Figure 4.

The ECG pair encoder is adapted from Bidirectional Encoder Representations from Transformers (BERT)’s sequence pair encoder. It extracts joint features from the two input ECG sequences. ${T^{″}}_{k1},{T^{″}}_{k2},\dots ,{T^{″}}_{k22}$ are the final hidden states that correspond to ${T^{\prime }}_{k1},{T^{\prime }}_{k2},\dots ,{T^{\prime }}_{k22}$ respectively; ${T^{″}}_{q1},{T^{″}}_{q2},\dots ,T{T^{″}}_{q22}$ are the final hidden states that correspond to ${T^{\prime }}_{q1},{T^{\prime }}_{q2},\dots ,{T^{\prime }}_{q22}$ respectively. $L_{kq\left(\mathrm{VE}\right)}$ is the final hidden state that correspond to $cl{s}_{\mathrm{VE}}$; $L_{kq\left(\mathrm{ID}\right)}$ is the final hidden state that correspond to $cl{s}_{\mathrm{ID}}$.