Description:

Objectives: Prediction of disease progression in PD patients holds significant value. Individual knowledge of disease progression can help to make proper social and occupational decisions in association with future physical functioning of recently diagnosed patients. The analysis proposed in this work can help detect variables relevant to identification of disease progression and thus guide the design and interpretation of clinical trials involving neuroprotective and symptomatic therapy [58]. In this study, we investigate the optimal use of different hybrid machine learning (HMLS) and deep learning techniques for prediction of cognitive decline in year 4 from screening datasets. Montreal cognitive assessment (MoCA) is introduced to be a good screening test for cognitive function in PD. There are number of benefits for using MoCA that consists of it 1) being a quick measure of global cognitive function with short administration time compared with other tests, 2) covering a wide range of function in the cognitive domain, 3) having sensitivity to milder cognitive deficits in PD, 4) capturing executive dysfunction, 5) being widely adapted and utilized both clinically and in clinical research, 6) being widely adopted in many disease states, and 7) having alternate variants proposed for multiple languages and a more specific subtest weighting for PD. In this effort, we aim to investigate the optimal use of different hybrid machine learning (HMLS) and deep learning techniques for prediction of cognitive decline (MoCA) in year 4 from screening datasets. Methods: In our study, we selected over 650 PD subjects who had DaT SPECT images and clinical data from the Parkinson’s Progressive Marker Initiative (PPMI) database. First, DaT SPECT images are enhanced, normalized and cropped. Subsequently, we aim at applying deep learning algorithm (Autoencoder) to extract deep features from the preprocessed images. After that, deep features are preprocessed via some filters such as standard deviation and correlation filters. The reduced features are applied to HMLS including dimension reduction algorithms linked with classifiers to enhance prediction performance. In the end, we apply raw images to deep learning algorithms such as convolutional neural networks algorithms to make a comparison with above results. In this study, we employ 80% out of patients to train learner algorithms as 5-fold cross-validation and select the best modes. Furter, remaining 20% are employed to validate selected model.