Dalarna University's logo and link to the university's website

Neurodegenerative disorders such as dementia and Parkinson's disease (PD) affect millions of individuals globally and are characterized by progressive cognitive decline and motor impairments. As life expectancy and the number of older people increases, the number of people with these disorders is expected to increase. Currently, neurodegenerative disorders have no cure, making early diagnosis crucial for effective management and timely intervention. Gait analysis offers a non-invasive, inexpensive, and useful method for neurodegenerative disorders detection. Gait abnormalities, particularly under dual-task (dt) conditions, are early cognitive and motor decline indicators.

This thesis aims to investigate the potential of movement analysis for the discrimination of neurodegenerative disorders compared to healthy control (HCs) persons, with a specific focus on dementia and PD. By employing machine learning techniques, the research evaluates the effectiveness of these methods in distinguishing between HCs and those with dementia or PD. This thesis utilized various traditional machine learning and deep learning models applied to the movement data. The models implemented across the studies are Support Vector Machines (SVM), Logistic Regression (LR), Random Forest (RF), Decision Trees (DT), Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and a hybrid CNN-LSTM architecture. Movement features extracted from the datasets were applied to those models.

For dementia, utilizing video-based data obtained from the Uppsala-Dalarna Dementia and Gait project (UDDGait™), the study performed pose estimation using YOLOV8, followed by feature engineering. In the current study, movement features, including velocity, acceleration, number of steps, cadence, stride length, total time, and joint angles (knee and hip) were computed and used in the machine learning algorithms to differentiate the groups. The dataset comprised 64 individuals with dementia and 67 HCs. The participants performed the Time-Up-and-Go tests (TUG) under single task and dt paradigms. Following the UDDGait study protocol, the test performance was documented with two synchronized video cameras. In the dt conditions, participants completed the TUG test while simultaneously performing a verbal/cognitive task, which involved naming animals (TUGdt-NA) and reciting the months in reverse order (TUGdt-MB). For PD, gait features were extracted from a sensor-based dataset comprising 93 individuals with the disease and 73 HCs. The vertical ground reaction force (VGRF) was recorded for nearly two minutes using 16 sensors placed beneath each foot (8 per foot).

The results demonstrate that movement features extracted from video data, especially under dt conditions, are effective in distinguishing between HCs and those with dementia. The SVM algorithm achieved the highest accuracy of 88.5% and recall of 92.5% in dt animal naming (TUGdt-NA). For the PD study, the results demonstrate that RF obtained the highest accuracy and recall of 96%. The findings from these studies suggest that movement analysis using machine learning models offers a promising non-invasive, automated, and simple tool for the discrimination of dementia and PD compared to HCs. Future research could explore multimodal fusion approaches (i.e., speech and gait analysis) to enhance the accuracy and generalizability of these methods in clinical settings.