Classification of 12-lead ECGs Using Intra-Heartbeat Discrete-time Fourier Transform and Inter-Heartbeat Attention

Abstract

In this work, we built a model to classify 12-lead ECGs using attention for the PhysioNet/Computing in Cardiology Challenge 2020. Since information about different classification outcomes might be present only in specific segments, we tune our feature representation to show the frequency distribution shift as we move through time. This is done by first representing the original signal as a spectrogram, which shows the signal's frequency spectrum during different time windows (heartbeats). The frequency spectrum at each heartbeat is extracted using discrete-time Fourier transform. The spectrogram is then inputted to a bidirectional LSTM network where each heartbeat vector represents a time step. The outputs of the bidirectional LSTM network at each stage are then used as attention vectors. The attention vectors are then multiplied with the original signal window embeddings, which are used to generate the final output. Our approach achieved a challenge validation score of 0.416 and a test score of 0.024 but were not ranked due to omissions in the submission (team name: SBU_AI).