A Cluster Analysis for Parkinson's Disease Phenotyping with Gait Parameters

Gait impairment and postural instability can lead to dangerous conditions for Parkinson's disease (PD) patients. Gait Analysis combined with current machine learning (ML) techniques may help the clinicians to improve the prediction of an outcome or response to rehabilitation treatments. This study aims to define whether a dataset of gait parameters acquired in patients with idiopathic PD can be used to identify homogeneous groups separated from each other and corresponding to different PD phenotypes. An optoelectronic motion analysis system was used to obtain spatial-temporal parameters of PD patients during a single walking task. An unsupervised ML technique, namely clustering, was employed on gait parameters extracted from gait analysis to find different motor-phenotypes in PD patients. In particular, the
$\boldsymbol{k}-\mathbf{means}$
clustering individuated two groups (Cluster 1 and Custer 2) with a specific gait-pattern. Cluster 2 was characterized by an increase of double support phase, stance phase and stance duration and a decrease of velocity, cadence, step and mean cycle length. These findings suggest that abnormalities in gait parameters may provide a data-driven PD phenotyping, which identify a worse motor and non-motor PD phenotype.