Comparison of neural basis expansion analysis for interpretable time series (N-BEATS) and recurrent neural networks for heart dysfunction classification

Abstract

Objective. The primary purpose of this work is to analyze the ability of N-BEATS architecture for the problem of prediction and classification of electrocardiogram (ECG) signals. To achieve this, performance comparison with various types of other SotA (state-of-the-art) recurrent neural network architectures commonly used for such problems is conducted. Approach. Four architectures (N-BEATS, LSTM, LSTM with peepholes, GRU) were tested for performance and dimension reduction problems for different number of leads (2, 3, 4, 6, 12), both in variants consisting of blended branches, allowing retaining accuracy while reducing the computational capacity needed. The analysis was performed on datasets and using metrics from Challenges in Cardiology (CinC) 2021 competition. Main results. Best results were achieved for LSTM with peepholes, then LSTM, GRU and the worst for N-BEATS (challenge metrics respectively: 0.42, 0.40, 0.39, 0.35; for times: 0.0395 s, 0.0036 s, 0.0027 s, 0.0002 s). Commonly used LSTM outperforms N-BEATS in terms of multi-label classification, data set resilience, and obtained challenge metrics. Still, N-BEATS can obtain acceptable results for 2 lead classification (metric of 0.35 for N-BEATS and 0.38 for other networks) and outperforms other solutions in terms of complexity and speed. Significance. This paper features a novel approach of using the N-BEATS, which was previously used only for forecasting, to classify ECG signals with success. While N-BEATS multi-label classification capacity is lower than LSTM, its speed obtaining results with a reduced number of leads (faster by one to two degrees of magnitude) allows for arrhythmias detection and classification while using off-the-shelf wearable devices (Holter monitors, sport bands, etc).