Epilepsy detection across patients with different electrode layouts: MOSAIC infers channel graphs and steadies them by self-distillation
Synopsis
Intracranial EEG electrode count and placement differ from patient to patient, so detectors with a fixed channel topology generalize poorly under subject-independent evaluation. MOSAIC encodes each channel with a shared convolutional encoder, then a single channel-attention layer infers a weighted inter-channel adjacency from the encoded features without any preset montage graph; an exponential moving average self-distillation term pulls each window's adjacency toward a slowly updated per-patient teacher. Across four public iEEG corpora, MOSAIC's F2 on both the SWEC-ETHZ short-term and long-term corpora exceeds the listed baselines, by 1.8 points over the strongest baseline on the short-term corpus and 3.3 points on the long-term one.
Interpretation
Detecting seizures in a new patient means facing a different number and arrangement of electrodes, and MOSAIC treats the inter-channel connectivity itself as something to be inferred rather than a montage graph fixed at training time. Earlier graph-neural-network detectors fixed the channel adjacency during training and assumed it held at deployment; Transformer-based approaches avoided fixed topology through self-attention but paid the quadratic compute and inference latency of full self-attention. On the SWEC-ETHZ short-term corpus MOSAIC reaches F2 51.1, 1.8 points above the strongest baseline in the table; on the long-term corpus it reaches F2 50.6, 3.3 points above the strongest baseline, while training in 100 minutes versus 300 minutes for the Multitask Transformer.
One shared one-dimensional convolutional encoder extracts local temporal features from every channel, channel attention turns those features into a per-window weighted adjacency, and graph convolution updates node features on that graph before pooling them into a window-level decision. The adjacency comes from a single shared self-attention layer whose queries and keys are normalized before the dot product, rescaled by a learned temperature and passed through a row-wise softmax, so no fixed electrode layout graph is required. Evaluation covers four public corpora: SWEC-ETHZ short- and long-term under a leave-one-patient-out protocol, and the single-channel Mayo and FNUSA clips under a grouped leave-one-group-out protocol, with all results averaged over 5 random seeds.
Because a per-window adjacency fluctuates with the non-stationarity of the signal, a self-distillation term pulls each window's adjacency toward a moving average of that patient's own past predictions, stabilizing structure without external supervision. Without self-distillation the short-term F2 is 48.3, comparable to the 48.8 of the uniform-adjacency control, so learning the adjacency alone brings no gain; only the regularized adjacency beats the uniform control, by 2.3 points. On the SWEC-ETHZ short-term corpus full MOSAIC reaches F2 51.1, macro-F1 64.1 and accuracy 73.1; removing the temporal encoder drops F2 to 31.2, the largest fall in the ablation.
Perspective
Evaluation is limited to four public intracranial EEG corpora: SWEC-ETHZ short- and long-term are multi-channel continuous recordings, while Mayo and FNUSA are single-channel fixed-length clips on which the channel-attention adjacency degenerates to the shared encoder. All results are averaged over 5 random seeds, and the cross-patient protocols are leave-one-patient-out or grouped leave-one-group-out.
The self-distillation teacher is only a smoothed copy of the model's own predictions, so the distillation loss reduces variance in the adjacency but cannot correct a systematic bias such as persistent concentration on a single dominant channel. The authors list extending the distillation target, for example with external supervision, and evaluating on scalp EEG as future work.