Figure 2.

Architecture of the future prediction framework. The model predicts the correct future frame among negative samples (right) based on the present frames (left). Encoder E extracts features from ECG frames, producing a feature vector for each frame. Attention pooling summarizes feature vectors into a single context vector c describing the present. A dot product between c and frame encodings $h_i$ gives the similarity between the context and future frames. The entire model is trained end-to-end with gradients backpropagated from the cross-entropy loss of classifying the future frame correctly.