. Eligible participants were ages 18 to 49, pregnant or up to 12 months postpartum.

The intervention added two AI-assisted checks. One used a 12-lead electrocardiogram whose prediction was returned asynchronously, usually within a week. The other combined heart sounds with a single-lead ECG in a digital stethoscope and returned an AI result at the point of care. A positive screen still required echocardiographic confirmation and clinical judgment.

By the end of the study, reduced left-ventricular systolic function—defined as an ejection fraction below 50%—had been identified in 24 of 587 participants in the intervention group and 12 of 608 in the control group: 4.1% versus 2.0%. The odds ratio was 2.12 (95% confidence interval 1.05–4.27), and the estimated number needed to screen to detect one additional case was 47.

Those numbers support “more cases detected.” They do not support “disease prevented,” “lives saved” or “better treatment.”

Research references: [1] [2]

The useful signal came with a follow-up burden

At study entry, the digital-stethoscope model had an area under the receiver operating characteristic curve of 0.976 for identifying ejection fraction below 50%. Using the maximum prediction across recording positions, it identified 22 of 23 cases found by echocardiography, a sensitivity of 95.7%.

High sensitivity came with a tradeoff. Under that maximum-prediction strategy, positive predictive value was 18%, and 100 of 580 participants without the target condition screened positive—a 17% false-positive rate. The device worked as a broad net, not a verdict. Any health system using a similar screen would need enough echocardiography, clinicians and follow-up capacity to resolve the alerts it creates.

The second AI pathway was less conclusive. The 12-lead AI-ECG identified 20 cases in the intervention arm versus 12 in the control arm, an odds ratio of 1.75 (95% confidence interval 0.85–3.62; P=0.125). That comparison did not reach statistical significance. Treating “AI” as a single uniform intervention would hide a central result: the real-time digital-stethoscope pathway met the primary endpoint; the asynchronously returned 12-lead pathway did not.

Research references: [1]

Detection improved; health outcomes did not

The trial showed more diagnoses, not better health outcomes. There was no statistically significant difference in the composite adverse outcome or the composite cardiovascular outcome. The screening intervention did not standardize treatment, and the study was not designed to establish an effect on hospitalization, mortality, cost or utilization.

An all-cause mortality imbalance requires direct attention. Fifteen participants died during follow-up: 12 of 587 in the intervention arm and 3 of 608 in the control arm, corresponding to a hazard ratio of 4.20 (95% confidence interval 1.18–14.87). Cardiovascular mortality was numerically higher in the intervention arm but not statistically significant. The authors discuss possible differences in ascertainment; the study was not powered to explain this imbalance. It is therefore neither evidence that the AI caused deaths nor a result to dismiss. Later trials should be designed to examine downstream care and outcomes.

Research references: [1]

The comparison was not symmetrical

Intervention participants received a baseline echocardiogram; control participants did not receive one routinely. The true prevalence of reduced heart function in the control group is therefore unknown. Randomization helps balance groups, but unequal access to imaging can still shape which cases are recognized. The authors also note that out-of-pocket costs may have affected whether control participants received an echocardiogram during usual care.

Generalizability is limited too. The study took place at six teaching hospitals with cardiologists and echocardiography capacity. All participants identified as Black women in Nigeria, only 61% completed a second visit, and the trial’s ejection-fraction threshold did not exactly match the lower thresholds used to train the underlying algorithms. These facts do not erase the result; they define the setting in which it was observed and the validation still needed elsewhere.

The registered protocol confirms that the study was designed around detection of ejection fraction below 50%, with a modified intention-to-treat analysis and prespecified secondary and exploratory outcomes. It also shows that the digital-stethoscope pathway was an exploratory point-of-care component of the broader AI-ECG program.

Research references: [1] [2]

The human system is the intervention

The most useful description of this technology is not “AI diagnoses heart disease.” It is “a health worker gains a better reason to look.”

A positive alert needs confirmation. A confirmed diagnosis needs a clinician, an affordable treatment plan and continuity after delivery. A false positive needs a clear explanation and timely resolution. A negative result cannot erase symptoms that still warrant clinical evaluation.

Implementation work from the SPEC-AI team emphasizes the same point: deploying healthcare AI in resource-constrained settings depends on infrastructure, workflow, data governance, workforce capacity and local trust—not model performance alone.

Research references: [4]

Funding and commercial context

The published trial reports funding from Mayo Clinic programs, including work with Dalio Philanthropies, and support from National Institutes of Health grants. Eko Health extracted portable recordings and AI predictions for analysis; the paper says Eko and Anumana did not design the study, collect or analyze the clinical data, interpret results or write the report. Several authors and Mayo Clinic reported patents, licenses, advisory roles, consulting, stock options or potential equity and royalty interests involving companies including Anumana, AliveCor and Eko. The current article includes a publisher-recorded Author Correction dated February 4, 2025; the updated Nature Medicine version is the controlling report.

The trial offers credible evidence that a real-time AI-enabled digital stethoscope can increase detection inside this obstetric workflow. It does not yet show that the technology improves maternal outcomes. The next generation of studies should ask both questions at once: not only whether the model hears a signal, but whether the health system can turn that signal into safer care.

Research references: [1] [3]