Work

Automated Assessment of Parkinson's Disease from Gait Parameters

Year: 2025
Type: dissertation
Abstract: Parkison's disease (PD) is a neurodegenrative disorder characterized by a set of motor symptoms, including bradykinesia, rigidity, tremor, and postural instability, which arise due to degeneration of dopaminergic neurons in the substantia nigra of the brain.These motor manifestations significantly impact daily life activities, often impairing gait and balance.Given the growing need for tools to automatically detect PD from gait-related data, the aim of this study was to apply machine learning to spatiotemporal gait parameters from a clinical cohort in order to distinguish PD patients from healthy controls.Gait data were collected for four walking conditions, which include for four walking conditions: Free Walking (FrW), Fast Walking (FaW), Dual-Task Walking (DTW), and Obstacle Walking (OBW).Various classifiers were trained on these data, including support vector machines (SVM), k-nearest neighbors (k-NN), random forests (RF), decision trees (DT), naïve Bayes (NB), logistic regression (LR), multi-layer perceptrons (MLP), and one-dimensional convolutional neural networks (1D CNN), based on their common use in the literature for PD detection through gait data.Two experimental setups were tested.In the first one, the full dataset for each task was used, while in the second, only a subset of common patients across all tasks were comprised.The dataset with the intersection of common patients allowed evaluating different task combinations and three fusion strategies: (1) early fusion, (2) intermediate fusion, and (3) late fusion.The results of the first experiment revealed that the highest AUC across all models and tasks was achieved with the OBW task using RF, with an AUC of 0.963, sensitivity of 1.00, and specificity of 0.917.In addition, FrW achieved an AUC of 0.951 with 1D CNN, DTW achieved 0.924 with MLP, and FaW demonstrated lower performance with an AUC of 0.778.The second experiment showed that classification performance varied significantly by task, with OBW providing the most consistent and accurate PD identification.Late fusion of all tasks resulted in the highest performance, reaching AUC = 1.000 for several models.Interestingly, certain single-task models, such as FaW with RF, also attained high AUCs, highlighting the value of individual tasks.Given the small test cohort, the results suggest that future research should extend this multi-task approach to larger, more diverse cohorts and explore the integration of additional data modalities to further improve performance and clinical applicability.
Author: Constança Inês Gomes da Silva
Language: English
Open Access status: green
OpenAlex ID: W7112698779
Cites: 216
Cited by:
Related to: 10