A Pairwise Deep Ranking Model for Relative Assessment of Parkinson's Disease Patients from Gait Signals

Abstract

Continuous monitoring of the symptoms is crucial to improve the quality of life for patients with Parkinson's Disease (PD). Thus, it is necessary to objectively assess the PD symptoms. Since manual assessment is subjective and prone to misinterpretation, computer-aided methods that use sensory measurements have recently been used to make objective PD assessment. Current methods follow an absolute assessment strategy, where the symptoms are classified into known categories or quantified with exact values. These methods are usually difficult to generalize and considered to be unreliable in practice. In this paper, we formulate the PD assessment problem as a relative assessment of one patient compared to another. For this assessment, we propose a new approach to the comparative analysis of gait signals obtained via foot-worn sensors. We introduce a novel pairwise deep-ranking model that is fed by data from a pair of patients, where the data is obtained from multiple ground reaction force sensors. The proposed model, called Ranking by Siamese Recurrent Network with Attention, takes two multivariate time-series as inputs and produces a probability of the first signal having a higher continuous attribute than the second one. In ten-fold cross-validation, the accuracy of pairwise ranking predictions can reach up to 82% with an AUROC of 0.89. The model outperforms the previous methods for PD monitoring when run in the same experimental setup. To the best of our knowledge, this is the first study that attempts to relatively assess PD patients using a pairwise ranking measure on sensory data. The model can serve as a complementary model to computer-aided prognosis tools by monitoring the progress of the patient during the applied treatment.