ECGNextFormer: A Lightweight, Scalable, and Explainable Architecture for Estimating Cardiovascular Well-being from Electrocardiograms.

An electrocardiogram (ECG) is a widely utilized non-invasive test in healthcare for diagnosing cardiac diseases. Recent studies have validated that the delta between neural network-derived ECG age and chronological age represents a measure of cardiovascular well-being. Typically, ECG age estimation studies rely on neural networks designed for ECG disease diagnosis, resulting in networks that are often computationally heavy, difficult to scale, and challenging to interpret. To address this gap, we propose ECGNextFormer, a lightweight hybrid CNN and transformer architecture for efficient ECG age estimation. ECGNextFormer features a novel shallow backbone with modified ECGNext blocks for ECG feature extraction and integrates a custom self-attention layer with a multiscale pooling block to enhance the receptive field. The bottleneck drop key self-attention provides a global receptive field and reweighs the feature maps, while multiscale pooling retains fine-grained spatial details crucial for age estimation. These innovations make ECGNextFormer a lightweight and scalable alternative to existing networks with superior performance. The lite variant of ECGNextFormer surpasses the best network in the literature by 14.7% in performance and 53× lower parameter count. ECGNextFormer is comprehensively explained through fine-grained heatmaps generated using guided backpropagation. Subsequently, a novel lead scoring scheme quantifies lead significance in healthy and diseased groups, and human-interpretable concepts highlight essential ECG features for age estimation.