Feature Selection of Input Variables for Diagnosis of Patellofemoral Pain Syndrome based on Random Forest and Multilayer Perceptron

Patellofemoral pain syndrome (PFPS) is a common knee disease in the clinic. Its etiology is various, involving a variety of biomechanical variables of lower limbs. Most of the traditional diagnostic methods are subjective and the diagnostic accuracy mainly depends on the experience of doctors. A machine learning method is proposed in this paper to objectively analyze the related variables of PFPS and classify it to assist doctors in diagnosis. The proposed method was tested on a running data set of forty-one subjects, which included seven surface electromyography (sEMG) and three joint angles. Firstly, the importance of ten biomechanical features related to PFPS was compared by the analysis of variance and mean combined with random forest (RF), and then the six most important features were selected. Finally, the 100-time sampling points of each feature selected were input into the multilayer perceptron (MLP) for classification. The classification accuracy is 75% with a 40% reduction of input variables, which is not much different from the 76% accuracy before feature selection. Compared with previous work, the proposed method explores the importance of features related to PFPS from a new perspective, which can assist doctors in the diagnosis of PFPS.