Clip-to-trial: EEG-based decoding network with long short-term self-attention for DRDS recognition

Dynamic random-dot stereograms (DRDS) elicit stereoscopic perception associated with high-level binocular visual processing, which can be decoded from EEG signals. However, existing EEG-based approaches predominantly rely on clip-level learning, leading to feature-label misalignment and limited modeling of trial-level spatiotemporal dependencies. To address these limitations, this paper proposes a clip-to-trial long short-term self-attention network (C2T-LSTSA) for EEG-based DRDS recognition. The proposed method first extracts discriminative spatiotemporal features from EEG clips using a multi-scale local feature extraction module. A trial-level feature reconstruction module is then introduced to explicitly map clip-level representations into semantically consistent trial-level features. Furthermore, a long short-term self-attention module is employed to jointly model local temporal dependencies within each reconstructed trial representation and global dependencies among multiple reconstructed trial representations derived from the same original EEG trial. Extensive experiments on three EEG datasets collected from strabismus patients demonstrate that the proposed method consistently outperforms state-of-the-art approaches under both subject-dependent and subject-independent evaluation protocols, indicating its effectiveness and robustness for objective assessment of stereoscopic visual function. The reference datasets are available at https://github.com/YANGeeg/TJU-SRD-datasets and the code is at https://github.com/wu860898-design/C2T-LSTSA.