Structured Section Plan
This article is organized into the following sections to match the clinical and scientific priorities of the study: Highlights; Background and unmet need; Study design and methods; Key results; Clinical interpretation; Strengths and limitations; Implications for practice and health systems; Conclusion; Funding and trial registration; Citation.
Highlights
First, in eight outpatient facilities across Kenya, an AI-enabled 12-lead ECG model demonstrated very strong performance for detecting left ventricular systolic dysfunction (LVSD) when benchmarked against echocardiography.
Second, among 1444 adults who completed paired echocardiography, LVSD defined as left ventricular ejection fraction below 40% was present in 14.1%, indicating a substantial burden of undiagnosed or under-recognized systolic dysfunction in this ambulatory population.
Third, the AI-ECG algorithm achieved 95.6% sensitivity, 79.4% specificity, an area under the receiver operating characteristic curve of 0.96, and an exceptionally high negative predictive value of 99.1%, making it particularly attractive as a rule-out screening tool.
Fourth, diagnostic performance remained consistent across prespecified cardiovascular risk strata, supporting the possibility that AI-ECG screening could be deployable beyond only the highest-risk groups.
Background and Clinical Need
Heart failure with reduced ejection fraction remains a major cause of morbidity and mortality globally, and its burden is increasingly recognized across sub-Saharan Africa. A central challenge is that left ventricular systolic dysfunction often precedes overt clinical heart failure, yet it may remain undetected until patients present with advanced symptoms. Earlier diagnosis matters because guideline-directed medical therapy can reduce hospitalization and improve survival once reduced ejection fraction is identified.
In high-resource environments, echocardiography is the standard tool for evaluating left ventricular function. In many low- and middle-income settings, however, access to echocardiography is constrained by cost, availability of machines, trained personnel, referral delays, and geographic concentration of diagnostic services in urban centers. These barriers make population-scale or routine opportunistic screening difficult.
This is the clinical space in which artificial intelligence applied to the electrocardiogram has generated enthusiasm. AI-ECG models, especially convolutional neural networks trained on large labeled datasets, can detect latent structural and functional cardiac abnormalities from standard ECG waveforms, including reduced left ventricular ejection fraction. Previous studies from high-income settings have reported promising performance, but translation into real-world resource-limited settings is not guaranteed. Differences in comorbidity patterns, ECG acquisition quality, disease spectrum, care pathways, and referral infrastructure may alter both feasibility and test characteristics.
The present study by Pandey and colleagues is therefore important not simply because it tests an algorithm, but because it asks a health-system question: can AI-ECG meaningfully expand access to LVSD screening where echocardiography is limited?
Study Design and Methods
This investigation was a cross-sectional study conducted from June to December 2024 at eight outpatient health care facilities across Kenya. Adults aged 18 years and older who were seeking routine care were eligible for enrollment, provided they could give informed consent. Participants underwent baseline assessment and a 12-lead ECG. A subset then completed echocardiography within seven days, allowing comparison between AI-ECG predictions and the imaging reference standard.
The AI tool used was a validated convolutional neural network algorithm, AiTiALVSD, designed to estimate risk of LVSD from ECG data. The reference standard outcome was LVSD defined as left ventricular ejection fraction less than 40% on echocardiography.
The echocardiography subset was drawn from three prespecified strata: patients with prior cardiovascular disease, patients at high cardiovascular risk defined by a Framingham Risk Score of at least 10%, and those at lower risk with a Framingham Risk Score below 10%. This stratified approach is clinically relevant because it allows assessment of whether performance is maintained across different baseline risk profiles rather than being driven only by sicker patients.
The abstract reports 1444 participants who completed paired echocardiography. Their mean age was 59.0 years, 62.8% were female, and 77.4% were categorized as high risk. The primary diagnostic metrics were sensitivity, specificity, positive predictive value, negative predictive value, and area under the receiver operating characteristic curve.
Key Results
Study population and disease prevalence
Among the 1444 participants with ECG and echocardiography, 204 had LVSD, yielding a prevalence of 14.1%. This is a clinically meaningful prevalence for an outpatient cohort and immediately affects interpretation of predictive values. In populations where disease prevalence is moderate, a screening test with high sensitivity and good specificity can still have only modest positive predictive value, while maintaining a very high negative predictive value.
Diagnostic performance of AI-ECG
The AI-ECG algorithm performed strongly against echocardiography. Sensitivity was 95.6% with a 95% confidence interval of 91.8% to 97.7%. This means that the algorithm correctly identified nearly all participants who truly had LVSD, an essential property for a screening test intended to minimize missed cases.
Specificity was 79.4% with a 95% confidence interval of 77.0% to 81.5%. Although not perfect, this level of specificity is acceptable in many screening contexts, particularly when the downstream confirmatory test is echocardiography and the clinical priority is not to overlook patients with potentially treatable systolic dysfunction.
The positive predictive value was 43.2% with a 95% confidence interval of 38.7% to 47.9%. This lower figure should not be read as a weakness in isolation. Instead, it reflects the mathematics of screening in a population where most individuals do not have LVSD. In practical terms, fewer than half of AI-positive patients would be confirmed to have LVEF below 40% on echocardiography, so AI-ECG is not a substitute for imaging-based diagnosis.
The negative predictive value was 99.1% with a 95% confidence interval of 98.3% to 99.5%. This is arguably the most clinically useful result in the paper. A negative AI-ECG result makes LVSD very unlikely, suggesting that the tool could help avoid unnecessary echocardiograms in low-probability individuals and help prioritize imaging capacity for those most likely to benefit.
Overall discrimination was excellent, with an area under the receiver operating characteristic curve of 0.96 and a 95% confidence interval of 0.95 to 0.97. An AUC at this level indicates that the algorithm was highly effective at ranking patients with and without LVSD across the spectrum of predicted risk.
Performance across risk strata
A notable strength is the consistency of performance across cardiovascular risk strata. The reported AUC ranged from 0.96 to 0.98 in the prespecified groups. This suggests that the algorithm’s signal was not confined to patients with known cardiovascular disease or very high traditional risk. If confirmed in broader implementation studies, that consistency could support flexible deployment strategies, including screening of targeted outpatient populations rather than referral only after clinical suspicion is already high.
Clinical Interpretation
The most important practical message is that AI-ECG appears well suited as a triage or screening tool, not as a stand-alone diagnostic test. Its very high sensitivity and negative predictive value mean it may be most useful for ruling out LVSD in ambulatory patients when echocardiography is scarce. In resource-limited systems, this is not a trivial gain. If a standard ECG can be transformed into a high-performing front-end screening test, clinicians may be able to direct confirmatory imaging toward those with a positive AI signal, symptoms, prior disease, or other markers of elevated risk.
This matters because conventional clinical assessment alone can miss asymptomatic or mildly symptomatic LVSD. Many patients with hypertension, diabetes, ischemic heart disease, or previous myocarditis may not undergo imaging until they are significantly ill. AI-ECG offers a pathway to earlier risk detection using equipment that is already far more available than echocardiography.
The modest positive predictive value also has practical consequences. An AI-positive ECG should not trigger heart failure treatment solely on its own. Rather, it should trigger confirmatory echocardiography when feasible, or at least structured clinical evaluation and follow-up where imaging access remains delayed. Programs adopting such tools will need clear care pathways to prevent overdiagnosis, unnecessary anxiety, or diversion of limited resources toward false positives.
The consistency of performance across risk strata is particularly encouraging for public health planning. It raises the possibility of embedding AI-ECG into routine chronic disease clinics, hypertension clinics, diabetes follow-up services, or even broader internal medicine outpatient workflows. The operational advantage is that ECG acquisition requires less infrastructure and can often be decentralized more readily than echo.
Strengths and Limitations
Strengths
This study has several strengths. It was conducted in a real-world African outpatient setting rather than an exclusively high-income tertiary environment. It used echocardiography as the gold standard within seven days of ECG acquisition, limiting temporal misclassification. The sample size was substantial, and the stratified design allowed examination of performance in clinically distinct subgroups. Most importantly, the study addresses a translational gap between algorithm development and implementation where evidence is often thin.
Limitations
At the same time, caution is warranted. First, the study was cross-sectional, so it does not answer whether AI-ECG screening improves patient outcomes, treatment initiation, hospitalization rates, or mortality. Diagnostic performance is necessary but not sufficient for clinical utility.
Second, the echocardiography subset was selected using prespecified risk strata rather than as a purely consecutive universal imaging sample. That is a reasonable design choice in a resource-constrained study, but it can influence prevalence estimates and predictive values, and may complicate extrapolation to lower-risk general populations.
Third, the abstract provides limited information on calibration, threshold selection, ECG quality control, intersite variability, missing data, and subgroup performance by sex, age, rhythm abnormalities, or specific underlying causes of cardiomyopathy. These details matter for implementation.
Fourth, LVSD was defined as LVEF below 40%, which aligns with clinically important reduced ejection fraction but does not address mildly reduced ejection fraction or broader structural heart disease. A screening strategy focused only on this threshold may miss other actionable pathology.
Finally, external validity beyond the participating Kenyan sites remains to be tested. Performance in rural facilities, community screening environments, or populations with different etiologic patterns of heart disease may differ.
Implications for Practice and Health Systems
If integrated carefully, AI-ECG could become a pragmatic gatekeeper to echocardiography. In settings where only a fraction of patients can access imaging, high-sensitivity screening can help prioritize those most likely to harbor reduced ejection fraction. This may shorten time to diagnosis for patients who would benefit from angiotensin receptor-neprilysin inhibitors, beta-blockers, mineralocorticoid receptor antagonists, sodium-glucose cotransporter-2 inhibitors, or etiologic evaluation for ischemic and nonischemic cardiomyopathy.
For health systems, the question is not simply whether the algorithm works, but whether it changes workflow efficiently. Key implementation needs include reliable ECG digitization, standardized acquisition, local regulatory oversight, integration into clinical decision pathways, training of frontline staff, and protection against automation bias. A positive result should trigger an explicit next step; a negative result should not override strong clinical suspicion when symptoms or signs suggest heart failure.
Future research should address several issues: prospective impact on patient outcomes; cost-effectiveness compared with usual referral strategies; performance in primary care and rural facilities; fairness across sex and age groups; and utility in patients with arrhythmias, conduction disease, or prior myocardial infarction. Head-to-head comparisons with natriuretic peptide-based screening may also be informative in low-resource settings where biomarker availability varies.
Conclusion
This study provides persuasive evidence that AI-enabled ECG screening for left ventricular systolic dysfunction is feasible and diagnostically robust in Kenyan outpatient care. The combination of 95.6% sensitivity, 99.1% negative predictive value, and AUC of 0.96 suggests strong value as a rule-out and triage tool where echocardiography capacity is limited. The results do not eliminate the need for confirmatory imaging, but they support a scalable strategy to expand earlier detection of reduced ejection fraction in resource-constrained settings. The next step is to determine whether embedding AI-ECG into real care pathways improves access, accelerates treatment, and ultimately reduces heart failure morbidity and mortality.
Funding and Trial Registration
The abstract and citation provided do not specify funding information or a ClinicalTrials.gov registration number. These details should be confirmed from the full-text article.
Citation
Pandey A, Keshvani N, Segar MW, Kwon JM, Lee HS, Bhograj C, Mashilane KI, Jain N, Mwiti W, Wambari E, Nguchu H, Wagana-Muriithi LN, Anyira E, Namasaka P, Mbau L, Wairagu A, Muthui-Mutua B, Bikoro M, Mwita MCR, Njeri I, Gituma B, Mbogo D, Ngolobe A, Nabiswa H, Samia B. Artificial Intelligence Electrocardiogram and Left Ventricular Systolic Dysfunction in Kenya. JAMA Cardiology. 2026-05-06. PMID: 42090146. URL: https://pubmed.ncbi.nlm.nih.gov/42090146/
Selected Contextual References
Attia ZI, Kapa S, Lopez-Jimenez F, et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nature Medicine. 2019;25(1):70-74.
Heidenreich PA, Bozkurt B, Aguilar D, et al. 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure. Circulation. 2022;145(18):e895-e1032.
McDonagh TA, Metra M, Adamo M, et al. 2023 Focused Update of the 2021 ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure. European Heart Journal. 2023;44(37):3627-3639.
