Epileptic EEG classification faces significant challenges in practical applications, particularly due to limited sample sizes and substantial inter-subject variability. This issue is even more pronounced under few-shot conditions, where the model’s stability and generalization ability still need improvement. To address this, we propose an Adaptive Task-aware Multi-Scale Convolutional Network (ATMSNet) designed for few-shot EEG classification tasks in epilepsy diagnosis. The model leverages a multi-scale convolutional structure to perform time-frequency feature modeling of EEG signals, and incorporates a task-driven adaptive optimization mechanism to effectively model the differences in EEG feature distributions across different subjects. ATMSNet employs a multi-path parallel feature extraction strategy to capture discriminative features at various time scales in epileptic EEG, while introducing a feature-level adaptive modulation mechanism to adjust intermediate feature representations in a lightweight manner, enhancing the model’s adaptability in few-shot task scenarios. The model's training process consists of two stages: full-data supervised pre-training and task-driven adaptive optimization. This ensures feature discriminability while improving the model's ability to quickly adapt to new subjects. Experimental results on the Bonn and CHB-MIT epilepsy EEG datasets show that ATMSNet outperforms baseline few-shot learning methods in classification performance, demonstrating superior stability and generalization ability, thereby validating its applicability to few-shot epileptic EEG classification scenarios.
Chen et al. (Fri,) studied this question.