TL-P3GAN: An Efficient Temporal-Learning-Based Generative Adversarial Network for Precise P300 Signal Generation for P300 Spellers

• Published in: IEEE transactions on cognitive and developmental systems Vol. 16; no. 2; pp. 692 - 705
• Main Authors: Bhandari, Vibha, Londhe, Narendra D., Kshirsagar, Ghanahshyam B.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.04.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Brain modeling; Brain–computer interface (BCI); Classification algorithms; Data augmentation; Datasets; Electroencephalography; generative adversarial network (GAN); Generative adversarial networks; Generators; Machine learning algorithms; Morphology; P300 speller; Qualitative analysis; Signal classification; Signal generation; Standard data; temporal learning; Visualization
• ISSN: 2379-8920; 2379-8939
• DOI: 10.1109/TCDS.2023.3288201

The problem of data imbalance among target and nontarget classes is inherent in the oddball paradigm-based P300 speller. This class imbalance is a critical issue and requires advanced rigorous learning to deal with. Conventionally, data level-like sampling approaches and algorithmic level-like ensemble approaches had been attempted in past research for data augmentation. However, information loss, overfitting, and subject variability were their major pitfalls. Alternatively, generative adversarial network (GAN)-based data augmentation managed to alleviate information loss but exhibits problems of overfitting and subject variability due to lack of temporal learning. To compensate for those problems, the authors have proposed novel temporal learning-based GAN (TL-P3GAN) to generate precise P300 signals and augment the minority class, i.e., P300. The TL-P3GAN comprises a novel contribution of multiscale morphological learning in both generator and discriminator. Moreover, the multiscale hybrid model in the generator learns multiresolution morphological information and considers sample-wise latency information from the original P300. The effectiveness of TL-P3GAN was confirmed by qualitative and quantitative evaluation metrics with two standard data sets. Further, the work is extended to analyze the effect of generated P300 signals on P300 classification and found significant performance improvement of 8%-16% with both data sets in comparison with existing conventional and GAN-based data augmentation approaches.