| Title | Neural Net for Auditory Evoked Potentials Analysis |
| Author(s) | R. N. Moreira; M. I. Fasolo; P. J. Abatti; J. C. Nievola; R. R. Seixax |
| Source | Proceedings of the World Congress on Medical Physics and Biomedical Engineering, Page 475 |
| Publication Date | 1994 |
| Abstract | This work presents auditory brain-stem responses' characteristics identification using backpropagation recurrent network [CAUDILL92]. The evaluation of an evoked potential consists of 4 steps: identifying the peaks (or troughs) in the waveform; measuring the amplitude and latency at these peaks; comparing these measurements to sets of normal values; and relating any abnormal measurements to patterns caused by various lesions [PICTON88]. The effective applying of these procedures needs an expert neurophysiologist that actually does the evaluation. This imposes some limits on this valuable diagnostic technique in medical medium. Classical procedural algorithms have been proposed to solve automatic identification in evoked potentials equipments. The main difficulty on these methods is that the human expert sometimes cannot explain the mental sequence of evidences that support his decision. The neural net coneccionist approach is interesting provided that systems built on this paradigm can "learn" to identify signal parameters by actual examples. The success of a neural net design is based on proper choose of net topology, learning algorithm and mainly on data base for training and validate. The system's data base has generated from actual paper records exams. Scanning process convert recorded image to a function that preserves original amplitude and time scales [MOREIRA93]. A set of discrete time series that becomes the "training set" to feed the net is attained on applying this procedure. How nervous system answer is time-locked the net used to analyze the evoked potential biological signal has one neuron for each time series element. Every signal in the training has a fix size of 10 ms and was sampled at 25 KHz that means 250 neurons in the input layer. |