Artificial intelligence guided screening for cardiomyopathies in an obstetric population: a pragmatic randomized clinical trial

Cardiomyopathy is a leading cause of maternal mortality in the United States and the primary cause of death in the postpartum period, with substantially higher incidence reported in Nigeria (approximately 1 in 96 deliveries). Symptoms of heart failure can mimic normal pregnancy physiology, often delaying diagnosis and worsening outcomes. Hemodynamic changes in pregnancy may unmask previously undiagnosed LV systolic dysfunction (LVSD). While AI-enabled ECG-based tools have shown strong retrospective and pilot prospective performance for detecting low ejection fraction, it was unknown whether AI-guided screening improves detection of pregnancy-related cardiomyopathy beyond usual care. This study aimed to determine whether AI-guided screening using a digital stethoscope and 12-lead ECG increases identification of LVSD among pregnant and postpartum women in Nigeria compared with standard care.

Prior work has demonstrated AI-enabled ECG effectiveness in detecting various cardiovascular diseases, including low LVEF. Retrospective analyses and a pilot prospective study in obstetric populations reported high AUCs (up to 1.00 with 12-lead ECG and 0.98 with a digital stethoscope) for detecting LVSD (LVEF <45%). Additional retrospective studies in the United States and Korea validated separate AI-ECG models for perinatal LVSD. Prospective multicenter studies using ECG-enabled stethoscopes in general populations also showed feasibility. These findings suggested AI could enable early identification of LVSD in pregnancy/postpartum, motivating a randomized pragmatic evaluation in a high-incidence, low-resource setting.

Design: SPEC-AI Nigeria was an investigator-initiated, pragmatic, multicenter, open-label, randomized clinical trial conducted at six Nigerian hospitals. Participants were randomized 1:1 to AI-guided screening (intervention) versus usual care with 12-lead clinical ECG (control). Randomization used dynamic minimization with site stratification via a web application. Participants: Women aged 18–49, pregnant or within 12 months postpartum, receiving obstetric care at one of six sites. Key exclusions: complex congenital heart disease, significant conduction abnormalities (e.g., complete heart block, pacemaker), and inability to consent. All participants provided informed consent. All sites had cardiologists and echocardiography capability. Interventions: All participants received a standard 12-lead ECG at enrollment. Control arm: usual care plus AI-based age/sex estimation (attention control). Intervention arm: in addition, 15-second ECG/phonocardiogram recordings with an AI-enabled digital stethoscope at predefined chest positions (V2, angled, handheld) with real-time binary AI predictions for LVSD; 12-lead AI-ECG for LVSD (predictions provided asynchronously, typically within 1 week); and a baseline echocardiogram for AI validation. For the primary endpoint in the intervention arm, a positive AI screen was required and echocardiography confirmation was mandated; recordings deemed poor quality by in-built checks were treated as negative. AI algorithms: Convolutional neural networks trained on >100,000 adults. The original Mayo Clinic 12-lead AI-ECG targeted LVEF ≤35%, later retrained to detect LVEF <40% (US FDA-cleared version with data quality checks). The stethoscope model was adapted for single-lead ECG plus phonocardiogram to detect LVEF <40% (ELEFT 7.2.0) with device-level quality checks. Outcomes: Primary outcome was identification of LVSD defined as LVEF <50% on echocardiography. Control arm counted clinically recognized/documented LVSD per usual care. Intervention arm counted AI-positive screens (digital stethoscope maximum across locations or any positive 12-lead AI-ECG per encounter) confirmed by echocardiography at the time of acquisition. If the AI screen was negative/not computed/poor quality, a protocol echocardiogram with LVEF <50% did not count toward the primary endpoint. Secondary outcomes: subgroup performance (age, ethnicity, region, hypertensive disorders, pregnancy/postpartum) and diagnostic performance at LVEF thresholds (≤35%, <40%, <45%, <50%) in the intervention arm at baseline, plus exploratory composites of adverse and cardiovascular outcomes and all-cause mortality. Follow-up: Participants could have up to seven visits from early pregnancy through 12 months postpartum; follow-up continued through study end. Testing frequency aligned with routine obstetric schedules; home visits were used at some sites. Statistics: Sample size assumed 4% vs 1% LVSD prevalence (intervention vs control) to achieve 80% power at α=0.05; target 1,200 participants. Modified intention-to-treat (mITT) excluded those without baseline testing, pre-baseline death, or withdrawal; two unplanned analyses included a full intention-to-treat (ITT, assuming excluded participants negative for LVSD) and site-adjusted logistic regression. Diagnostic metrics and ROC AUC were reported per STARD for intervention baseline echocardiogram-validated assessments. Analyses used R v4.1.2 with two-sided P<0.05 considered significant. Ethics approvals were obtained locally and at Mayo Clinic; trial registered (NCT05438576).

- Enrollment/randomization: 1,232 women (616 per arm) across six Nigerian sites (Aug 2022–Sep 2023); mITT included 1,195 (587 intervention; 608 control) with follow-up through May 2024. - Primary outcome (LVSD, LVEF <50%): - Digital stethoscope AI: 24/587 (4.1%) vs 12/608 (2.0%); OR 2.12 (95% CI 1.05–4.27); P=0.032. Estimated number needed to screen (NNS) to detect one additional case: 47. Results robust to site adjustment (OR 2.25, 95% CI 1.09–4.66; P=0.029) and full ITT analyses (unadjusted OR 2.04, 95% CI 1.01–4.12; P=0.042; site-adjusted OR 2.13, 95% CI 1.03–4.41; P=0.041). - 12-lead AI-ECG (US FDA-cleared): 20/587 (3.4%) vs 12/608 (2.0%); OR 1.75 (95% CI 0.85–3.62); P=0.125 (not statistically significant). - 12-lead AI-ECG (original Mayo model): 18/587 (3.1%) vs 12/608 (2.0%); OR 1.57 (95% CI 0.75–3.29); P=0.227 (not significant). - Subgroups: Digital stethoscope and 12-lead AI-ECG showed consistent directional benefit across prespecified subgroups; stronger effect for the 12-lead model in participants ≥30 years (OR ~4.2) compared to <30 years (OR ~0.9–1.4). - Diagnostic performance (intervention arm, baseline): - Digital stethoscope (max across locations): AUC 0.976 (95% CI 0.953–0.998) for LVEF <50%; AUC 0.985 (95% CI 0.974–0.996) for LVEF <40%. Sensitivity for LVEF <50% was 95.7% with max prediction; PPV 18.0% and higher false positives vs single best position (angled). - 12-lead AI-ECG (US FDA-cleared): AUC 0.928 (95% CI 0.875–0.981) for LVEF <50%; AUC 0.928 (95% CI 0.865–0.990) for LVEF <40%. - 12-lead AI-ECG (original Mayo): AUC 0.892 (95% CI 0.825–0.960) for LVEF <50%; AUC 0.921 (95% CI 0.864–0.979) for LVEF <40%. - Exploratory outcomes: No significant difference in composite adverse events (100/587 vs 104/608; OR 1.00, 95% CI 0.74–1.35; P=0.975) or composite cardiovascular events (56/587 vs 53/608; OR 1.10, 95% CI 0.74–1.64; P=0.621). All-cause mortality was higher in intervention (12/587) vs control (3/608); HR 4.20 (95% CI 1.18–14.87). Cardiovascular mortality was numerically higher (5 vs 3) but not statistically significant (HR 1.75, 95% CI 0.42–7.33). - Safety: No serious adverse events related to study participation; five reports of skin irritation from ECG electrodes (1 intervention; 4 control).

The trial directly addressed whether AI-guided screening improves detection of pregnancy-related LV systolic dysfunction beyond standard obstetric care. AI-enabled digital stethoscope screening, with real-time point-of-care predictions and confirmatory echocardiography, approximately doubled LVSD detection compared to usual care, reflecting improved case ascertainment in a high-risk, low-resource setting. While AI-ECG (12-lead) showed a consistent directional trend, statistical significance was not reached, potentially related to threshold alignment, data quality handling, or age-related performance differences. High AUCs across models indicate strong discriminative capability, with the stethoscope model achieving the highest. These findings support the feasibility and clinical utility of AI-guided point-of-care screening to reduce diagnostic delay for peripartum cardiomyopathy, facilitate earlier referral in settings with limited cardiology capacity, and enable scalable, low-cost screening workflows. Observed higher all-cause mortality in the intervention arm likely reflects ascertainment and follow-up differences rather than harm from screening; cardiovascular mortality differences were not statistically significant. Overall, the study demonstrates that integrating AI into routine obstetric workflows can enhance detection of a life-threatening but treatable condition, with potential to reduce disparities in cardio-obstetric care.

Among pregnant and postpartum women in Nigeria, AI-guided screening using a digital stethoscope significantly increased detection of cardiomyopathy associated with LVSD compared with usual care. The intervention showed excellent diagnostic performance and feasibility at the point of care, suggesting an impactful, scalable approach for early identification in low-resource settings. Future research should assess impacts on treatment decisions, healthcare utilization and costs, maternal and infant outcomes, and mortality; optimize model thresholds for obstetric populations; evaluate implementation strategies across diverse care settings and earlier pregnancy time points; and conduct population-level studies to refine prevalence estimates and screening policies.

- Pragmatic design at tertiary teaching hospitals with cardiology access may limit generalizability to lower-tier facilities. - Majority enrolled in late pregnancy or postpartum, reducing possible follow-up visits and limiting longitudinal assessment; only 61% completed a second study visit, with attrition thereafter. - Control arm echocardiography was not mandated; true LVSD prevalence in control is uncertain and could be influenced by physician discretion, out-of-pocket costs, and return logistics. Socioeconomic status was not collected. - Mortality ascertainment challenges (no national registry; reliance on medical records and phone follow-up) may introduce reporting bias; more frequent interactions in the intervention arm could increase detection of deaths. - Primary endpoint defined LVSD as LVEF <50%, which does not fully align with model training thresholds (≤35% or <40%), potentially affecting sensitivity/specificity estimates and between-model comparisons. - Limited follow-up window and variable entry time points resulted in end-of-study detection rates similar to baseline. - Some AI predictions treated as negative due to poor-quality recordings may bias detection downward.