Context-Conditioned Spatio-Temporal Modeling of Multimodal Neurophysiological Signals for Interictal Epileptiform Discharge Detection
Abstract
Interictal epileptiform discharges (IEDs) are the principal electrophysiological biomarker of epilepsy, yet their automatic detection from short physiological windows remains difficult: the discharges are rare, morphologically diverse, buried in ocular and muscular artifacts, and their statistics vary strongly across patients and acquisition sessions. We tackle this problem using multimodal recordings that pair scalp electroencephalography (EEG) with electrocardiography (ECG) and electromyography (EMG), and argue that a fixed feed-forward network is poorly matched to such heterogeneous, shifting inputs. We instead propose a context-conditioned spatio-temporal network that decouples what is computed from how it is adapted. A signal trunk first extracts per-channel temporal morphology with a shared convolutional frontend and a dilated temporal convolutional network then models cross-channel and cross-modal dependencies with a channel-wise Transformer over the 26 physiological nodes. In parallel, a lightweight context branch summarizes each recording into a latent vector by fusing a variational time-domain statistical embedding and a modality-aware spectral embedding through a learned per-dimension convex combination. This latent conditions the entire trunk through feature-wise linear modulation (FiLM) at twenty-three normalization and feature points, allowing the network to re-parameterize its computation per recording without any test-time weight update. A five-way discharge-subtype head provides auxiliary weak supervision on positive windows. On the NeuroMM 2026 Track 1 benchmark, the proposed model reaches 0.9531 area under the precision–recall curve (AUPRC) as a single model, and multi-split probability averaging further raises it to 0.9693. Ablations show that temporal modeling, cross-channel aggregation, fused context conditioning, and moderate auxiliary supervision each contribute, with context conditioning being the single largest architectural factor.