Introduction
Deep learning has shown significant potential in detecting seizures from electroencephalogram (EEG) signals. However, real-world deployment faces two critical hurdles: data privacy and inter-patient variability. Medical data is strictly protected by regulations like GDPR and HIPAA, preventing the centralization of large datasets. Furthermore, seizure patterns are often "non-IID" (not identically and independently distributed), meaning a model trained on one patient may perform poorly on another.
To address these challenges, researchers have developed a two-stage Federated Few-Shot Learning (FFSL) framework. This system enables collaborative training across multiple institutions without moving sensitive patient records off-site, while allowing for rapid, patient-specific adaptation using minimal labeled data.
Methodology: A Two-Stage Framework
The proposed framework utilizes a Biosignal Transformer (BIOT) architecture and operates in two distinct phases to balance global knowledge with local personalization.
Stage 1: Federated Fine-Tuning (E1)
The process begins with a BIOT model pre-trained on a massive corpus of 5 million EEG samples. This model is then fine-tuned across simulated hospital sites using Federated Averaging (FedAvg). During this stage, hospitals share only model weight updates with a central server, ensuring raw EEG recordings remain local. This phase establishes a "global" encoder capable of identifying general seizure features across a broad population.
Stage 2: Federated Few-Shot Personalization (E2)
To account for individual patient variability, the framework employs Few-Shot Learning (FSL). When the model encounters a new patient, it is adapted using a "support set" of just five labeled 5-second segments (typically 4 seizure segments and 1 background segment).
A critical component of this stage is the weighted local-global update. By using an interpolation parameter ($\alpha = 0.8$), the system prioritizes local, patient-specific signals while retaining the foundational knowledge gained from the global model.
Experimental Setup
The study was evaluated using the Temple University Hospital Event Corpus (TUEV), a standard benchmark for seizure detection. The dataset includes six categories: spike and sharp wave (spsw), generalized periodic epileptiform discharges (gped), periodic lateralized epileptiform discharges (pled), eye movement (eyem), artifact (artf), and background (bckg).
The researchers split the data into TUEV I for initial federated training and TUEV II for few-shot personalization and final evaluation, ensuring no patient overlap between the training and testing phases.
Key Results and Insights
The transition from a purely federated model to a personalized one yielded significant performance gains. In the initial federated stage (E1), the model achieved a balanced accuracy of 0.43, reflecting a "privacy tax" compared to the 0.52 achieved by a centralized model.
However, after the few-shot personalization stage (E2), performance improved dramatically. Across four heterogeneous sites, the client-specific models reached an average balanced accuracy of 0.77, a Cohen’s $\kappa$ of 0.62, and a weighted F1-score of 0.73.
As shown in the results, Client 2 achieved a balanced accuracy of 0.912, demonstrating the framework's effectiveness when local data aligns well with global features. Conversely, clients with unique seizure types not well-represented in the global training set saw more modest gains, highlighting that model performance is closely tied to the separability of seizure vs. background clusters in the embedding space.
Conclusion and Future Directions
The FFSL framework demonstrates that personalization is not merely beneficial but essential for clinical reliability in seizure detection. By combining federated learning with few-shot adaptation, the model achieves high accuracy while strictly adhering to privacy constraints.
Future work will focus on moving beyond simulated environments to real-world clinical networks, which often feature asynchronous participation and varying hardware. Additionally, researchers aim to explore adaptive mechanisms for the mixing parameter ($\alpha$) and extend the framework to support online adaptation, allowing models to evolve alongside patients in real-time.
Summary updated