Abstract :
[en] Introduction
Parkinson's disease (PD) is the fastest-growing neurodegenerative disease [1]. As PD affects mobility, it significantly impacts the patient's quality of life (QoL) [2]. Although the association between QoL and mobility has been established, the relevance of sensor-derived gait features to model self-reported mobility is yet to be defined [3].
Research Question
This study explores the predictive value of combining gait features from different test paradigms with demographic and clinical variables to model the 39-item Parkinson's Disease Questionnaire (PDQ39) mobility sub-score.
Methods
People with PD (PwPD) were enrolled from the Luxembourg Parkinson’s Study established within the National Centre of Excellence in Research on Parkinson's Disease [4, 5]. Participants performed a standard Time-Up and Go (TUG) and a dual-task motor test (TUG with tray) with gait features derived from a medical-grade IMU sensor attached to the patient's shoes.
Analysis included a nonparametric Wilcoxon signed-rank test to assess significant differences across gait tests, a Kruskal–Wallis test for group-based differences, and bivariate correlation analysis through Spearman's rank to assess the significance of all variables and mobility. The association between significant correlations was assessed via multiple linear regression, with optimal features selected through repeated sub-setting of a global model.
Three models were implemented to predict PDQ39’s mobility sub-score: a base model incorporating solely demographic and clinical variables, an extended model integrating TUG gait features, and a complex model with dual-task features.
Results
Data from 164 PwPD was analyzed (111 men, age: 66.4 ± 9.8 y, PDQ39 Mobility: 8.2 ± 8.1, MoCA: 24.67 ± 3.8, UPDRS III: 35.3 ± 13.8). Gait features significantly differed across TUG and dual-task tests (p < 0.05), except for Turn Angle and Heel Strike Degree. Bivariate correlations between mobility and clinical/demographic variables were constant across models, with medication, depression, and cognition being significant. Additionally, significant gait correlations with mobility were seen for gait speed and maximum lateral excursion for the extended model, and maximum lateral excursion for the complex model.
The best-performing linear model was the extended model predicting mobility, with a test set R² of 0.42 (p < 0.05) and a root mean square error (RMSE) of 5.93 points. The base (R² = 0.25, RMSE = 6.49) and complex (R² = 0.24, RMSE = 6.63) models performed similarly on the same test set.
Discussion
These findings highlight the relevance of gait parameters in modeling mobility-related QoL in PD and suggest their potential to outperform demographic and clinical metrics alone, thereby confirming their clinical utility. Furthermore, results suggest that complex gait tests may be unnecessary, as they do not outperform the predictive power of a standard gait test.
Further studies are needed to compare our results to alternative tests, like dual-task cognitive tasks, model non-linear combinations, and power for sub-group comparisons.
