In this paper, we present a novel interpretable framework for electrocardiogram (ECG)-based disease detection by combining HDC and learnable neural network encoding. Unlike conventional HDC methods that rely on static random projections, we introduce a rhythm-aware learnable encoding pipeline based on RR intervals, a physiological signal segmentation strategy that matches the cardiac cycle. The core of the HDC architecture is a neural distillation HDC architecture featuring a learnable RR-block encoder and a BinaryLinear high-dimensional projection layer, which jointly optimizes cross-entropy and proxy-based metric losses. This hybrid framework enables task-adaptive representation learning while maintaining the symbolic interpretability of HDC. Experimental results on Apnea-ECG and PTB-XL datasets show that it outperforms conventional HDC and classical machine learning baseline models, achieving a precision of 73.09% and an F1-score of 0.626 on Apnea-ECG, and similar robustness on PTB-XL. This framework provides an efficient, scalable, edge computing compatible ECG classification solution with strong potential for interpretable and personalized health monitoring.
