Gait classification of parkinson’s disease patients using efficient representations from autoencoders

Abstract

ABSTRACT : Parkinson’s Disease (PD) is one of the most common neurodegenerative diseases. Patients manifest a progressive degeneration of dopamine, which plays a key role in abilities such as the locomotion, cognitive capabilities, sleep regulation and mood. One of the symptoms of the disease is the progressive gait impairment, resting tremors, slowness of movement, shuffling steps, among others. There is interest among the scientific community to develop automatic classification systems to support the diagnosis. The goal is to properly discriminate the disease and to predict the neurological state of the patients. This work focuses on the use of Convolutional Auto-Encoders to obtain efficient representations from multi-channel gait signals from Smartphones and sensors to classify PD patients vs. Healthy subjects. The channels represent the acceleration in the 3-dimensional plane (X, Y, Z). The proposed experiments consist of three models using 64, 128, and 256-dimensional bottlenecks to compress the information of gait signals. The accuracy and unweighted average recall are used to evaluate the classification performance over the PC-GITA database, from which 38 controls and 38 subjects were used for training the neural networks, and 30 patients and healthy subjects were used as test dataset. The subjects were asked to perform the 4x10 gait task, which consists of four repetitions of walking for 10 meters, stop and perform a 180° turn. A Stratified 5-Fold-Cross-Validation strategy is used to evaluate the performance of a Support Vector Machine over the testing dataset. The results indicate that the 64-dimensional bottlenecks provide enough information to properly differentiate between patients and controls. The results report accuracy of up to 85%, and Unweighted average recall values of 93%. Additionally, the area under the ROC curve is reported for each fold. There is no variation in the results when considering gait signals with non-randomized and randomized channels. It is concluded that the methodology is suitable to classify patients vs. healthy subjects, despite of the different origins from the signals and the challenges that different sampling frequencies impose for such a methodology.