PhD Thesis
Dissertation Title: Computer-Aided Early Detection of Parkinson’s Disease (PD) through Multimodal Data Analysis
Abstract of the work:
Early detection of PD is a challenging and an important problem for early management of the disease and for better treatment regimens. There are no definitive diagnostic tests for PD and the clinical diagnosis relies on presence of classic symptoms, and by the time a patient shows these symptoms, more than 60% of dopaminergic neurons are lost. In this work, various analytics and predictive modeling using multimodal data has been carried out for the early detection of PD. The data used for the study are neuroimaging data (SPECT and MRI), clinical examination data and biological (cerebrospinal fluid) data. Image processing was carried out to process the images and extract the features, and along with clinical and biospecimen data were used to develop predictive models. Machine learning and statistical tests were used to analyse the features. Various techniques such as logistic regression, Naive Bayes, support vector machines, random forest were used in the analysis.
Highlights from this research are:
- 3 international journals and 3 conferences
- Among the winners of the Best Paper at the IEEE EMBC 2014 conference (A prestigious conference of the IEEE).
Summary of the work:
Early detection of Parkinson’s disease (PD) is an important and challenging problem for effective management and treatment. There are no definitive diagnostic tests for PD and the clinical diagnosis relies on presence of classic symptoms, and by the time a patient shows these symptoms, more than 60% of dopaminergic neurons are lost. In this thesis, approaches based on multimodal data analysis and machine learning for early detection of PD are presented.
Recent neuroimaging techniques, such as SPECT, have shown to be sensitive in discriminating PD from healthy normal even at the early stages of the disease. Striatal Binding Ratio (SBR) values are well-established and clinically used quantification measures obtained from SPECT. This work carries out a novel use of these features in developing classification and prediction models for discriminating early PD from healthy controls. A high accuracy is achieved in classification.
Along with discriminating PD from normal, SPECT imaging have also indicated their usefulness in detecting a group of subjects, called Scans Without Evidence of Dopaminergic Deficit (SWEDD) who show normal scans who otherwise were clinically misdiagnosed as PD. This work carries out analysis of SPECT images, for segmenting and quantifying the high activity regions that correspond to striatal uptake, for extracting discriminatory features that eventually are used for classifying degenerative parkinsonism such as PD from healthy normal and non-degenerative conditions such as SWEDD. These computed features show good variation in PD as compared to normal and SWEDD, and they are statistically significant as well. The classification accuracy is also observed to be very high.
Research studies show that there exist a phase called the premotor or the prodromal phase before the motor phase in PD and that this phase is mostly dominated by nonmotor symptoms such as smell loss and REM sleep behavior disorder. This work carries out a novel use of these features along with other relevant features such as imaging and cerebrospinal fluid measurements to classify early PD from healthy normal. A very high accuracy is obtained in classification using an SVM classifier. As of now, there exists no standard screening procedures for detecting PD, and patient questionnaire are easy tools that might have the potential for a screening instrument. This work carries out the novel use of using the patient questionnaire portion of the widely-used Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS) to classify early PD from healthy normal. We observe a very high accuracy in classification.
Therapeutic options in PD depend on the stage and severity of the disease. This work carries out estimation of stage and severity of PD using MDS-UPDRS which evaluates the most pertinent features in PD. The AdaBoost-based classifier ensemble gave a high accuracy in classification. We infer from the study that such models have the potential to aid clinicians in the diagnostic process.
A copy of my thesis is available here.