AUTOMATIC RECOGNITION OF FAST AND SLOW PHASES IN NYSTAGMIC EYE-MOVEMENTS USING A NEURAL-NETWORK

This paper presents an algorithm for the automatic classification of f ast and slow phases during nystagmic eye movements, using a neural net work. When a patient is presented with a nonstationary visual field, t he resulting eye movements may be used to determine valuable clinical information about patients with vertigo and balance disorders. When th e nonstationary visual field is induced by sinusoidally rotating the p atient in a chair, the eye movements-collectively referred to as nysta gmus-typically consist of short, high-velocity movements (fast phases) which are in the direction of the stimulus and longer, low-velocity m ovements (slow phases) which are in the direction opposite to that of the stimulus. The slow phases are produced to compensate for the movin g visual field. By extrapolating over the fast-phase segments, the slo w-phase segments can be pieced together to form a slow-phase response. When the stimulus is sinusoidal, the slow-phase response is also sinu soidal and the magnitude and phase relationships between the stimulus and response may be used to help identify the source of the patient's disorder. Thus, the ability to accurately reconstruct the response fro m the slow-phase segments is extremely important. This, in turn, neces sitates the ability to accurately determine the locations of the fast and slow phases of nystagmus. For the neural network used here, the op timal input feature set and number of hidden units are determined, alo ng with the necessary preprocessing of the network inputs and the post processing of the network output data. It is also shown that an effect ive error-correction algorithm can be applied to the outputs of the ne ural network to improve its classification ability. Finally, results a re presented for the performance of the network on independent sets of test data. The classifications obtained from the neural network when applied to the test data are much more accurate than those obtained us ing two current classifiers: an algorithm proposed by Wall and Black a nd an algorithm proposed by Jell, Turnipseed and Guedry.