Discrimination of healthy controls and selected visually impaired through visually evoked potentials

Abstract

This thesis presents a digital signal processing based detection of healthy controls and selected visually impaired through visually evoked potentials (VEP). Visual impairment is a term used by ophthalmologist to describe any kind of vision loss, whether it's partial or total vision loss. Some of the conventionally used techniques for the investigation of vision impairments include fundoscopy imaging, ultrasound imaging, and manual inspection of retina. These techniques have several disadvantages such as poor quality of images produced by the ultrasound imaging, require experts, and are prone to error in manual inspection. The VEP provides an objective method for the diagnostics of vision impairments in patients. VEP is an electrical signal generated by the brain (Occipital Cortex) in response to a visual stimulus. By analyzing these responses, the abnormalities in the visual pathways of a person can be detected. The development of feature extraction and classification algorithms for investigation of vision impairments through VEPs however is still at an infancy level. Therefore, this study was carried out to investigate the time, frequency, and time-scale/frequency characteristics of the single trial transient VEPs, and propose an efficient feature extraction and classification algorithm for distinguishing the vision impairments. Four different feature extraction methods based on time, frequency, wavelet, and Stockwell transform were explored and statistical features were proposed for the VEP analysis. A new feature augmentation technique was proposed to enhance the variation of the data prior to the analysis. Three different feature reduction techniques were used to reduce the dimensional space of the features. Extreme learning machine, least square support vector machine and probabilistic neural networks were employed to evaluate the performance of the features in discriminating the vision impairments. Statistical analysis were used to demonstrate the significance of the preprocessed features, while performance measures such as sensitivity, specificity, positive predictivity, negative predictivity, and overall accuracy was considered for the evaluation of the classifiers. The dataset from two different experimental settings were used in the analysis. The first experiment was conducted to investigate the effect of different sizes of checkerboard stimulus to the resulting evoked responses while the second experiment was perpetrated to investigate the performance of the new colour fusioned checkerboard stimulus in elicitating reliable VEP responses. The experimental investigation elucidate that features derived from the VEP elicited by the proposed stimulus performed well in classifying the vision impairments. Promising 100% accuracy was achieved using the combinations of the proposed stimulus and feature extraction methods.