Chagas disease is a tropical parasitic disease that is widespread in some regions. Since widespread serological testing for Chagas disease is impossible due to limited capacities, one of the solutions is to make an initial screening based on electrocardiogram (ECG) results. To classify Chagas disease, we propose an extension of our modified two-branch Neural Basis Expansion Analysis for Interpretable Time Series (N-BEATS) architecture, previously used for fast classification of cardiovascular diseases. N-BEATS is rarely used for classification purposes, but we want to show that with proper modifications, it can be a valid and fast solution and alternative to other state-of-the-art networks. The paper proposes a novel, multi-branch architecture for the N-BEATS network. It allows for better separation of datasets with different labels and credibility, making it highly suited for this year's George B. Moody challenge, in which one of the presented problems is using small, strongly-labeled datasets along large, weakly-labeled datasets. This modification also allows for better use of the metadata and signal modifications (like wavelet transformations or drift removal) while not losing the essential elements of pure ECG signal. For performance comparison, we also show the effects of the exact modification on two other popular state-of-the-art networks: long-short-term memory (LSTM) and gated recurrent unit (GRU).
Our team (WEAIT) sent the first simplified (for performance reasons) version of the proposed N-BEATS architecture during the unofficial phase, and it was scored 17th (out of more than 173 entries) with a 0.605 challenge score. LSTM in the same architecture achieved a challenge score of 0.294.