ECG Feature Learning by Using Rational Variable Projection Autoencoders

Szabari, Mátyás Márton and Bognár, Gergő and Kovács, Péter (2024) ECG Feature Learning by Using Rational Variable Projection Autoencoders. In: 51st Computing in Cardiology. IEEE, Karlsruhe. (In Press)

Text CinC24_Matyas3.pdf - Published Version Restricted to Registered users only Download (181kB) | Request a copy

Abstract

In this paper, we propose a model-based shallow autoencoder structure to automatically extract features from electrocardiogram (ECG) data. The encoding path in our model employs parametrized orthogonal transformations by means of rational function systems, and utilizes Variable Projections (VP) to compute low-dimensional representations of individual heartbeats. After the global training of this rational VP autoencoder, we used the linear coefficients of the projections in the encoding as ECG heartbeat features. We evaluated the performance of the proposed feature learning scheme on the standard 5-class AAMI heartbeat classification problem using the benchmark MIT-BIH Arrhythmia Database, training separate support vector machine and random forest classifier models on the extracted features. Employing the subjectoriented (inter-patient) evaluation scheme, we achieved an accuracy exceeding 94%. This performance is comparable to other state-of-the-art ECG classification approaches, while providing a computationally simple and explainable method for learning features from raw ECG data.

Actions (login required)

Edit Item