Enhancing Explainability and Performance in ECG Analysis: The Role of R-Peak Alignment in Supervised and Unsupervised Machine Learning

Background: The integration of machine learning (ML) models in cardiovascular diagnostics has revolutionized the use of the electrocardiogram (ECG) for patient care, emphasizing the need for both accuracy and interpretability. Traditional convolutional neural networks (CNNs), while robust in performance, often fall short in providing the transparency required for confident clinical decision-making. This has led to increased interest in enhancing model explainability. This study investigates the effectiveness of R-Peak alignment in enhancing both the explainability and the performance of shallow ML models for ECG analysis. We evaluate its impact on classification accuracy and the coherence of clustering within ECG data, utilizing both supervised and unsupervised learning approaches.

Methods: Utilizing the PTB-XL ECG dataset, containing 21799 12-lead resting-state ECGs (10s with a 500 HZ sampling rate) from 18,869 patients, we conducted experiments to compare the performance of logistic regression and support vector machines (SVMs), both with and without R-Peak alignment, against standard CNN architectures. We also examined the dynamics of clustering using hierarchical and k-means clustering on aligned and non-aligned data. Performance metrics included macro-averaged area under the curve (AUC) for classification and V-measure scores for clustering.

Results: R-Peak alignment markedly improved classification outcomes of shallow machine learning models, with SVMs reaching an AUC up to 93.4%, notably surpassing CNNs especially in scenarios with limited data. Logistic regression also showed significant gains, reducing expected calibration error (ECE) from 0.2 to 0.05, enhancing the reliability and interpretability of probability predictions. In unsupervised learning, R-Peak alignment enabled clustering of ECG signals, shifting V-measure scores of near zero for unaligned data to 15%, indicating reasonable alignment of cluster formations with expert-annotated diagnostic categories. Importantly, the use of shallow machine learning models like SVMs allowed for detailed analyses of feature importance with permutation importance techniques. Application of R-peak alignment also enabled aggregated importance analyses for CNNs with integrated gradients. Without alignment IGs can only be calculated for individual ECGs, limiting global explainability across different recordings. Using R-peak alignment on individual IG maps to transform these maps to an R-peak aligned space allowed global aggregation of importance metrics across all instances of a dataset.

Conclusion: R-Peak alignment substantially enhances both the performance and the explainability of ML models in ECG analysis. In supervised learning, it empowers shallow models to meet or exceed the performance of more complex models, especially in data-limited settings. In unsupervised learning, it unveils data structures, improving the interpretability of clustering outcomes. The results strongly support the broader adoption of R-Peak alignment in clinical ML applications, promoting enhanced diagnostic accuracy and deeper insights into the models' decision-making processes.