Detecting load conditions in human walking using expectation maximization and neural networks

Weight carrying conditions in human subjects is classified using a mixture of Gaussian based classifier and neural networks. Expectation maximization (EM) algorithm is used to learn the model parameters for the mixture of Gaussian (MOG) based classifier. Since variables change in time as the human subjects walk, and likelihood score in a mixture model is usually calculated for stationary data, a scoring system is developed to calculate how likely a time sequence of variables belongs to a particular distribution. A neural network (NN) based classifier is also developed which uses the traditional back-propagation algorithm for training. The results obtained show overall 74.1% accuracy using MOG and 66.4% using NN for the test set in a binary classification task of detecting “load” or “no-load” conditions using just two variables. The lower accuracy using NN is not surprising as averaged variables are provided as inputs to the NN while MOG is able to use the dynamic information. In another classification task with four classes and using two variables, the accuracy was 37.2% using MOG and 34.2% using NN on the test set which are both better than chance. Accuracy of NN using 7 variables was 81.4% for binary classification and 41.3% for four-class classification. An interesting finding from these results is that NN (using 7 averaged variables) performed better than human perceivers who were asked to judge based on stick figure animations of subjects walking. Another interesting revelation from this study was that the covariance matrices from the MOG model revealed “anti-phase” locking of the elbow and stoop angle as the subjects walk.