A comparative study on neural networks for paroxysmal atrial fibrillation events detection from electrocardiography.

OBJECTIVE

This work conducts a comparative study on the effect of neural networks of different architectures on the detection of paroxysmal atrial fibrillation (PAF) events from dynamic electrocardiography (ECG) recordings, a problem raised in the 4th China Physiological Signal Challenge 2021 (CPSC2021).

APPROACH: We proposed 3 neural network models and an auxiliary one for QRS detection to tackle the problem. A convolutional recurrent neural network (CRNN) model and a U-Net model that accepts ECG waveform input make sample-wise predictions. This regards the PAF events detection as a segmentation task. A stacked bidirectional long short-term memory (LSTM) model takes the sequence of RR intervals, which is derived from the output of the QRS detection model and makes beat-wise predictions. The QRS detection model also has a CRNN architecture, which is slightly different from the model for the AF segmentation task. Final predictions are merged by outputs from models making sample-wise predictions and making beat-wise predictions. Finally, the locations of QRS complexes are used to filter out segments (both normal and AF) shorter than 5 beats. In order to make the neural network models more sensitive to the critical sample points (onsets and offsets) of the AF events, we proposed a novel masked binary cross-entropy (MaskedBCE) loss function for training the models. This loss function is the conventional BCE loss multiplied by a mask, whose values in a neighbourhood of critical sample points are significantly larger than elsewhere.

MAIN

RESULTS

Our method received a score of 1.9972 on the first part of the hidden test set of CPSC2021 and a score of 3.0907 on the second part. The average score was 2.5440, ranked 5th out of 17 teams with successful official entries.

SIGNIFICANCE: This work proposed an effective solution to the problem of the detection of PAF events from dynamic ECGs and validated the efficacy of several neural network architectures on this problem.