Classification of Parkinson’s Disease Patients—A Deep Learning Strategy

Abstract

(1) Background and objectives: Parkinson’s disease (PD) is one of the most prevalent neurodegenerative diseases whose typical symptoms include bradykinesia, abnormal gait and posture, shortened strides, and other movement disorders. In this study, we present a novel framework to evaluate PD gait patterns using state of the art deep learning algorithms. A comparative analysis with three different approaches is presented and evaluated upon three groups of subjects: PD patients, Young Healthy Controls (YHC), and Elderly Healthy Controls (EHC). (2) Methods: The three approaches used in the study include: (i) The energy content of the gait signals in the frequency domain is captured with spectrograms that are used to feed a CNN model, (ii) Temporal information is incorporated by creating GRU networks, (iii) Temporal and spectral information is simultaneously captured by creating a new architecture based on CNNs and GRUs. (3) Results: Accuracies of up to 83.7% and 92.7% are found for the classification between PD vs. EHC and PD vs. YHC, respectively. According to our observations, the proposed approach based on the combination of temporal and spectral information, yields better results than others reported in the state of the art. (4) Conclusions: The results obtained in this study suggest that the combination of temporal and spectral information is more accurate than individual approaches used to classify and evaluate gait patterns in PD patients. To the best of our knowledge, this is the first study in gait analysis where temporal and spectral information is combined in an architecture of deep learning.