| Schizophrenia(SCZ)is a disease whose cause is still unknown.It affects about 1%of the world population.Doctors’ diagnosis of schizophrenia is mainly based on the assessment of the patient’s course of disease,mental state,and the diagnosis results,which will be subject to subjective influences of doctors to a certain extent.Currently,more and more studies are devoted to searching for effective biomarkers to distinguish patients with schizophrenia,but no unified conclusion has been reached.This is partly because our understanding of the brains of schizophrenics is not clear,and the source of the study’s participants varied.With the robust development of neuroimaging,imaging data can be used to improve the understanding of schizophrenia.There have been many results showing that the brain of schizophrenics has structural and functional differences compared with that of healthy people.Such differences can be detected by MRI(Magnetic Resonance Imaging)technology.Meanwhile,Electroen-cephalography(EEG),another non-invasive technique for studying the brain,has also shown differences in brain waves in certain frequency bands of the cerebral cortex.Therefore,this study aims to combine MRI and EEG techniques to search for reliable neural markers based on data-driven,as well as to predict the mental state of patients with schizophrenia.This study mainly includes the following two parts:1.In this study,45 normal subjects and 49 schizophrenic patients were included,and the classification model of these two groups was constructed.The data collected included rs-f MRI(Resting-state f MRI),T1-weighted data,DTI(Diffusion Tensor Imaging)data and EEG data.Fraction Amplitude of Low Frequency Fluctuations,Regional Homogeneity and Degree Centrality were calculated based on resting state data;white matter volume,gray matter volume and cortical thickness were calculated based on T1-weighted data;Fraction Anisotropy value,Mean Diffusivity value and structural connection matrix were calculated based on DTI data;and power spectrum under five different frequency bands were calculated based on EEG data.Based on the above ten indicators,Logistic Regression(LR)and Support Vector Classification(SVC)were respectively used to construct the single mode Classification model,and most of the Classification effects of the two model test sets reached 80%.And the classification effect of SVC is generally 2～5 percentage points higher than that of LR classifier.This research took SVC classification model as the learning machine and adopted Bagging algorithm to carry out the next step of multi-mode fusion.Through the ensemble learning algorithm,the classification accuracy of the model reached 95% and the generalization ability of the model was improved at the same time.2.For the SCZ group,this research continued to explore the predictive ability of different feature spaces on Positive and Negative Syndrome Scale(PANSS).Also used ten kinds of feature space like single mode classification,using Linear regression,Bayesian Ridge regression,SVR regression and Ridge regression algorithm to build regression model,the positive symptoms rating scale,negative symptoms rating scale,general psychopathology scale score and total scale score,respectively made a forecast.The results showed that the minimum mean absolute error of the SVR model was generally 0～2 less than the MAE value of the other regressors.Therefore,this reasarch chose SVR algorithm as the base learning machine of our Bagging integrated algorithm.It was found that the prediction error was reduced for the positive symptom scale and the total score of the scale.To sum up,in this paper,in the process of constructing classification model and regression model are found that the effect of a variety of modal data fusion is better than the single mode condition.Important features in the process of model construction can be used as effective neural markers,which can provide an auxiliary means for clinical diagnosis and prediction of schizophrenia patients. |
PDF Full Text Request