Fig 1.
Individual EEG electrode signals are fed through a 1D CNN (left). The sequences of representations are fed through the BLSTM layer and then classified for seizure activity in each electrode.
Fig 2.
Localization zones and electrode connectivity graph.
Partition of EEG electrodes into zones to train our network based on coarse hemisphere (a) andanterior and posterior head regions (b).
Fig 3.
Electrode connectivity graph used in GCN baselines.
Table 1.
Patient demographics and clinical attributes for our JHH evaluation dataset (N = 34) and UWM generalization dataset (N = 15).
Table 2.
Performance on the JHH dataset.
Seizure detection results on the JHH dataset. Window-level metrics are aggregated across one-second segments of the EEG. Seizure level results are calculated over the duration of the seizure interval.
Table 3.
Generalization detection results on the UWM dataset.
Seizure detection performance when applying the JHH models to data from UWM. We ran a LOPO-CV on UWM to calibrate the seizure versus baseline detection threshold. However, we did not retrain the neural network weights.
Fig 4.
Seizure activity tracking in two JHH patients. Clinical SOZ annotations are given for each patient. Where clinical annotations are provided, images show seizure activity tracking corresponding to annotation times.
Fig 5.
SZTrack and No-BLSTM output comparison.
Channel-wise predictions for the fronto-temporal seizure shown on the top row of Fig 4 are superimposed on the EEG signal. In (a) SZTrack makes a confident prediction of seizure onset in the temporal channels which spreads to the parietal and frontal areas. In (b) No-BLSTM responds to isolated seizure activity at the onset but does not provide a temporally stable prediction.
Fig 6.
Average localization accuracy in JHH when varying the weight on the detection loss. Boxplots are shown for the SZTrack, No-BLSTM, and TGCN models. A horizontal dashed line shows performance for the CNN-BLSTM model.
Fig 7.
Localization results from the JHH dataset.
Patient-wise lateralization and lobe classification for SZTrack in JHH. Predicted SOZ locations are superimposed on the head figure in red. The small circle indicates the coarse clinical SOZ annotation, where green indicates concordance with clinical annotations and red circle indicates disagreement. SZTrack correctly localizes both the hemisphere and lobe in 21 of 34 patients. In 12 of 34 patients, SZTrack correctly localizes either hemisphere or lobe; it misses completely in just one patient.
Fig 8.
UWM dataset generalization results.
Lateralization and lobe classification results when applying a SZTrack model trained on JHH to data from UWM. Predicted SOZ locations are shown superimposed on the head figure in red. The small circle indicates the coarse clinical SOZ annotation.