Design and test of neural networks and statistical classifiers in computer-aided movement analysis: a case study on gait analysis

Objective. To describe the methods for designing and testing diagnostic systems in movement analysis and to verify the clinical usefulness of neural networks and statistical classifiers in a case study. Design. Connectionist and statistical models trained and tested with measured data. Background. A basic need in rehabilitation and related fields is to efficiently manage the vast information obtained from a movement analysis laboratory. Many studies have dealt with the interpretation of measured variables in order to correlate objective descriptors to the presence and/or severity of specific neuromusculoskeletal disorders or their consequences. This traditional analytical approach has been complemented in the last decade by new non-linear classification tools called neural networks. Methods. A gait analysis study on 148 lower limb arthrosis patients and 88 age-matched control subjects. Pathological and healthy gait patterns obtained from force plates were discriminated by means of multilayer perceptrons and statistical classifiers. Results. Ten input features were enough to train a multilayer perceptron with six hidden neurons. The discrimination rate of the neural net was 80% after cross-validation, significantly higher (P < 0.05) than the performance of a Bayes quadratic classifier (about 75%). A great variance due to a small cross-validation set could be demonstrated. Conclusions. Strict statistical requirements must be observed for designing a neural network. Although these models attain a better performance than conventional statistical approaches, the benefits they bring are sometimes not sufficient to justify their use. Further-more, clinicians routinely involved in critical decisions may not consider such diagnostic systems reliable enough.