ARTIFICIAL INTELLIGENCE-BASED TREMOR QUANTIFICATION IN CLINICAL VIDEOS OF PARKINSON'S DISEASE PATIENTS

Abstract

Over 10 million people around the world suffer from Parkinson¿s disease, a progressive neurodegenerative disorder. Its main symptoms are slowness of movement, stiffness, and tremors. For its diagnosis and to check its progression, the MDS-UPDRS scale is commonly used, which is a standardized rating system that is based on clinicians¿ visual observations. However, the inherent subjectivity of this method results in inconsistent scoring, which affects treatment choices and patient care. To address this, researchers are investigating more objective ways of quantifying tremors using various tools like computer vision and wearable sensors. Sensor systems can, however, be expensive, invasive, and not practical for regular use. Video-based methods, on the other hand, are gaining attention because they are accessible and non-invasive. Despite this, most conducted studies have either focused on highly controlled environments or on different items of the MDS-UPDRS, like finger-tapping tests, leaving postural tremors in real world settings mostly unexplored. This thesis aims to bridge that gap by offering a low-cost, non-invasive way to evaluate postural tremors in Parkinson¿s disease with video recordings taken in natural environments with the least constraints possible. Patients were filmed performing the MDS-UPDRS postural tremor test using just one camera, and the videos were then segmented into frames. Keypoints corresponding to the positions of the fingers and hands were manually labeled. By using these labeled points, a deep learning YOLO model was trained to detect fingertips. From the resulting finger trajectories, tremor-related data was obtained, including displacement distance, frequency, and peak shapes. These features were then used to train a binary classifier to predict Parkinson¿s Disease presence. The classifier achieved an F1 score of 0.93 and an AUC score of 0.962 on the test set, with a precision for the PD class of 100%, and recall at 80%, missing one case. Model training and detection were successful in many cases. However, the classification results were restricted by the small dataset size and the lack of publicly available data for this specific hand position and angle. The model also faced issues with imbalanced data, meaning it found it hard to perform certain classifications. As a result, even though the methodology proved feasible, the current findings are not yet conclusive. In conclusion, AI-based video analysis shows strong potential for objective postural tremor evaluation in Parkinson¿s disease. To guarantee clinical applicability, future work should focus on expanding the dataset, enhancing the robustness of keypoint detection, and validating the approach in a larger population.