Beat Analysis of Dimensionality Reduced Perspective Streams from Electrocardiogram Signal Data

Abstract

Exploration of ECG traces for their spectral characteristics, distinctly discloses a random component that pertains to the triggering of a new cardiac cycle in the inter-beat interval. Yet the stream consistently shows impressive reproducibility of its intrinsic core waveform. Respectively, the presence of close to deterministic structures firmly contends for representing a single cycle ECG wave by a state vector in a low dimensional embedding space. Rather than performing arrhythmia clustering directly on the high dimensional state space, our work first reduces the dimensionality of the extracted raw features. Analysis of heartbeat irregularities becomes then more tractable computationally and thus increasingly relevant to run on emerging wearable and IoT devices that are severely resource and power constraint. In contrast to prior work that constructs a two dimensional embedding manifold, we project feature vectors onto a three coordinate frame of reference. This merits an essential depth perception facet to a specialist that qualifies cluster memberships, and furthermore, by removing stream noise, we managed to retain a high percentile level of source energy. We performed extensive analysis and classification experiments on a large arrhythmia dataset, and report robust results to support the intuition of expert-neutral similarity.