Ong, Jia Heng (2023) Development of functional near-infrared spectroscopy-based brain-computer interface model using principal components and latent variables artificial neural networks. Masters thesis, Universiti Tun Hussein Onn Malaysia.

Preview	Text 24p ONG JIA HENG.pdf Download (777kB) | Preview
	Text (Copyright Declaration) ONG JIA HENG COPYRIGHT DECLARATION.pdf Restricted to Repository staff only Download (447kB) | Request a copy
	Text (Full Text) ONG JIA HENG WATERMARK.pdf Restricted to Registered users only Download (3MB) | Request a copy

Abstract

Functional near-infrared spectroscopy (fNIRS) is a non-invasive brain imaging technology in brain-computer interface (BCI) applications. Poor signal quality and massive irrelevant signals are the shared challenges of fNIRS signal post-processing in real-life. Principal component analysis (PCA) and partial least squares (PLS) extracted principal components (PCs) and latent variables (LVs) are the practical feature extractors in near-infrared chemometric analysis, but both features were rarely studied for fNIRS-BCI application. Thus, this study investigates the feasibility of PCs and LVs as the signal features in fNIRS-BCI model development. First, fNIRS signals were analysed using different signal denoise filters, channel compressions, and feature extraction approaches to produce optimal PCs and LVs features. Next, both features were applied as ANN training inputs to produce PCs-ANN and LVs-ANN in 10-folds cross-validation for fNIRS-BCI applications. The PCs-ANN outperformed benchmark statistical approach literature with average classification accuracy of 71.67% in four class classification. The optimal LVs-ANN outperformed partial least square discriminant (PLS-DA) and traceable best-performed literature of bagging classifiers. The classification accuracies of LVs-ANN, PLS-DA, and the bagging classifier were 84.94%, 77.05%, and 74.00%, respectively, in binary classification. Findings show the PLS is effective in eliminating less relevant training features to improve fNIRS-BCI performance. In conclusion, this study validated the feasibility of PCA and PLS feature extractors by demonstrating reliable performance for fully imagine tasks-based fNIRS-BCI applications, including hand motion imagine and word generation. This demonstrates that both feature extractors were essential to the development of the fNIRS-BCI device for patients with total paralysis

Actions (login required)

View Item