The purpose of this study is to identify novel phenotypes of freezing of gait using a deep neural network model and whole-body kinematics data.

Freezing of gait (FOG) in parkinsonism contributes to significant morbidity and is a challenge to treat due to its complex representation. Prior studies explored objectively phenotyping FOG primarily by spectral analyses on kinematics during gait assessments, which lack details to capture complex whole-body movements. In this work, we demonstrate the development of a deep learning model that analyzes complex body movements from whole-body 3D kinematics during time-up-and-go tests (TUG) for individuals with parkinsonism to detect FOG phenotypes. 

We collected whole-body kinematic marker time series from 2015 to 2017 using a motion capture system from 57 patients, including 5 patients with atypical parkinsonism, assessed with TUG tests in the off and on medication state. FOG score was measured with Movement Disorder Society-Unified Parkinson’s Disease Rating Scale Part III motor score (MDS-UPDRS-III). In our experiment, we designed a deep neural network model to simultaneously predict levodopa medication state (ON/OFF), FOG score (0-4), and MDS-UPDRS-III total score from the collected kinematics data. 

As compared to formal clinical assessments by a movement disorders specialist, our model classified levodopa medication state and FOG score with 96.4% and 96.2% F1 scores respectively and regressed MDS-UPDRS-III total score with root mean square error (RMSE) of 2.7 points.

This model detects time segments having characteristic movements of FOG during walking, e.g. small shuffling steps, akinesia, and tremulousness. Additional findings demonstrate that FOG is not limited to the lower extremity, and also significantly involves movements in the upper body, further supporting that FOG requires phenotyping using whole-body kinematics. Findings from our analysis may lead to novel hypotheses to define more granular FOG phenotypes.