Hsiam: A Generalized Siamese Network for Efficient Multi-Resolution EEG Analysis

High-frequency electroencephalogram (EEG) offers richer spectral representations that can reveal neural features invisible in low-resolution recordings. However, its practical deployment is constrained by several factors, including high computational cost, limited clinical availability of high-sampling-rate EEG systems, and susceptibility to amplified high-frequency noise. To address these issues, we propose the High-Frequency Siamese Network (Hsiam), a generalized dual-branch Siamese architecture designed for efficient cross-resolution EEG analysis. Hsiam processes inputs of different frequency resolutions through weight-sharing parallel branches and enforces cross-resolution consistency via a dedicated alignment loss. Within this framework, the High-Frequency Activity Enhanced Module (HEM) facilitates discriminative learning by emphasizing task-relevant high-frequency components, while the Frequency-Domain Dropout Transformer (FD-former) models temporal-spectral dependencies in the frequency domain to enhance robustness against noisy and redundant frequency information. Importantly, Hsiam maintains high practicality by requiring only a single-branch input during inference, significantly reducing computational overhead without compromising accuracy. Extensive experiments on one self-built and two public EEG datasets demonstrate that Hsiam achieves strong performance across both treatment efficacy prediction and seizure detection tasks. Further branch-, frequency-, and channel-level analyses show that its performance gains are associated with training-time cross-resolution alignment, task-dependent spectral utilization, and non-uniform spatial channel contributions.