Adopting Machine Learning Technology for the Classification of Parkinson’s Disease
- Publication Type:
- Thesis
- Issue Date:
- 2020
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Parkinson’s Disease (PD) is the second most common neuro-degenerative disorder affecting approximately 1% of the population. Major symptoms include tremor, bradykinesia and freezing of gait. The precise diagnosis of PD remains a challenge for clinicians due to the similarity of PD symptoms with other disorders. Although diagnosis is based on clinical symptoms, PD is associated with a plethora of non-motor symptoms adding to its overall disability.
Research into early diagnosis of PD has taken advantage of machine learning-based image analysis. Neuroimaging modalities, like Single Photon Emission Computed Tomography (SPECT), have shown to aid in early diagnosis of PD. SPECT images are powerful tools that depict dopaminergic deficits in brain. Dopamine transporter (DAT) loss is a crucial feature necessary for PD identification. Regular scans are identified by intense and symmetric DAT binding that appear as two “comma-shaped” regions. Any asymmetry of this shape implies an abnormal finding. This thesis proposes three machine learning approaches for the classification of PD.
Firstly, we developed a neural network that classified PD patients from healthy controls. SPECT images were used to train our network. 10-fold cross-validation was used to evaluate our network’s performances. Experimental results indicated that our approach outperformed the benchmark studies.
To classify PD patients into different stages, PDStageNet was implemented which learned several features from the images and their associated stages. The affected regions were enhanced using image segmentation process. Experimental results showed that the proposed approach achieved a very high accuracy in classifying PD patients into five clinical stages of PD progression.
Finally, a classification model was utilised to streamline the process of identifying PD patients for surgical treatment using only clinical data. A feature selection process was used to identify the essential features and determine if the accuracy could be improved. Two experiments were carried out to test this hypothesis. In experiment 1, the best classifier for PD classification was identified. In experiment 2, feature selection process was implemented to determine the essential features. Experimental results indicated that, with only 60% of the features, the accuracy was higher than the benchmark studies.
The significance of these studies is that we propose effective machine learning-based approaches to diagnose PD at its earlier stages, so that the management and prognosis of PD patients can be significantly improved. Given the high performance of our approaches, we believe that the early diagnosis of PD can be done, which will revolutionize PD diagnosis and management.
Please use this identifier to cite or link to this item: