Figure 4.

Real-time prediction comparison of a sample patient session without AURA learning (left panels) and with AURA learning (right panels). The dashed vertical red line represents the ground truth seizure onset time, and the solid green line is the weak label generated by the seizure detection model. In the example, these two lines overlap which means the detection model correctly classified the actual seizure onset. The peak that precedes the seizure event corresponds to the prediction system forecasting a high probability that a seizure will occur within the next hour. The peak in AURA learning is far more distinguishable than that without AURA learning. The sample is taken from the second session of EEG recordings, resulting in different probabilities at t = 0 s as the model has undergone one session of AURA training.