ECG-Based Artificial Intelligence for Screening of Diastolic Dysfunction in Kenya: a prospective cohort study
A Pandey, M Segar, H Seung Lee, J Kwon, C Bhograj, P Golecha, B SamiaAbstract
Background
Diastolic dysfunction (DD) with elevated left atrial pressure (LAP) represents an important precursor of the development of heart failure with preserved ejection fraction. However, it remains challenging to detect in resource-limited settings where access to echocardiography is constrained. Electrocardiogram (ECG)-based artificial intelligence algorithms offer potential for scalable screening, but their performance for identifying DD with elevated LAP (DD-eLAP) in low- and middle-income countries is unknown.
Purpose
To evaluate the diagnostic performance of an AI-ECG algorithm for detecting DD-eLAP among adults seeking routine care in Kenya, and to assess associations between AI predictions and echocardiographic markers of cardiac dysfunction.
Methods
This prospective cross-sectional study enrolled adults ≥18 years from eight outpatient facilities in Kenya. Participants underwent cardiovascular risk assessment, 12-lead ECG, and echocardiography within 7 days. DD-eLAP and its components were defined by guideline criteria (Figure 1). An AI-ECG algorithm predicted high-risk DD, and diagnostic performance was evaluated against echocardiography. Logistic regression assessed associations with echo-confirmed abnormalities, adjusting for age, sex, Framingham risk score, and atrial fibrillation.
Results
Among 1,447 participants (age 58 years, 62% female), echo-confirmed DD-eLAP was present in 6.6% (95/1,447). The AI-ECG identified 52.2% as high-risk, with sensitivity 93.7% (95% CI 86.9-97.1%), specificity 50.7% (95% CI 48.0-53.3%), positive predictive value 11.8% (95% CI 9.7-14.3%), and negative predictive value 99.1% (95% CI 98.1-99.6%). The algorithm achieved an AUC of 0.84 (95% CI 0.81-0.88) for DD-eLAP, with similar discrimination for its individual components (Figure 1). In multivariable-adjusted models, positive AI-ECG screening was independently associated with DD-eLAP (OR 14.8, 95% CI 6.4-34.3, p<0.001) and its echocardiographic components (Figure 2).
Conclusion
AI-enabled ECG screening demonstrated high sensitivity and negative predictive value for identifying DD-eLAP, with positive predictions independently associated with echocardiographic markers of elevated filling pressures. These performance characteristics suggest feasibility as a screening tool to identify patients who warrant confirmatory echocardiography in resource-limited settings.For image description, please refer to the figure legend and surrounding text.For image description, please refer to the figure legend and surrounding text.