Objective: To find hidden gait patterns associated with freezing of gait (FoG) in Parkinson’s disease (PD) by using machine learning techniques on the spatial and temporal parameters of gait analysis.

Background: FoG is a common and debilitating phenomenon in patients with PD. FoG refers to paroxysmal events, in which patients feel “their feet glued to the floor”¹. FoG is associated with an increased prevalence of falls, disability and loss of independence, adversely affecting patients’ quality of life (QoL)^2,3.

Method: We consecutively enrolled 41 patients with PD. All patients were evaluated with both the MDS-UPDRS and a gait analysis system. The presence of FoG was assessed using the item 13 of the MDS-UPDRS Part 2. Gait was investigated in the following conditions: 1) normal gait; 2) motor dual task; and 3) cognitive dual task. After applying the Smote technique⁴ to balance the dataset, four following supervised and tree-based algorithms were implemented on the spatio-temporal gait parameters in order to predict the presence of FoG.

Results: Based on the item 13 of the MDS-UPDRS Part 2, 10 out of the 41 patients reported to experience FoG, so these were classified as freezers, whereas the remaining 31 patients were classified as non-freezers. The analysis was performed per each gait condition with all features. For a fourth analysis nineteen gait features were included after a selection through the computation of a matrix and the choice of a threshold as regards correlation. Evaluation metrics were computed using a leave one out cross-validation. The evaluation metrics of accuracy, precision, recall, sensitivity, specificity and Area Under the Receiver Characteristics Curve were computed. All the algorithms obtained an accuracy greater than 80%, with one oh them, namely the Gradient Boosted Tree, achieving the best accuracy (93.5%) and specificity (93.5%) in predicting the presence of FoG.

Conclusion: The present pilot study shows that a machine learning approach on gait parameters could differentiate PD patients with ad without FoG. Importantly, in our patients FoG was only amnestic and never detected during gait analysis acquisitions. This enrollment choice suggests that a machine learning approach can help to identify patients affected by mild form of FoG that, even though less clinically significant at this stage, however exposes those patients at a major risk of developing more severe FoG.

References: 1. Giladi N. et al. Freezing of gait in PD: prospective assessment in the DATATOP cohort. Neurology. 2001 Jun 26;56(12):1712-21 2. Rahman S. et al. Quality of life in Parkinson’s disease: the relative importance of the symptoms. Mov Disord 2008;23(10):1428-34. 3. Amboni M. et al. Prevalence and associated features of self-reported freezing of gait in Parkinson disease: The DEEP FOG study. Parkinsonism Relat Disord. 2015 Jun;21(6):644-9. 4. Chawla N. et al. SMOTE: Synthetic Minority Over-sampling Technique. JAIR. 2002;16: 321-357.

« Back to MDS Virtual Congress 2020

MDS Abstracts - https://www.mdsabstracts.org/abstract/using-machine-learning-methods-to-explore-gait-features-associated-with-freezing-of-gait-in-parkinsons-disease/