| Popis: |
The purpose of the current study was to classify people with autism spectrum disorder (ASD) using resting state functional magnetic resonance imaging data. Toward this direction, data were retrieved from the Autism Brain Imaging Data Exchange initiative, based on the Harvard–Oxford parcellation scheme, focusing on the default mode network, which has been previously reported to be related with ASD. The extracted time series were used to calculate plethora of functional features quantifying within-network connections, which were set as input to a medical decision support system (DSS). These interactions were evaluated based on a broad variability of methods, such as static functional connectivity (sFC) and dynamic functional connectivity (dFC) analysis, information-based metrics, and adjusted Haralick texture features. Finally, after extensive trials, head motion parameters, age, sex, and information regarding the acquisition protocol were found to improve the overall performance of the DSS. Internal parameters of the DSS were chosen based on a Bayesian optimization framework, which aimed to maximize the area under curve. The DSS comprised of a common support vector machine classifier featuring several kernels such as linear, polynomial, and radial basis function (RBF). In this study, several biomarker combinations were made to construct a suitable feature vector, in terms of higher classification performance. The best results were obtained using the combination of sFC and dFC, head motion parameters, handedness, sex, and information concerning the hardware setup. Using the RBF kernel the above combination resulted in accuracy 69.77%, sensitivity 77.08%, and specificity 60.51%. Through the current study, it is shown that it is feasible to achieve high classification performance despite the majority of acquisition parameters and different demographics or other information present in the data, by setting them as features to the DSS. |
