A multi-source classification framework with invariant representation reconstruction for dual-target RSVP-BCI tasks in cross-subject scenario

• Published in: Neurocomputing (Amsterdam) Vol. 620; p. 129239
• Main Authors: Chen, Hongying, Wang, Dan, Xu, Meng, Chen, Jiaming, Zhang, Yueqi, Chen, Yuanfang
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2025
• Subjects: Autoencoder; Brain-Computer Interface; Calibration; RSVP; Transfer Learning; Brain-Computer Interface; Transfer Learning; Calibration; Autoencoder; RSVP
• ISSN: 0925-2312
• DOI: 10.1016/j.neucom.2024.129239

The Rapid Serial Visual Presentation (RSVP) is a widely used paradigm for target detection tasks in Brain-Computer Interface (BCI) by decoding Electroencephalogram (EEG) signals. One major issue concerns the time-consuming calibration in cross-subject scenarios, which worsens in dual-target RSVP-BCI tasks. A new method is desperately needed to detect two targets further with less calibration time. This paper proposed a novel framework named Cross-subject Invariant Representation Extraction-Targeted Stacked Convolutional Autoencoder (CS-IRE-TSCAE) based on reconstructing the invariant representation. After filtering the source subjects, the CS-TSCAE alleviates the subject-dependent effect by reconstructing the invariant representation generated by CS-IRE. It was validated on the ERP datasets from the BCI Controlled Robot Contest 2022. The experimental result showed that CS-IRE-TSCAE obtained the highest Recall, F1 and Average ACC with significant differences both in subject-dependent and inter-subject experiments. It demonstrated that CS-IRE-TSCAE achieved a higher classification performance for dual-target RSVP with less calibration time. Our framework drives the application development of target detection in RSVP-BCI by facilitating multiple target detection in cross-subject scenarios, which has practical significance, especially in fast-deployment scenarios.