Comparação de dois métodos de classificação na análise do padrão dinâmico da marcha

Abstract

Gait can be described mechanically by the time series of a set of biomechanical variables, e.g. the time series of the ground reaction forces. These time series can be analyzed by their correlations with orthogonal functions, serving as criteria for comparison of different movements. Quantitative analysis of human gait has been used by physicians and physiotherapists for diagnostics of pathologic gait pattern and identification of gait pattern changes due to therapeutic and orthopedic interventions. In order to classify and distinguish different groups (normal gait versus pathologic gait), non-linear classification methods (Artificial Neural Networks) have been applied and the results compared to linear statistic classificators. Therefore, the objective of this study was to verify the capacity of two classification methods, Artificial Neural Networks and Least Squares Method, to distinguish two different situations, barefoot gait and gait with shoe wear. Twenty-four male and female, individuals, averaging 23.2 years of age, walked on a force platform at the most economic velocity of 1.3 m s'^ and selfdetermined velocity, 40 trials barefoot and 40 times with their own shoe wear. The velocities were monitored by two photo-cells. The three components of the ground reaction force were registered with a frequency of 1 kHz. For each group of trials (barefoot and with shoe ware) the optimal number of orthogonal functions was calculated. The next step was the presentation of the orthogonal functions coefficients as a standard entry for the Artificial Neural Networks and the Least Square Method in order to identify the two distinct situations (barefoot gait and gait with shoe ware). The results showed that both methods obtained very high recognition rate. For the vertical component of the ground reaction force the recognition rate of the validation process was 99.27% for Least Square Method and 99.65% for the Artificial Neural Networks. The recognition rate for the anterior-posterior component was 94.06% for the Least Square Method and 96.06% for the Artificial Neural Networks. For the medium-lateral component the recognition rate was a bit lower, 87,46% for the Least Square Method and 93,62% for the Artificial Neural Networks. Based on these results it can be concluded that it might be possible to apply these methods also for other classification problems of human gait and that the Least Square Method, which is much easier to be applied, provides recognition rate close to the ones from the Artificial Neural Networks.