Introduction: The large availability of biomedical data allows to develop reliable medical tools for detecting cardiovascular diseases using data-driven machine learning approaches. In fact, state-of-the-art deep learning (DL) methods provide promising results in the classification of various arrhythmia types. On the other hand, in clinical applications, improving classification accuracy alone is usually not enough, as physiological interpretation of the results are also important. Besides the lack of explainability, high computational complexity is another drawback of the top performing DL models.
Methods: In this paper, we alleviate the previously mentioned problems by combining variable projections (VP) with spiking neural networks (SNN). VPs are nonlinearly parametrized orthogonal projections whose weights have physical meaning, whereas SNNs are brain-inspired network topologies, which can be implemented with ultra-high speed and ultra-low energy consumption on neuromorphic devices. In order to combine the advantages of VPs and SNNs, we design a hybrid neural network model called VPSNN. This architecture is adapted specifically to ECG data analysis such that the first VP layer servers as an automatic feature extractor and spike encoder whose weights represent the positions and the widths of the clinically relevant ECG waveforms (P-QRS-T). The VP layer constructs a clinically interpretable latent feature space and is trained together with the subsequent SNN layers, which are responsible for the feature analysis and the actual classification of the data.
Results: As a case study, we consider the classification of normal and ventricular ectopic beats (VEBs) in real ECG recordings of the PhysioNet MIT-BIH Arrhythmia Database. Our experiments show that the proposed VPSNN architecture can be effectively used for detecting VEBs with an average classification accuracy of 96%, comparable to the state-of-the-art. Moreover, due to highly compact topology of VPSNN, it offers a low computational cost inference ability suitable for edge computing in clinical applications.