Seizure detection using hierarchical temporal trend integration and self-supervised learning

Early, reliable detection of epileptic seizures from electroencephalography (EEG) remains challenging due to label scarcity, inter patient variability, and the non stationary nature of clinical recordings. This work introduces Federated Temporal Learning (FTL), an annotation efficient framework that couples a supervised seizure classifier with a self supervised pretext task, Temporal Sequence Consistency Learning (TSCLM), over a shared backbone. The backbone is built from Hierarchical Temporal Trend Integration Modules (HTTIMs) that extract and top down mix multi scale long and short term trends; a Temporal Self Supervision Head (TSSH) learns order relations between subsequences to inject temporal positional awareness without manual labels. Training optimizes a joint objective with a mixing coefficient δ and split ratio τ; at inference the self supervised head is discarded, yielding a purely supervised predictor. We theoretically analyze the algorithmic stability of our algorithm, and indeed, provide a bound on the gap between the empirical and the expected risk under the semi supervised objective, thereby clarifying the interplay between more labeled data and well conditioned optimization. From an empirical perspective, it also performs seamlessly at state of the art (SoA) on two datasets: TUSZ v2.0.1 (accuracy 99.16%, precision 99.19%, recall 99.65% and F1 macro 97.17%), Zenodo NICU EEG (accuracy 95.61%, precision 96.78%, recall 94.67% and F1 macro 96.87%). The results suggest that order aware self supervision in conjunction with hierarchal trend integration can be a viable solution to the label efficient and clinically reliable, EEG seizure detection task. The proposed approach has the advantage of explicitly modelling multi-scale temporal patterns by adopting a top-down strategy, using Hierarchical Temporal Trend Integration Modules (HTTIM), which is not used in existing approaches. In addition, Temporal Sequence Consistency Learning (TSCLM) is designed as an order-aware self-supervised task to capture temporal relationships without labels. The integration of these components within the Federated Temporal Learning (FTL) framework enables improved performance under limited annotated data.