A geometric insight into ECG classification: leveraging phase space reconstruction and distance to measure

We propose a geometric framework for robust ECG classification by combining phase space reconstruction (PSR) and the distance-to-measure (DTM) function to capture intrinsic cardiac dynamics. One-dimensional ECG signals are embedded into high-dimensional phase spaces, forming point clouds that reflect rhythm characteristics. Using median-of-means kernel density estimation, we identify high-density regions and construct DTM-based pseudo-distances, enhancing noise resilience and discriminative power. Evaluated on MIT-BIH and PTB datasets, our method achieves 89.08% accuracy with logistic regression and 91.42% with Gradient Boosting Trees under Gaussian noise. DTM features complement traditional statistical ones, demonstrating strong potential for nonlinear, interpretable, and noise-robust ECG analysis.