Parkinson's Disease (PD) is a progressive neurodegenerative disorder characterised by both motor impairment and non-motor symptoms, including cognitive impairment. PD presents significant challenges for reliable diagnosis and accurate symptom assessment. The current "gold standard" clinical assessments rely on visual judgement, introducing subjectivity. This thesis aims to mitigate these limitations by applying objective machine learning methodologies to two distinct types of movement data, simple hand motor tasks and neuropsychological graphmotor assessments, with the objective of modeling motor severity and a potential for a more granular approach for assessing cognitive impairment for people with PD.
For the hand motor tasks, end-to-end time series classification models were used to analyse positional data collected from 47 healthy controls and 148 PD patients. These models were applied for the diagnosis of PD and for the detection of clinically slight bradykinesia. After employing a 5-fold nested cross-validation strategy, the top-performing models achieved an accuracy rate of 84% for PD diagnosis and 82% for bradykinesia detection. These models provide an agile, objective, and rapid framework for hand kinematic assessments, negating the need for domain-specific knowledge. They have the potential to serve as essential tools for preliminary research in the field of kinematic evaluations.
For the drawing assessments, the structural components of the Benson Complex Figure were identified with a top accuracy rate of 96%, following the novel investigation of encoding pen-dynamics. This enables the extraction of cognitive features related to the organisational strategy employed by the subjects.
Collectively, these findings introduce promising new data-driven approaches for the modeling of PD diagnosis and cognitive states. Importantly, the research is designed with the aim of integrating these methodologies into routine clinical practice and aligning with current research interests, thus laying the groundwork for future domain-specific studies in PD assessment.
