[The use of artificial neural networks in evaluation of posturographic data]

Posturography methods have been applied in clinical neurootology to evaluate the equilibrium function of patients. Methods of statistical analysis play an important role for improving data processing and to support the interpretation of the results. In contrast to conventional statistics, artificial neural networks are model-free and non-parametric. The aim of the presented study was to investigate how accurately these methods are able to discriminate between healthy and equilibrium-disturbed subjects.

51 healthy volunteers participated in this study. 2 static posturography measurements were recorded before and 40 min after alcohol intake (0.4‰-0.6‰). Recorded signals were processed by 4 different methods in order to estimate power spectral densities (0 Hz-25 Hz). 11 different methods of artificial neural networks were investigated. The ability of artificial neural networks for classification was evaluated in patients with an acute unilateral vestibular loss.

It turned out that estimating power spectral densities by means of autoregressive modelling and subsequent classification by Support-Vector Machine or by Learning Vector Quantization Networks are most accurate. Validation analysis yielded mean classification errors for the test set of 4.2 ± 2.2%.

Analysis of neurootological data by artificial neural networks proved to be a sensitive recognition method of even small changes of the postural system.