Unsupervised and scalable low train pathology detection system based on neural networks

Abstract

Currently, there exist different technologies applied in the world of medicine dedicated to the detection of health problems such as cancer, heart diseases, etc. However, these technologies are not applied to the detection of lower body pathologies. In this article, a Neural Network (NN)-based system capable of classifying pathologies of the lower train by the way of walking in a non-controlled scenario, with the ability to add new users without retraining the system is presented. All the signals are filtered and processed in order to extract the Gait Cycles (GCs), and those cycles are used as input for the NN. To optimize the network a random search optimization process has been performed. To test the system a database with 51 users and 3 visits per user has been collected. After some improvements, the algorithm can correctly classify the 92% of the cases with 60% of training data. This algorithm is a first approach of creating a system to make a first stage pathology detection without the requirement to move to a specific place.

Keywords

• Pathology detection
• Recurrent neural network
• Gait analysis
• Biomechanics
• Signal processing
• Pattern recognition