Graph theoretical metrics and machine learning for diagnosis of Parkinson's disease using rs-fMRI

In this study, we investigated the suitability of graph theoretical analysis for automatic diagnosis of Parkinson's disease. Resting state fMRI data from 18 healthy controls and 19 patients were used in the study. After data preprocessing and identifying 90 regions of interest using the AAL atlas, average time series of each region was obtained. Next, a brain network graph was constructed using the regions as nodes and the Pearson correlation between their average time series as edge weights. A percentage of edges with the highest magnitude were kept and the rest were omitted from the graph using a thresholding method ranging from 10% to 30% with 2% increments. Global graph theoretical metrics for integration (Characteristic path length and Efficiency), segregation (Clustering Coefficient and Transitivity) and small-worldness were extracted for each subject and their between group differences were subjected to statistical analysis. Local metrics, including integration, segregation, centrality (betweenness, z-score, and participation coefficient) and nodal degree, were also extracted for each subject and used as features to train a support vector machine classifier. We have shown a statistically significant increase in characteristic path length as well as a decrease in segregation metrics and efficiency in Parkinson's patients. A floating forward automatic feature selection method was used to select the 5 best features from all extracted metrics to classify patients. Our classifier was able to achieve a diagnosis accuracy of ~95% when subjected to a leave-one-out cross-validation test. These features belonged to cuneus (right hemisphere), precuneus (left), superior (right) and middle (both) frontal gyri which were all previously reported to undergo alterations in Parkinson's disease. This investigation confirmed that global brain network alterations are associated with Parkinson's patients' symptoms and showed the potency of using graph theoretical metrics and machine learning for diagnosing the disease.