Clinical gait analysis by neural networks: issues and experiences

Clinical gait analysis is an area aiming at the provision of support for diagnoses and therapy considerations, the development of bio-feedback systems to train patients, and the recognition of eeects of multiple diseases and still active compensation. The data recorded with ground reaction force measurement platforms is a convenient starting point for gait analysis. We argue in favor of using the raw data from such force platforms and apply artiicial neural networks for gait malfunction identiication. In this paper we discuss our latest results in this line of research by using a supervised learning rule. The employed classiication approach is learning vector quantization which proved to be highly robust in the training process yielding a remarkably high recognition accuracy of gait patterns. Human gait analysis is particularly attractive to model because of its importance in everyday life and its complexity as a total body movement. In particular, pathological gait in humans following disease or injury is the subject of much contemporary reseach 10]. Gait analysis is traditionally the domain of dynamic electromyography, (ground reaction) force platforms, and joint kinematics and kinetics 2]. We perform a medical computing project based on ground reaction force measurements of patients' gait. The project aims at an assessment of gait after injury that is useful, on the one hand, as a support of diagnosis and therapy considerations. On the other hand, it is intended to give clues to a model of gait, to help developing bio-feedback systems to train patients, and to nd eeects of multiple diseases and still active compensation. In this paper we report on supervised identiication of gait malfunction with respect to the location of ailment. The location is provided by labelling the measured input patterns with the location of the injured body region as known from the patient's record. The employed classiication algorithm is learning vector quantization 6] as provided in the freely available LVQ PAK 7]. In our previous work we have demonstrated that employing a supervised neural network model (simple three layer MLP 1, 11]) for discriminating \healthy" and \pathological" gait patterns leads not only to a classiication scheme but also to a fairly good assessment of the gait patterns 4]. In 9] we classiied a number of \pathological" gait patterns by using one of the most prominent unsupervised learning strategies, self-organizing maps 8], yielding a topographic arrangement of the diverse patients' injuries, and thus a detailed classiication means. …