This dissertation, "Fast Evoked Potential Estimation by Artificial Neural Networks" by 馮順明, Shun-ming, Fung, was obtained from The University of Hong Kong (Pokfulam, Hong Kong) and is being sold pursuant to Creative Commons: Attribution 3.0 Hong Kong License. The content of this dissertation has not been altered in any way. We have altered the formatting in order to facilitate the ease of printing and reading of the dissertation. All rights not granted by the above license are retained by the author. Abstract: Abstract of thesis entitled Fast Evoked Potential Estimation by Artificial Neural Networks Submitted by Fung Shun Ming For the degree of Doctor of Philosophy at the University of Hong Kong in January 1999 Clinical measurements of the electrical signal in human body are significant issues in the study of health and disease. Evoked potential (EP) is the gross electrical potential generated by the central nervous system (CNS) in response to sensory stimulation. EP measurement is one important non-invasive technique of probing the brain and it has many potential applications in studying and monitoring the main functions of the system. Changes in EP waveform could well be indicative of disorders in the CNS. Since the EPs are severely corrupted by the background spontaneous activities of the brain - electroencephalogram QEEG), it is difficult to detect the EP. In many clinical situations, like operation room or intensive care unit, tune is critical. It is essential to develop fast EP estimation algorithms that could reduce the measurement time and track the changes across trials. In this thesis, artificial neural networks (ANNs) were studied for estimating EP. ANNs are powerful computational architecture having networked processing nodes, which could be non-linear hi nature, for solving many complex and mathematically ill defined problems like non-linear problems. EPs are fed via a tapped delay line to the input nodes of the multilayer perceptron (MLP) having a non-linear hidden layer and a linear output node. Using experimental VEP, the MLP was trained by back-error- propagation training. After training, the MLP achieves the SNR enhancement characteristic and performs as an artificial neural network filter (ANNF). The results of simulation experiments showed that the ANNF could boost the SNR of observed visualevoked potential (VEP) as much as 10 dB. Using the ANNF, the major VEP peak component, PI00, is readily detected in a single-trial. Therefore, peak latency and peak amplitude could be determined. Applying the concept of latency corrected average (LCA), VEP is estimated by aligning the enhanced trials at PI00 and then average. The results showed that, compared to EA, a clearer and sharper PI00 was obtained by LCA with the same ensembles of data. An adaptive radial-basis-function neural network (RBFNN) filter is then developed for EP tracking. A RBFNN was specifically designed for modelling EP. Taking the time-locked property of EP, the underlying EP was adaptively tracked without the need of reference signal. Its performance in SNR enhancement was theoretically investigated and verified by simulation experiments. The technique was applied to track VEP, brainstem auditory evoked potential (BAEP), and somatosensory evoked potential (SEP). It was found that the technique could greatly reduce the measurement time, significantly improve the SNR, and track the change across trials. Comparative studies showed that the technique outperforms moving-window average, exponentially weighted average, and adaptive filter (AF) in terms of tracking speed and accuracy. Further improvements are described to reduce the network size, shorten the duration of initial convergence and pre-enhance EP for AF. Fast computation of the RBFNN was formulated to reduce the computing load to sixth of AF. Using the method EP can be tracked across trials, which can provide potentially useful information for clinicians and physiol