Aims: We aimed to develop 4 multi-label classification models on 12-lead, 6-lead (I, II, III, aVL, aVR, and aVF), 3-lead (I, II, V2), and 2-lead (II and V5) ECG recordings, which can automatically identify cardiac abnormalities out of 27 ECG abnormalities designated by Challenge organizers.
Methods: Recordings without a label in the 27 scored classes were excluded and recordings were downsampled from 1,000 Hz to 500 Hz to make the sampling frequency of all training data consistent. Signals were fixed at the same length of 30 seconds. Biorthogonal wavelet transformation (bior2.6) was applied to reduce the noise in ECG signals. We trained a baseline model using 12-lead signals to re-label the data due to the uneven distribution of labels and missing labels. Considering that different ECG abnormalities may be more apparent in specific leads and an equal importance of different leads could cause information losses leading to misdiagnosis, we set up our models using SE_ResNet which is a variant of a ResNet that employs squeeze-and-excitation blocks to enable the network to perform dynamic channel-wise feature recalibration. For the class imbalance issue, we designed an improved multi-label Sign Loss: for the correctly classified labels, a coefficient smaller than 1 was multiplied to the default binary cross-entropy loss. By doing so, the accumulated loss from a large number of true negative labels became smaller, and the loss from the misclassified labels became more prominent.
Results: Our models achieved a Challenge metric score of 0.352, 0.350, 0.341, and 0.331 on 12-lead, 6-lead, 3-lead, 2-lead on the released training data and a Challenge metric score of 0.340, 0.340, 0.342, and 0.342 on 12-lead, 6-lead, 3-lead, 2-lead validation datasets respectively. Our team name is HeartBeats.
Conclusion: Our models can automatically classify multi-label cardiac abnormalities with good performance on standard twelve-lead EGCs and reduced-lead ECGs.