Introduction: State-of-the-art deep learning methodologies offer highly effective ways for extracting optimal representations from medical data. However, these are typically model-agnostic, end-to-end methods that demand substantial computational power. Therefore, in this paper, we propose a lightweight model-based learning approach, consisting of only 6 trainable parameters, to automatically extract features from ECG data.
Methodology: We developed a model-based shallow autoencoder structure, where the encoder employs parametrized orthogonal transformations by means of the so-called real Malmquist-Takenaka basis. Utilizing the Variable Projections (VP) in the encoding path, our approach computes low-dimensional representations of individual heartbeats, incorporating both linear and non-linear parameters. The trainable non-linear parameters comprise the poles of rational basis functions, represented in hyperbolic geodetic polar coordinates.
Results: We evaluated the efficiency of the proposed feature learning scheme in the standard 5-class AAMI heartbeat classification benchmark problem using the records of the MIT-BIH Arrhythmia Database. To investigate the generalization ability of the proposed feature learning method, the training was not patient-specific; instead, non-linear parameters were globally trained and tested on patient-wise distinct subsets of the entire dataset. After training the rational VP autoencoder, we used the linear coefficients of the projections in the encoding as the features extracted from the ECG measurements. Subsequently, a separate classifier model, such as SVM or Fisher discriminant analysis, was trained on the extracted features to distinguish normal and abnormal heartbeat signals. Employing the subject-oriented evaluation scheme proposed by de Chazal, we achieved an accuracy exceeding 94%. This performance is comparable to other state-of-the-art ECG classification approaches, while providing a computationally simple method for learning features from raw ECG data.