Rheumatic Heart Disease (RHD) affects approximately 55 million people worldwide, with the Global South bearing the greatest burden, particularly among children, adolescents, and young adults. Recently, the WHO declared RHD a global health priority, targeting a 25% reduction in premature deaths by 2030. Achieving this goal requires cost‑effective and scalable interventions for early detection in at‑risk populations. However, in low-resource settings, effective RHD prevention is hindered primarily by the absence of periodic community-based screening programs. As a result, patients often present late with severe complications, overwhelming cardiology services and making valve surgery unaffordable or inaccessible for most. One potential solution is the use of machine learning methods applied to electrocardiogram (ECG) and phonocardiogram (PCG) recordings from low-cost, portable devices as a screening tool. This research investigates whether portable ECG and PCG devices can serve as an effective, low-cost screening modality for early, asymptomatic RHD in resource‑constrained settings. To design and validate a scalable, AI‑enabled screening framework using such low-cost devices, we first established a novel, comprehensive multimodal dataset consisting of ECG, PCG, and echocardiography recordings from 627 schoolchildren in high‑risk communities. Statistical analysis identified ECG biomarkers associated with early RHD, including PR interval (p<0.001), Pw/PR ratio (p<0.001), and P‑wave dispersion (p=0.008), demonstrating significant electrophysiological differences between asymptomatic RHD cases and healthy individuals. Furthermore, an automated ECG analysis pipeline using efficient heuristic waveform delineation algorithms was validated against deep learning models and manual PQRST annotations, achieving onset delineation errors of 11.9 ms (P-wave), 9.8 ms (QRS complex), and 17.0 ms (T-wave), with corresponding offset errors of 11.3 ms, 7.5 ms, and 13.2 ms, respectively. These results demonstrated comparable performance to deep learning approaches thereby feasible for real-time implementation that can be operated by local healthcare staff. Machine learning models trained on automate ECG‑derived temporal, time‑frequency, wavelet, and graph‑based features achieved an F1‑score of 62.8%, sensitivity of 59.6%, and PPV of 66.7% for asymptomatic RHD detection, while maintaining detection of 51.1-65.7% of true RHD cases across clinically relevant operating thresholds. In contrast, PCG features alone achieved an F1‑score of only 22.3%, and combining ECG with PCG did not improve performance, confirming ECG as the most informative low‑cost screening modality in asymptomatic cohorts. The approach also demonstrated strong clinical utility for symptomatic RHD triage, achieving 93.6% sensitivity and 68.5% PPV using time-frequency ECG features, increasing to 97.2% sensitivity and 92.1% PPV when trained using attention‑based Inception‑Time convolutional neural networks (CNNs). In summary, this study demonstrated that affordable portable ECG and PCG devices, combined with machine learning, can serve as a practical first‑line screening tool without requiring specialist operators or echocardiography at the point of care. Rather than replacing echocardiography, the system identifies children who are highly likely to require confirmatory assessment, enabling more efficient allocation of scarce healthcare resources while reducing unnecessary referrals. Future implementation should focus on expanding training datasets, adopting the updated 2023 WHF diagnostic criteria, and integrating multimodal clinical data through foundation models to improve screening sensitivity. These advances would enable community health workers to perform large‑scale decentralized screening. Consequently, this supports early detection, timely antibiotic prophylaxis, and ultimately reduces the need for costly valve surgery in low‑resource settings.
Source
Amsalu Thomas Chuma. Lirias (KU Leuven), 2026.