Machine Learning-Based Analysis of ECG and PCG Signals for Rheumatic Heart Disease Detection: A Scoping Review (2015-2025)

Technical Summary: Machine Learning-Based Analysis of ECG and PCG Signals for Rheumatic Heart Disease Detection (2015–2025)

Problem Statement
Rheumatic heart disease (RHD) remains a significant global health burden, disproportionately affecting low- and middle-income countries (LMICs) where it accounts for up to 1.5% of cardiovascular mortality. While the World Heart Federation has set a "25 by 25" goal to reduce RHD-related premature mortality by 25% by 2025, early detection remains a critical bottleneck. The diagnostic gold standard, echocardiography, is largely inaccessible in resource-limited settings due to high costs and workforce shortages. Conversely, traditional clinical auscultation suffers from declining proficiency and high false-negative rates, potentially missing up to 90% of cases. This review addresses the need for accessible, scalable screening alternatives using Electrocardiogram (ECG) and Phonocardiogram (PCG) signals enhanced by machine learning (ML) to bridge the diagnostic gap.

Methodology
This study is a scoping review conducted in accordance with PRISMA-ScR guidelines and registered with the Open Science Framework (OSF). The authors systematically searched PubMed, IEEE Xplore, Scopus, and Embase for peer-reviewed studies published between January 2015 and March 2025.

• Selection Criteria: The review focused on original research and conference proceedings applying AI/ML to analyze ECG and PCG signals for diagnosing RHD and its variants (e.g., aortic stenosis, mitral regurgitation, mitral stenosis).
• Process: Two independent reviewers screened titles, abstracts, and full texts, with agreement assessed via Cohen's Kappa coefficient.
• Scope: A total of 37 studies were selected for synthesis. The analysis categorized studies by disease target, ML methodology, signal processing techniques, validation strategies, and performance metrics.

Key Contributions
The paper provides the first comprehensive mapping of the ML landscape for RHD detection using ECG and PCG over a decade. Its primary contributions include:

1. Categorization of Techniques: It documents the evolution from classical algorithms (SVM, Random Forest) to deep learning architectures, specifically Convolutional Neural Networks (CNNs), and hybrid models (CNN-LSTM).
2. Signal Processing Standards: It identifies consensus preprocessing techniques, noting that bandpass filtering (20–250 Hz), downsampling, and normalization are standard, while feature extraction has shifted from manual engineering (MFCCs) to automated feature discovery via spectrograms and wavelets in deep learning models.
3. Critical Gap Analysis: The review systematically identifies methodological weaknesses in current literature, including a lack of external validation, demographic biases, and insufficient severity stratification.
4. Implementation Reality Check: It highlights the disconnect between high laboratory performance metrics and the lack of cost-effectiveness analysis or feasibility studies for real-world deployment in endemic regions.

Results

• Performance Metrics: The selected studies reported high technical performance, with a median accuracy of 97.75% (IQR: 88.4–97.1%), a median F1-score of 0.95, and an AUROC of 0.89. Notable outliers achieved accuracies up to 99.90% using Vision Transformers.
• Model Evolution: There was a clear shift in methodology. Early studies (2015–2019) relied on Support Vector Machines (SVM) and manually engineered features. By 2022–2025, CNNs became the dominant architecture (51.4% of studies), with hybrid models emerging to improve accuracy by 3–5% over single-algorithm approaches.
• Data Characteristics:
  • Sample Size: Studies suffered from small sample sizes, with a median cohort of only 244 subjects.
  • Validation: A critical limitation was the lack of external validation; only 10.8% of studies validated models on independent cohorts. 73% of studies used single-center datasets, and 81.1% relied on private data.
  • Demographics: 48.6% of studies reported significant age and gender disparities between case and control groups, introducing potential confounding factors.
• Modality Comparison: PCG models demonstrated superior performance for detecting specific valve pathologies (e.g., murmurs) due to direct acoustic signatures. ECG models detected RHD indirectly through structural consequences (e.g., atrial enlargement), showing lower efficacy for early or subclinical stages.
• Geographic Distribution: While 45.9% of studies originated from endemic regions (notably Sub-Saharan Africa and South Asia), the review found a critical disconnect between technical development and implementation science. No study provided a comprehensive cost-effectiveness analysis, and only 8.1% addressed practical deployment factors like device durability or training requirements.

Significance and Claims
The paper claims that while ML-based analysis of ECG and PCG signals presents a "transformative opportunity" for accessible RHD screening, the field is currently hindered by a "disconnect between model performance and clinical readiness."

• Technical Feasibility vs. Clinical Readiness: The authors assert that while deep learning models have achieved high accuracy in controlled settings, they have not yet been proven effective in real-world, resource-constrained environments. The lack of external validation and prospective trials means these tools are not yet ready for widespread clinical adoption.
• Strategic Recommendations: The review concludes that bridging the diagnostic gap requires a shift toward implementation science. Specifically, the authors call for:
  • Standardized benchmark datasets to facilitate cross-study comparison.
  • Prospective, multi-center trials in endemic areas.
  • Research into severity stratification (moving beyond binary classification) to guide clinical intervention.
  • Development of Explainable AI (XAI) to build clinician trust.
• Global Health Impact: The paper emphasizes that addressing these gaps is essential to achieving the World Heart Federation's "25 by 25" goal. Without rigorous validation and feasibility studies, the potential of AI-augmented auscultation to prevent thousands of disability-adjusted life years in underserved populations remains unrealized. The authors note that current global R&D funding for RHD is negligible compared to other infectious diseases, further complicating the path to scalable solutions.