| BackgroundSchizophrenia(SZ)is one of the most disabling neuropsychiatric conditions in the world and is associated with distributed neural abnormalities.It is well established that SZ is a disorder of distributed brain “dysconnectivity” emerging from complex biological alterations across multiple neural systems.Its symptoms lead to profound economic cost and lifelong disability for most patients.Therefore,characterizing disturbances of brain networks in SZ constitutes a critical research goal and requires the identification of both pathophysiological mechanisms and better neural markers to guide intervention and treatment.There is growing evidence that schizophrenia is associated with disrupted functional brain connectivity,involving interconnections between diverse brain regions.Functional studies have shown that functional connectivity within large-scale brain networks such as Default Mode Network(DMN),Executive Control Network(ECN),Visual Network,Auditory Network,Sensorimotor Network,Speech Network,Semantic Network and Salience Network are compromised in patients with schizophrenia.These studies improved our knowledge in functional impairment of brain networks in schizophrenia.However,it is not clear whether the structural integrity of these brain networks is altered in schizophrenia.To date however,these brain network results showed minimal clinical impact for diagnostic and prognostic purposes,traditional diagnostic and prognostic tools are still used by psychiatrists.The most important reason is that the differences between SZ and controls were reported at group level,which provided limited information to make inferences at the level of the individuals.With the continuous innovation of machine learning technology,pattern classification algorithms have become widely used in SZ research.Machine learning algorithms based on brain network data may be helpful in the diagnosis,prediction,classification and treatment of SZ.Schizophrenia is a devastating neurodevelopmental disorder with a complex genetic etiology.Multiplex family studies have established significant heritability for SZ,which is often summarized as 81%.First-degree relatives(FDRs)of patients with SZ are 10 times more likely to suffer from schizophrenia than healthy controls.Consequently,when evaluating an individual’s risk for schizophrenia,the familial risk is among the most important factors.Previous large-scale network analysis showed that FDRs of patients with SZ demonstrated similar deficits in connectivity metrics,interhemispheric functional connectivity abnormalities,default mode network dysfunction,and rich-club connectivity impairments as SZ.This sharing of disease-specific patterns indicates that brain network disturbances are likely to show familial associations,possibly reflecting a vulnerability for SZ.Consistent with these results,our previous study using stochastic dynamic causal modeling found similar anterior cingulate cortico-hippocampal dysconnectivity in unaffected FDRs and patients with SZ.With the help of magnetic resonance imaging and machine learning algorithms,we may be able to develop more sensitive predictive models to achieve early intervention for first-degree relatives with high-risk SZ.Medication treatment became available with the development of chlorpromazine in the1950 s,and antipsychotic medication development continues to this day.Unfortunately,not all patients respond to antipsychotic medications.Overall estimates suggest that one-fifth to one-half of patients have treatment resistant schizophrenia(TRS).TRS patients have poorer outcomes when compared to other patients with severe mental illnesses after treatment.They also have worse living quality.Furthermore,persistent positive,negative,and cognitive symptoms lead to worsened social functioning and long-term disability.Finally,TRS costs 3–11 fold more than schizophrenia patients in remission.Therefore,If the effect of antipsychotics on SZ patients can be predicted at the early stage of the disease using machine learning methods with network analyses,it will have good clinical significance and social benefits.ObjectiveAccording to the above research background,this thesis discusses three aspects from the pathogenesis,diagnosis to the prediction of curative effect in schizophrenia: first,based on large sample size of magnetic resonance imaging structural data,we discuss the changes of structural covariant network in patients with first-episode schizophrenia(FE-SZ),and with the confirmation that schizophrenia is a dysconnectivity condition from the brain structural network level;Second,based on a functional magnetic resonance imaging data,with the support of machine learning algorithms,we discuss whether the brain functional networks can be used to accurately classify schizophrenia patients and controls.If true,whether the classification model can be applied to the first-degree relatives of patients with schizophrenia to guide the early intervention;Thirdly,based on f MRI data and with the support of machine learning algorithms,we discuss whether the characteristics of brain networks can be used to predict the effect of antipsychotic drugs on patients with schizophrenia.Methods1)In this experiment,All FE-SZ patients were recruited form the department of psychiatry of Xijing Hospital.Age and gender matched health controls(HCs)were recruited through advertising from local communities.All FE-SZ patients were diagnosed by two senior clinical psychiatrists using the structured clinical interview for DSM-5.The severity of symptoms was assessed using the positive and negative syndrome scale.All MRI data were collected from the department of Radiology of Xijing Hospital.High resolution of T1 images of 107 FE-SZ patients and 92 HCs were obtained using 3.0-T Siemens Magnetom Trio Tim imaging unit and an eight-channel phased-array head coil,another high resolution of T1 images of 86 FE-SZ patients and 86 HCs were obtained using 3.0-T GE 750 imaging system.The MRI T1 data was preprocessed using FSL 5.0.9 software.General linear model was used to perform on the preprocessed gray matter data.We selected eight spherical regions with a radius of 4 mm in the gray matter image as the seed point region for this study.The eight regions of interest we selected are as follows: 1.Visual network(primary visual cortex,calcarine sulcus,9-81 7);2.Auditory network(primary auditory cortex,Heschl ’s gyrus,46-18 10);3.Sensorimotor network(primary motor cortex,precentral gyrus,28-16 66);4.Speech network(inferior frontal gyrus,pars opercularis,50 18 7);5.Semantic network(temporal pole,38 10-28);6.Salience network(fronto-insular cortex,3826-10);7.Execution control network(dorsal lateral prefrontal cortex,dl PFC,44 36 20)and8.Default mode network(angular gyrus,46-59 23).The specific coordinates are as above.We first calculate the average gray matter volume in the spherical region for each subject,and use these gray matter volumes as covariates of interest in the general linear model for independent correlation analysis.The statistical model has a column of ones,which is used to remove the mean effects,and then the age,gender,and MRI machine(1 for Siemens,2for GE)of each subject are included as covariates.First-episode schizophrenia patients and healthy subjects were modeled separately,so a total of 8 × 2 = 16 structural covariance network analyses were performed.The threshold for each group of statistics was specified at P <0.05,threshold-free cluster enhancement(TFCE)corrected.2)The study sample consisted of 40 first-episode SZ patients from early intervention services within the outpatient clinic and inpatient department of Xijing Hospital,36 FDRs of patients with SZ,and 40 HCs recruited from the local community by advertisements.Two senior clinical psychiatrists diagnosed schizophrenia using the structured clinical interview for DSM-5.All MRI data were collected on a 3.0-T Siemens Magnetom Trio Tim scanner at the department of Radiology of Xijing Hospital.The first 10 time points were discarded to ensure MRI data stability.Then the remaining 230 time points were slice timing corrected and realigned to the first image,during which the average frame-wise displacement(FD)was obtained.Inter-scan motion was assessed using the translation and rotation parameters,and an exclusion criterion of > 2.5 mm translation and/or > 2.5° rotation in each direction at each time point was set.Two SZ patients,two HCs,and three FDRs met the criteria and were excluded from further analyses,resulting in 38 SZ patients,33 FDRs and 38 HCs for final inclusion.As FC measures are sensitive to head motion,Friston-24 parameters were used to regress out their effects.Then,the DARTEL toolbox was used to normalize the data into MNI space.The Craddock atlas was used to parcellate the whole brain into 200 regions of interest(ROI).The time series within each region were first bandpass filtered(0.01–0.08 Hz)and then averaged.For each participant,FC was calculated between each region of interest using Pearson’s correlation coefficients,resulting in 19900([200 × 199]/2)dimensional functional connectivity feature vectors for each subject.Before building the classifier model,an initial feature selection step was performed for data dimension reduction.The current study used F-score for feature ranking,which was shown to be an effective method in previous studies.Leave-one-out cross-validation(LOOCV)was used to evaluate the performance of the classifier.In LOOCV,one subject is used as test data and the classifier is trained on the remaining dataset.For each LOOCV iteration,the features were ranked from highest to lowest according to their F-score,and the first 644 features were used to build the classifier.The above-mentioned classification model was used to classify SZ patients and HCs,and our next analysis was to investigate whether the final classification model could be used to determine whether the FDRs showed similar FC patterns to SZ patients or HCs.After the classification model was built,the functional connectivity of each FDR was input as test data into each iteration of the LOOCV,to obtain its corresponding prediction label(1 or-1).Therefore,each FDR was given 76 individual prediction labels.The classification score,which is the average of the 76 prediction labels,was used as a robust measure to characterize the similarity of each FDR’s FC pattern to an SZ pattern(in the range of-1 to 1,a positive score indicated a SZ pattern).Finally,a general linear model was used to investigate the correlations between classification scores and measures of cognitive function in FDRs,with age,sex,years of education as covariates.A semantic fluency test(animal version)was administered to evaluate the executive function and the semantic memory which are severely affected in schizophrenia,the performance was analyzed using the number of correct words within one minute.3)The study sample consisted of 38 first-episode SZ patients from early intervention services within the outpatient clinic and inpatient department of Xijing Hospital and 38 HCs recruited from the local community by advertisements.GE 3.0 Tesla Discovery MR scanner with eight–channel phased array head coil(EXCITE,General Electric,Milwaukee,Wisconsin)was used to collect the imaging data.The f MRI data were preprocessed by the Data Processing & Analysis for Brain imaging(DPABI,http://rfmri.org/dpabi)which is similar with study two.After getting the preprocessed f MRI data,we calculated the Pearson correlation coefficients between time courses of each pair of voxels,then,a correlation matrix was firstly obtained.To remove the weak correlations that might be induced by noise,a threshold of r > 0.25 was used to obtain the undirected adjacency matrix.Then,for each voxel,the degree centrality was calculated as the sum of the connections between this voxel with other voxels.For further statistical analysis,the weighted DC was converted into a zscore map.Finally,the DC map was smoothed with 6 mm FWHM Gaussian kernel.For the present study,treatment response was operationally defined using criteria described below:In brief,specific items from PANSS were chosen including the positive symptoms of delusions(P1),conceptual disorganization(P2)and hallucinatory behavior(P3),the negative symptoms of blunted affect(N1),social withdrawal(N4)and lack of spontaneity(N6),the general psychopathology of mannerisms/posturing(G5)and unusual thought content(G9).And any of the above-mentioned item with a score>3 was considered as nonresponse.Based on the response criteria,19 patients were classified as responders and 19 patients were classified as non-responders.SVM was applied by using the Pattern Recognition for Neuroimaging Toolbox(PRo NTo)(http://www.mlnl.cs.ucl.ac.uk/pronto)to investigate whether the baseline DC can classify antipsychotic drugs treatment outcomes.The first step was feature selection,the differences between SZ patients at baseline and HCs were firstly obtained and binarized as the mask.Then,feature selection consisted of identifying brain regions that belong to the mask.These procedures were processed in“Prepare feature set” program.Secondly,build the SVM classification model and use the Leave-one-out cross-validation(LOOCV)strategy to evaluate the performance of the classifier.In LOOCV,one subject is used as test data and the classifier is trained on the remaining dataset.These procedures were processed in “Specify model” program.Thirdly,a 1,000-times permutation test was used to evaluate the performance of the SVM model,the corresponding accuracy,sensitivity,specificity and AUC(the area under the receiver operating characteristic curve)were obtained.One advantage of the PRo NTo is that the weight map can be built at voxel level.According to the contribution in the classification model,the region contributions can be ranked and presented for illustration.Results1)The spatial distribution was roughly consistent between FZ-SZ patients and HCs.The visual network mainly includes the areas involved in vision processing such as the occipital lobe and the lingual gyrus.The auditory network mainly comprises areas including the heschl’s gyrus,superior frontal gyrus,superior temporal gyrus,and inferior temporal gyrus which are involved in auditory processing.The sensorimotor network mainly includes the central prefrontal and frontal lobe.The speech network mainly includes the temporal lobe,temporal pole,and inferior frontal gyrus,which are mainly involved in language organization and expression.The semantic network mainly includes the temporal pole and inferior temporal gyrus,which are mainly involved in semantic processing.The default mode network mainly includes the classical regions such as the middle temporal gyrus,postcentral gyrus and precuneus.The executive control network mainly includes areas of the dorsal lateral prefrontal cortex and inferior frontal gyrus that are mainly involved in high-order cognitive processing.The salience network mainly includes inferior frontal gyrus,insula,and angular gyrus.Slope differences were mainly found within auditory network and executive control network.For auditory network,FE-SZ patients showed altered association between seeding region Heschl ’s gyrus and middle frontal gyrus and between seeding region Heschl ’s gyrus and superior frontal gyrus.Positive correlations were found for the two pairs of regions in HCs(r = 0.27,P = 0.0002;r = 0.32,P < 0.0001,respectively),while negative correlations were found for the two pairs of regions in FE-SZ patients(r =-0.18,P = 0.009;r =-0.08,P = 0.24,respectively).Conversely,for executive control network,FE-SZ patients showed altered association between seeding region DLPFC and superior frontal gyrus and between seeding region DLPFC and supplementary motor area.Positive correlations were found for the two pairs of regions in FE-SZ patients(r =0.41,P< 0.0001;r=0.30,P < 0.0001,respectively),while no correlations were found for the two pairs of regions in HCs(r =-0.03,p=0.71;r =-0.06,p = 0.41,respectively).2)The accuracy of using the linear SVM classifier to distinguish SZ patients from normal subjects reached up to 88.15%(84.06% for sensitivity,92.18% for specificity).The discriminative score for each tested individual was acquired from the SVM classifier and a receiver operating characteristics(ROC)curve was created,which showed an area under the curve(AUC)of 0.93,indicating good classification power.A non-linear SVM classifier was also trained and showed similar results;however,to reduce the risk of overfitting and to directly calculate and exhibit the functional connectivity weights and ROI weights,the following analysis is based on the linear SVM classifier.In this study,397 consensus features were identified.Eighteen regions were identified as having weights that were at least one standard deviation greater than the average of the weights of all regions.The ROIs making the greatest contribution to the model were located within the default mode network(angular gyrus,middle temporal gyrus,orbital frontal gyrus,temporal pole,and inferior frontal gyrus),frontal-parietal network(superior parietal gyrus and parietal operculum cortex),auditory network(Heschl’s gyrus),and sensorimotor network(precentral gyrus and postcentral gyrus).Six FDRs were given a classification score of 1(SZ-specific)in all 76 LOOCV iterations,seven FDRs were classified as-1(HCs-specific)in all 76 LOOCV iterations,and the remaining FDRs were classified either as 1 or-1 in different iterations of the LOOCV.A significant negative correlation was found between the average classification scores of the FDRs and the semantic fluency scores.3)We adopted the DC study based on graph theory and found that patients had both areas of significantly increased DC and areas of significantly decreased DC compared with normal controls.According to the efficacy criteria defined in previous methods,we classified patients and normal controls using machine learning method based on DC.A total of 32 patients with SZ were successfully classified,with 17 of 19 patients who responded and 15 of 19 patients who did not respond.We obtained an accuracy of 84.2% with a sensitivity of 78.9% and specificity of 89.5% for classification of the two groups.The area under the curve was 0.94.The brain regions that contributed most to the classification are listed below,the top 10 regions are the right putamen(discriminative weight 4.51%),left inferior frontal gyrus(discriminative weight 4.21%),left putamen(discriminative weight4.19%),left middle occipital cortex(discriminative weight 4.13%),left middle frontal gyrus(discriminative weight 3.92%),left cerebellum(discriminative weight 3.81%),left medial frontal gyrus(discriminative weight 3.78%),right middle frontal gyrus(discriminative weight 3.74%),left inferior temporal gyrus(discriminative weight 3.41%)and left angular(discriminative weight 3.41%).ConclusionCompared with normal controls,we confirm that schizophrenia is a dysconnectivity disease using brain structural network.For the auditory network,the gray matter volume of the superior frontal gyrus and middle frontal gyrus were positively correlated with Heschl’s gyrus in the healthy population and negatively correlated in the patient.For the executive control network,there was no correlation between superior frontal gyrus and dorsolateral prefrontal cortex,but there was a positive correlation in patients in our study.The same condition can be found between supplementary motor area and dorsolateral prefrontal cortex.we found that using Multivariate Pattern Analysis,we could accurately classify SZ patients and HC based on whole-brain functional connectivity data.Functional connections within and between default model network,fronto-parietal network,auditory network,and sensorimotor network played the greatest role.The classification score obtained in the machine learning model can be used as an effective and sensitive biomarker to predict whether first-degree relatives of SZ are at high risk.The degree centrality of SZ patients was significantly different from that of normal people,and the difference between the degree centrality of SZ patients and normal people was significantly decreased after the treatment of the second generation antipsychotic drugs.Baseline degree centrality can be used to predict patient outcomes after antipsychotic treatment.The regions that contributed more to the classification effect included the right putamen,left inferior frontal gyrus,left putamen,left middle occipital gyrus,left middle frontal gyrus,left cerebellum,left medial frontal gyrus,right middle frontal gyrus,left inferior temporal gyrus and left angular.Our study first confirmed that schizophrenia is indeed a disorder of disconnection at the structural level of gray matter.With the help of magnetic resonance imaging and machine learning algorithms,the characteristics of brain networks in schizophrenia can be used as sensitive biomarkers for diagnostic classification and treatment prediction of SZ. |
PDF Full Text Request