Introduction
Schizophrenia is often a chronic, life-long condition with an early onset (Rabinowitz et al. Reference Rabinowitz, Levine, Haim and Häfner2007; Andreasen et al. Reference Andreasen, Nopoulos, Magnotta, Pierson, Ziebell and Ho2011). It accounts for 1.1% of the total disability-adjusted life years (DALYs) and 2.8% of years lived with disability (Levav & Rutz, Reference Levav and Rutz2002) and is the eighth leading cause of DALYs worldwide in the 15–44 years age group. Delayed diagnosis may result in brain structural/functional alterations (Penttilä et al. Reference Penttilä, Jääskeläinen, Haapea, Tanskanen, Veijola, Ridler, Murray, Barnes, Jones, Isohanni, Koponen and Miettunen2010; Guo et al. Reference Guo, Li, Wei, Fan, Kennedy, Shen, Chen and Zhao2013, Reference Guo, Liu, Xiao, Yu, Zhang, Liu, Zhang and Zhao2015), which complicate the treatment and may result in poor cognitive and social functioning (Malla et al. Reference Malla, Bodnar, Joober and Lepage2011). Therefore, the study of participants during their first episode of schizophrenia spectrum disorders (FES) is of high relevance, as it could improve early diagnosis. By limiting the effects of illness burden and medication exposure (Lieberman et al. Reference Lieberman, Chakos, Wu, Alvir, Hoffman, Robinson and Bilder2001; Smieskova et al. Reference Smieskova, Fusar-Poli, Allen, Bendfeldt, Stieglitz, Drewe, Radue, McGuire, Riecher-Rössler and Borgwardt2009; Ho et al. Reference Ho, Andreasen, Ziebell, Pierson and Magnotta2011), this approach could also help identify biological signatures of the illness.
Brain imaging provides an interesting tool for the study of psychiatric disorders. Yet, the diagnostic promise of neuroimaging in psychiatry has not yet been fully realized. The main reasons for this pertain to clinical heterogeneity and low sensitivity/specificity of brain imaging findings. The former issue can be addressed by studying more homogeneous patient populations, such as participants at the early stages of the illness. The later problem can be resolved by novel methods of magnetic resonance imaging (MRI) data analysis, such as machine learning (ML). Unlike conventional univariate methods, which yield significant results on a group level, multivariate ML classifiers are sensitive enough to accurately classify individual subjects (Haller et al. Reference Haller, Lovblad, Giannakopoulos and Van De Ville2014; Sundermann et al. Reference Sundermann, Herr, Schwindt and Pfleiderer2014, Koutsouleris et al. Reference Koutsouleris, Meisenzahl, Borgwardt, Riecher-Rössler, Frodl, Kambeitz, Köhler, Falkai, Möller, Reiser and Davatzikos2015). Moreover, multivariate patterns of brain changes, such as those detected by ML, may be more characteristic of psychiatric disorders (Davatzikos et al. Reference Davatzikos, Shen, Gur, Wu, Liu, Fan, Hughett, Turetsky and Gur2005). Thus, ML may help bring neuroimaging from the bench to the bedside (Hajek et al. Reference Hajek, Cooke, Kopecek, Novak, Hoschl and Alda2015).
Schizophrenia has been associated with aberrant functional connectivity (FC) within and between the default mode network (DMN), the central executive network (CEN) and the salience network (SN) (Palaniyappan et al. Reference Palaniyappan, Simmonite, White, Liddle and Liddle2013; Manoliu et al. Reference Manoliu, Riedl, Zherdin, Muhlau, Schwerthoffer, Scherr, Peters, Zimmer, Forstl, Bauml, Wohlschlager and Sorg2014; Nekovarova et al. Reference Nekovarova, Fajnerova, Horacek and Spaniel2014; Spaniel et al. Reference Spaniel, Tintera, Rydlo, Ibrahim, Kasparek, Horacek, Zaytseva, Matejka, Fialova, Slovakova, Mikolas, Melicher, Görnerova, Höschl and Hajek2016). Thus we investigated whether patterns of resting-state FC (rsFC), within these networks, would allow us to differentiate healthy controls from participants at the early stages of schizophrenia.
ML applied to resting-state functional MRI (rsfMRI) was successfully used to classify patients with an established illness (Kambeitz et al. Reference Kambeitz, Kambeitz-Ilankovic, Leucht, Wood, Davatzikos, Malchow, Falkai and Koutsouleris2015). However, to the best of our knowledge, this is the first study to use rsFC to differentiate participants with a FES from healthy controls. In addition, we compared the results obtained by ML with those derived from traditional between-group FC analyses.
Method
Study design
We recruited participants through the Early-Stage Schizophrenia Outcome study (ESO), a prospective trial of FES subjects, conducted in Prague and the Central Bohemia areas (Melicher et al. Reference Melicher, Horacek, Hlinka, Spaniel, Tintera, Ibrahim, Mikolas, Novak, Mohr and Hoschl2015). The study was carried out in accordance with the latest version of the Declaration of Helsinki. A written informed consent was obtained from all of the subjects and the local ethics committee approved the protocol.
Patients with FES
We recruited patients hospitalized at the Bohnice Psychiatric Hospital, Prague. The inclusion criteria were: (1) the diagnosis of schizophrenia or the diagnosis of an acute polymorphic psychotic disorder, as made by a psychiatrist, according to the International Classification of Diseases-10 criteria; (2) the first episode of psychotic illness; (3) the duration of untreated psychosis ⩽24 months. Any patients with psychotic mood disorders (including schizo-affective disorder, bipolar disorder, and unipolar depression with psychotic symptoms) were excluded from the study. All of the patients were treated with antipsychotic drugs at the time of the MRI scanning. We rated the symptom severity using the Positive and Negative Syndrome Scale (PANSS) (Kay et al. Reference Kay, Fiszbein and Opler1987).
The MRI scanning was performed during the first hospitalization, as soon as the patients were able to understand the purpose of the study and undergo the fMRI protocol. As participants were typically hospitalized shortly after developing symptoms, some of them did not meet the duration criteria for schizophrenia at the time of scanning. These patients received the working diagnosis of acute polymorphic psychotic disorder.
Healthy controls
The healthy control subjects were recruited via an advertisement from a similar sociodemographic background and were matched to FES participants by age and sex on an individual basis. The main exclusion criteria for the control subjects were a personal lifetime history of any psychiatric disorder, or any substance abuse, established by the Mini International Neuropsychiatric Interview (M.I.N.I.) (Lecrubier et al. Reference Lecrubier, Sheehan, Weiller, Amorim, Bonora, Harnett Sheehan, Janavs and Dunbar1997). We also excluded any family history of a psychiatric illness in first- or second-degree relatives (see Table 1 for details).
Table 1. Demographic and clinical characteristics of the participants
df, Degrees of freedom; s.d., standard deviation; n.a., not applicable; PANSS, Positive and Negative Syndrome Scale.
a A total of 13 patients received the working diagnosis of acute polymorphic psychotic disorder, because they did not meet the duration criteria for schizophrenia at the time of scanning.
b Data missing for six patients.
Further exclusion criteria, for both the patients and the healthy controls, included current neurological disorders, a lifetime history of seizures, or a head injury with altered consciousness, an intracranial hemorrhage, history of mental retardation, substance dependence, and any contraindications for MRI scanning.
fMRI data acquisition
The data were acquired by a 3 T Siemens Trio MRI scanner (Germany) equipped with a standard head coil. For the fMRI data pre-processing, the subjects were scanned using a structural T1-weighted three-dimensional MPRAGE sequence [repetition time (TR) 2300 ms, echo time (TE) 4.63 ms, bandwidth 130 Hz/pixel, field of view (FOV) 256 × 256 mm, matrix 256 × 256, 160–224 contiguous sagittal slices, a voxel size of 1 × 1 × 1 mm3, GRAPPA, and Acceleration Factor 2]. Functional images sensitive to the blood oxygen level-dependent (BOLD) contrast were measured with a gradient echo echo-planar sequence (GRE-EPI, TR = 2000 ms, TE = 30 ms, flip angle 90°, bandwidth 2232 Hz/pixel, without parallel acceleration, FOV = 192 mm × 144 mm, matrix size 64 × 48, a voxel size of 3 × 3 × 3 mm3, each volume with 35 axial slices without an inter-slice gap, and a total of 400 volumes).
fMRI data processing
Functional/structural data were pre-processed and analysed using tools implemented in the MATLAB 7.14 (R2012a) software. Slice-timing, realignment, regression of nuisance covariates (white matter and cerebrospinal fluid signal, voxel specific head motion, mean signal), spatial normalization and smoothing of the functional images, as well as spatial normalization of the structural T1 images, were performed by an SPM8-Data Processing Based Toolbox Assistant for the Resting-fMRI (DPARSF) (Chao-Gan & Yu-Feng, Reference Chao-Gan and Yu-Feng2010), and the Resting-State fMRI Data Analysis Toolkit (REST) (Song et al. Reference Song, Dong, Long, Li, Zuo, Zhu, He, Yan and Zang2011). The images were smoothed with an 8 × 8 × 8 Gaussian kernel. We applied temporal filtering over the frequency band of 0.008–0.09 Hz.
FC analysis
rsFC for all of the subjects was calculated between pre-selected regions of interest (ROIs) and the voxels in the rest of the brain (seed-based connectivity), using pre-processing pipelines which are well established and often used in the field (i.e. Craddock et al. Reference Craddock, Holtzheimer, Hu and Mayberg2009; Alonso-Solís et al. Reference Alonso-Solís, Corripio, de Castro-Manglano, Duran-Sindreu, Garcia-Garcia, Proal, Nuñez-Marín, Soutullo, Alvarez, Gómez-Ansón, Kelly and Castellanos2012; Venkataraman et al. Reference Venkataraman, Whitford, Westin, Golland and Kubicki2012). We selected three ROIs, which correspond to the three networks of interest, i.e. posterior cingulate gyrus for the DMN, dorsolateral prefrontal cortex (DLPFC), as represented by the middle frontal gyrus for the CEN and anterior insula for the SN. The posterior cingulate gyrus and the DLPFC were selected from the Automated Anatomical Labeling (AAL) atlas (Tzourio-Mazoyer et al. Reference Tzourio-Mazoyer, Landeau, Papathanassiou, Crivello, Etard, Delcroix, Mazoyer and Joliot2002). This atlas does not provide parcellation of insula, which we obtained from Freesurfer (sulcus circular insulae) (Fischl et al. Reference Fischl, Salat, van der Kouwe, Makris, Ségonne, Quinn and Dale2004; Sridharan et al. Reference Sridharan, Levitin and Menon2008; Destrieux et al. Reference Destrieux, Fischl, Dale and Halgren2010; Menon & Uddin, Reference Menon and Uddin2010; Palaniyappan et al. Reference Palaniyappan, Simmonite, White, Liddle and Liddle2013). As patients with established schizophrenia as well as those with FES show reduced asymmetry of rsFC, we used ROIs from both sides as seeds in each model (Damoiseaux et al. Reference Damoiseaux, Rombouts, Barkhof, Scheltens, Stam, Smith and Beckmann2006; Swanson et al. Reference Swanson, Eichele, Pearlson, Kiehl, Yu and Calhoun2011; Cabral et al. Reference Cabral, Kringelbach and Deco2014; Guo et al. Reference Guo, Xiao, Liu, Wooderson, Zhang, Zhang, Yu and Liu2014). Thus, for each subject, we calculated three connectivity maps, including connectivity between: (1) bilateral anterior insula; (2) bilateral posterior cingulate; (3) bilateral middle frontal gyrus and the rest of the brain. Correlation coefficients were transformed into Z-scores by Fisher's transformation. These first-level, individual subject connectivity maps were subjected to ML; see the support vector machines (SVM) classification paragraph below.
The second-level FC analysis was performed by the SPM8-Data Processing Based Toolbox Assistant. The differences in the seed-based FC between the patients and the controls were tested using a two-sample t test. The results were family-wise error corrected with a significance threshold of p < 0.05 on a cluster level. Only clusters exceeding 20 voxels were considered significant. In order to compare the uncorrected differences in the seed-based FC with the weight distribution obtained by the SVM, we performed a t test on an uncorrected level, with a cluster level of p = 0.001.
SVM classification
We applied a linear kernel support-vector classifier (SVM) implemented in the PRONTO toolbox v 1.1 (Schrouff et al. Reference Schrouff, Rosa, Rondina, Marquand, Chu, Ashburner, Phillips, Richiardi and Mourão-Miranda2013) to the individual subject FC maps. We estimated three separate SVM classification models (one for each of the three ROIs corresponding to the three pre-selected resting-state networks). In order to remove zero voxels, all normalized brains were extracted using a BET tool implemented in the FSL package (Smith et al. Reference Smith, Jenkinson, Woolrich, Beckmann, Behrens, Johansen-Berg, Bannister, De Luca, Drobnjak, Flitney, Niazy, Saunders, Vickers, Zhang, De Stefano, Brady and Matthews2004) and multiplied with each other in order to generate a common mask. The ML analyses were restricted to gray matter, by using a normalized gray matter mask provided by Dr Wager's laboratory (http://wagerlab.colorado.edu/wiki/doku.php/help/core/brain_masks). Each FC map comprised 52941 voxels (features).
This study utilized a linear kernel SVM, which is less prone to overfitting than non-linear SVMs. Linear kernel SVMs have a single parameter, C, that controls the trade-off between having zero training errors and allowing misclassifications. Similar to other studies, we used the default parameter C = 1 (Mourao-Miranda et al. Reference Mourao-Miranda, Reinders, Rocha-Rego, Lappin, Rondina, Morgan, Morgan, Fearon, Jones, Doody, Murray, Kapur and Dazzan2012; Rocha-Rego et al. Reference Rocha-Rego, Jogia, Marquand, Mourao-Miranda, Simmons and Frangou2014; Hajek et al. Reference Hajek, Cooke, Kopecek, Novak, Hoschl and Alda2015). It has been shown previously by LaConte et al. (Reference LaConte, Strother, Cherkassky, Anderson and Hu2005) that the SVM performance for whole-brain classification does not change for a large range of C values and only degrades with very small C values. Thus, modifying the C threshold was suggested only when the dimensionality of the data is smaller than the number of examples (e.g. classification based on small ROIs), which was not the case in our study. Others have suggested that using a sample-dependent optimization of the parameter C may improve the performance of the model (Franke et al. Reference Franke, Ziegler, Klöppel and Gaser2010; Nieuwenhuis et al. Reference Nieuwenhuis, van Haren, Hulshoff Pol, Cahn, Kahn and Schnack2012). However, the aim of this study was not to optimise the SVM methods. Our goal was to reduce the methodological heterogeneity and use a simple, ‘out of the box’ approach, which could be applicable in clinical setting (Mourao-Miranda et al. Reference Mourao-Miranda, Reinders, Rocha-Rego, Lappin, Rondina, Morgan, Morgan, Fearon, Jones, Doody, Murray, Kapur and Dazzan2012).
We performed a leave-one-subject-per-group-out cross-validation. This means, that on each run, one subject from each group was assigned to a testing set and the remaining subjects were assigned to a learning set. The classification was then performed on the two subjects in the test set. This was repeated until all of the subjects had been tested. During the cross-validation procedure, all of the patients were matched with the healthy controls according to age and sex. The classification accuracy was expressed as a total performance on all runs. The statistical significance of the obtained classification accuracy was tested on 1000 randomly permuted datasets, with a random assignment of the group class to the input image. A resulting null-hypothesis distribution was used to calculate the p value of the accuracies, i.e. the proportion of the permutations that yielded a greater accuracy than the accuracy found for the classification models.
Analysis of the effects of medication and symptoms
We attempted to clarify the contribution of medication and symptoms in several ways: (1) we compared symptoms and medication dose between correctly and incorrectly classified subjects using an independent-sample t test; (2) we used a linear regression to assess the association between classification accuracy (value of the SVM decision function) and medication dose [expressed as chlorpromazine (CPZ) equivalents] or symptoms; (3) we modeled the effects of covariates on FC. Of note, in ML, removal of confounding covariates can violate the basic train/test assumption by introducing the information about the whole dataset before introducing labels. Therefore covarying for medication dose or symptoms would not be optimal. To counter this problem, we thus used another ML approach – Gaussian process regression (GPR). Using the complex GPR model makes few assumptions about the shape of the possible relationship and is thus more powerful than linear models (Rasmussen & Williams, Reference Rasmussen and Williams2006; Schrouff et al. Reference Schrouff, Rosa, Rondina, Marquand, Chu, Ashburner, Phillips, Richiardi and Mourão-Miranda2013). Using linear model in this case could easily lead to underfitting in the case of a non-linear relationship. Therefore, by harvesting as much of the confounding relationship as possible, the GPR ensures that we exhaustively investigated and quantified the contribution of potential confounding factors to our findings. Specifically, we performed a GPR in order to estimate the CPZ dose, or the total PANSS and three subscales (positive, negative, general), on the day of scanning from the FC maps (Rasmussen & Williams, Reference Rasmussen and Williams2006). We only applied these analyses to networks, which differentiated FES from controls above chance level. All analyses were performed by the PRoNTo Toolbox v.1.1 (Schrouff et al. Reference Schrouff, Rosa, Rondina, Marquand, Chu, Ashburner, Phillips, Richiardi and Mourão-Miranda2013).
Ethical standards
All procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Results
Demographic data
We recruited 63 FES participants who were individually matched by age and sex to 63 healthy controls, without a personal or a family history of psychiatric disorder; see Table 1 for a description of the samples.
Support vector classification of FES and control participants
ML applied to rsFC within the SN differentiated FES from the control participants with specificity of 71.4% (p = 0.001), sensitivity of 74.6% (p = 0.001), and balanced accuracy of 73.0% (p = 0.001, area under the receiver operating characteristic curve = 0.80, Fig. 1). In other words, among the 63 FES subjects, 16 individuals were mislabeled as being controls, whereas 18 out of 63 controls were incorrectly classified as FES. The regions which most contributed to the discrimination of the two groups contained the anterior and posterior cingulate, precuneus, ventro- and dorsolateral prefrontal cortex, angular and supramarginal gyri, temporo-occipital regions, lingular gyrus, and thalamus. The FC within the other networks, i.e. the DMN and CEN, did not yield classifications above chance level (see Table 2).
Fig. 1. The receiver operating characteristic (ROC) curve for the classification model was based on functional connectivity maps of the anterior insula. The ROC curve shows the sensitivity v. specificity trade-off and the area under the curve (0.80). True positives = sensitivity; false positives = 1 – specificity.
Table 2. Results of the support vector machine classification of patients with a first-episode schizophrenia spectrum disorder and healthy controls from the whole-brain seed-based functional connectivity maps a
ROI, Region of interest; AUC, area under receiver operating characteristic curve.
a p Values were calculated from permutation testing with 1000 permutations.
b See Fig. 1 for details.
FC analysis – between-group comparisons
There were no significant differences in rsFC between the FES and the control participants on a corrected level for either of the seed regions, i.e. the posterior cingulate cortex, dorsolateral prefrontal cortex and anterior insula. We further tested differences in rsFC of the anterior insula on an uncorrected level. The uncorrected comparisons yielded between-group connectivity differences in the bilateral angular and supramarginal gyri. These regions closely overlapped with the maximum weight vectors obtained from the ML classification model based on SN connectivity (Fig. 2)
Fig. 2. Comparison of the regions contributing to the classification based on salience network connectivity obtained by machine learning (red) with differences between the patients with first episode of schizophrenia-spectrum disorder and controls in salience network connectivity (white). Differences in functional connectivity between the patients and the controls were obtained as F-contrast (p = 0.001, cluster-level uncorrected).
Effects of medication and psychotic symptoms
The correctly classified patients did not differ from the misclassified ones in PANSS scores [t = 0.08, degrees of freedom (df) = 61, p = 0.9] or medication dose (t = −1.66, df = 60, p = 0.1). There was no association between classification accuracy and CPZ equivalents (r = −0.21, p = 0.11) or symptoms (r = −0.07, p = 0.58). Lastly, ML (GPR) was unable to estimate the CPZ dose (r = 0.22, p = 0.09) or the current symptoms as measured by the PANSS total scores or subscales (PANSS-total, r = −0.03, p = 0.81; PANSS-pos, r = −0.09, p = 0.472; PANSS-neg, r = 0.17, p = 0.184; PANSS-gen, r = −0.03, p = 0.81) from the rsFC within the SN.
Discussion
In this study, whole-brain rsFC maps of the anterior insula/SN differentiated FES participants from controls with an above-chance accuracy of 73%, specificity of 71.4% and sensitivity of 74.6%. The classification was not significantly affected by medication dose, or by the presence of psychotic symptoms, and thus was probably based on trait rather than state markers.
Our findings provide a proof of concept that rsfMRI can be successfully used to differentiate participants with FES from healthy controls. rsfMRI is particularly suitable for diagnostic classification (Craddock et al. Reference Craddock, Holtzheimer, Hu and Mayberg2009; Haller et al. Reference Haller, Lovblad, Giannakopoulos and Van De Ville2014). Compared with task-based fMRI paradigms, the resting-state approach reduces the amount of bias introduced by task non-adherence. This is an important advantage in participants with FES, who tend to have distorted perception and impaired cognitive functioning. Also the rsFC seed regions can be selected a priori from pre-defined atlases, which makes this approach transferable among independent groups of subjects. Although it is difficult to directly compare the results due to the methodological differences, the classification performance obtained by using rsFC in this study was comparable with the sensitivity and specificity obtained from structural MRI and rsfMRI in established schizophrenia patients and FES as reported in a recent meta-analysis by Kambeitz et al. (Reference Kambeitz, Kambeitz-Ilankovic, Leucht, Wood, Davatzikos, Malchow, Falkai and Koutsouleris2015) (sensitivity 76% and specificity 79% for structural MRI, sensitivity 84% and specificity 77% for rsfMRI). Studies that focused on FES exclusively have reported accuracies ranging from 65.8 to 94% for structural MRI (Kasparek et al. Reference Kasparek, Thomaz, Sato, Schwarz, Janousova, Marecek, Prikryl, Vanicek, Fujita and Ceskova2011; Takayanagi et al. Reference Takayanagi, Takahashi, Orikabe, Mozue, Kawasaki, Nakamura, Sato, Itokawa, Yamasue, Kasai, Kurachi, Okazaki and Suzuki2011; Borgwardt & Fusar-Poli, Reference Borgwardt and Fusar-Poli2012; Pettersson-Yeo et al. Reference Pettersson-Yeo, Benetti, Marquand, Dell'Acqua, Williams, Allen, Prata, McGuire and Mechelli2013; Zanetti et al. Reference Zanetti, Schaufelberger, Doshi, Ou, Ferreira, Menezes, Scazufca, Davatzikos and Busatto2013).
The anterior insula is a crucial component of the SN. The disruption of anterior insula and SN connectivity has been well documented in schizophrenia (White et al. Reference White, Joseph, Francis and Liddle2010; Moran et al. Reference Moran, Tagamets, Sampath, O'Donnell, Stein, Kochunov and Hong2013; Palaniyappan et al. Reference Palaniyappan, Simmonite, White, Liddle and Liddle2013; Manoliu et al. Reference Manoliu, Riedl, Zherdin, Muhlau, Schwerthoffer, Scherr, Peters, Zimmer, Forstl, Bauml, Wohlschlager and Sorg2014; Iwabuchi et al. Reference Iwabuchi, Liddle and Palaniyappan2015). Additionally, a gray matter reduction within the insula has been consistently and robustly reported in meta-analyses of morphometric MRI studies in schizophrenia (Glahn et al. Reference Glahn, Laird, Ellison-Wright, Thelen, Robinson, Lancaster, Bullmore and Fox2008; Ellison-Wright & Bullmore, Reference Ellison-Wright and Bullmore2010; Bora et al. Reference Bora, Fornito, Radua, Walterfang, Seal, Wood, Yücel, Velakoulis and Pantelis2011; Palaniyappan et al. Reference Palaniyappan, Simmonite, White, Liddle and Liddle2013). However, the specificity of these changes to schizophrenia is unclear. Recent findings suggest that a general mapping exists between a broad range of psychiatric symptoms and the integrity of an anterior insula-based network across a wide variety of neuropsychiatric illnesses (Goodkind et al. Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata, Ortega, Zaiko, Roach, Korgaonkar, Grieve, Galatzer-Levy, Fox and Etkin2015).
Although the DMN and CEN exhibit functional abnormalities in schizophrenia (Ren et al. Reference Ren, Lui, Deng, Li, Li, Huang, Wang, Li, Sweeney and Gong2013; Spaniel et al. Reference Spaniel, Tintera, Rydlo, Ibrahim, Kasparek, Horacek, Zaytseva, Matejka, Fialova, Slovakova, Mikolas, Melicher, Görnerova, Höschl and Hajek2016), we did not achieve above-chance classifications when we focused on connectivity within these networks. We also did not identify any significant between-group FC differences in these networks. One explanation may lie in the dynamics of structural brain changes during the course of an illness. Some studies suggest that abnormalities of insular cortex in psychotic disorders may reflect pre-existing vulnerability. Later in the course of illness, the changes secondary to the illness burden may lead to the extension of these alterations to other neighboring structures (Takahashi et al. Reference Takahashi, Wood, Yung, Phillips, Soulsby, McGorry, Tanino, Zhou, Suzuki, Velakoulis and Pantelis2009; Chan et al. Reference Chan, Di, McAlonan and Gong2011). However, other studies do not support this view of dynamic pattern changes in brain morphology (Vita et al. Reference Vita, De Peri, Deste and Sacchetti2012).
Another explanation is that previous findings reflect clinical heterogeneity vis-à-vis the clinical course. Recent studies have demonstrated that classification accuracies are higher in participants with chronic course of illness, and lower in those with an episodic illness. Perhaps the presence of patients who will go on to develop an episodic illness could have decreased the prediction accuracies for some of the networks (Mourao-Miranda et al. Reference Mourao-Miranda, Reinders, Rocha-Rego, Lappin, Rondina, Morgan, Morgan, Fearon, Jones, Doody, Murray, Kapur and Dazzan2012; Gould et al. Reference Gould, Shepherd, Laurens, Cairns, Carr and Green2014).
It is of note that even the between-group differences in the anterior insula connectivity did not reach statistical significance. However, on an uncorrected level, the localization of group differences overlapped with the regions with a maximum contribution to the classification using ML. Due to the multivariate nature, which obviates the need to control for multiple comparisons, ML appears to be more sensitive than conventional mass univariate approaches.
This study has the following limitations. A common problem in ML is overfitting (Whelan & Garavan, Reference Whelan and Garavan2014). We used a linear-kernel SVM classifier, which was shown to have a low likelihood of overfitting in fMRI paradigms (LaConte et al. Reference LaConte, Strother, Cherkassky, Anderson and Hu2005; Mourao-Miranda et al. Reference Mourao-Miranda, Reinders, Rocha-Rego, Lappin, Rondina, Morgan, Morgan, Fearon, Jones, Doody, Murray, Kapur and Dazzan2012). In addition, two out of three of our models failed to produce significant results, which makes overfitting unlikely.
Unlike the healthy controls, all of the patients were on antipsychotic medication and experienced mild psychotic symptoms at the time of scanning. However, there was no significant association between symptoms or medication dose and classification accuracy. Likewise, ML was unable to estimate symptom levels or medication dose from rsFC data. Last but not least, there were no significant differences between the correctly and incorrectly classified subjects in symptom levels or medication dose. Thus, it is unlikely that medication or psychotic symptoms markedly contributed to the classification. There are a limited number of studies focusing on correlations of clinical symptom severity and/or medication dose with rsFC in FES. The results differ depending on the regions of interest. Some studies have found positive as well as negative correlations of striatal connectivity with symptom severity that was resolved with treatment (Sarpal et al. Reference Sarpal, Robinson, Lencz, Argyelan, Ikuta, Karlsgodt, Gallego, Kane, Szeszko and Malhotra2015). However, another study reported no correlations between symptom severity and rsFC of the DMN (Alonso-Solís et al. Reference Alonso-Solís, Corripio, de Castro-Manglano, Duran-Sindreu, Garcia-Garcia, Proal, Nuñez-Marín, Soutullo, Alvarez, Gómez-Ansón, Kelly and Castellanos2012).
We were primarily interested in subjects at the early stages of illness, as this is one of the few ways how to limit the effects of previous psychotic episodes and how to minimize exposure to medications or co-morbid conditions. This approach minimizes the effects of confounding variables, which could alter classification accuracy. Consequently, as many of the participants were hospitalized shortly after developing symptoms, some of them did not meet the duration criteria for schizophrenia at the time of scanning. These patients received the working diagnosis of acute polymorphic psychotic disorder.
As this is an emerging field, there is no standardization of the methods of data pre-processing and analyses for the ML studies. We do not know exactly the influence of individual processing steps or selection of ROI on the outcomes. At the same time, the SVMs are among the most used ML classifiers in psychiatric neuroimaging (Sundermann et al. Reference Sundermann, Herr, Schwindt and Pfleiderer2014).
Consequently, we used SVM with the default parameters in order to decrease the methodological heterogeneity and to ensure comparability with other studies. This way, we showed that even a simple ‘out of the box’, easily applicable classifier with linear (hyperplane) decision boundary could correctly distinguish between patients and controls. For the same reason, we used standard atlases to select the ROIs, which are representative of the individual networks. In future studies, it would be interesting to test other classifiers, such as random forest or discriminant analyses, or to focus on different ROIs.
To conclude, this study provides a proof of concept that ML of seed-based rsFC maps can be utilized for the diagnostic classification of participants early in the course of schizophrenia. We were able to discriminate patients with FES from healthy controls with an accuracy of 73%. Furthermore, our results emphasize the role of rsFC within the SN in the pathophysiology of FES. Future studies should attempt to use this approach in unaffected individuals at a genetic risk of schizophrenia, should aim to test the specificity of the results to schizophrenia and might benefit from novel classification algorithms.
Acknowledgements
The study was supported by: the Ministry of Health, Czech Republic (F.S., grant number 16-32696A); the Ministry of Health, Czech Republic – Conceptual Development of Research Organization 00023001 [Institut klinické a experimentální medicíny (IKEM), Prague, Czech Republic] – Institutional support; and Charles University, Prague (grant number 260168/SVV/2015).
The funding source had no role in the study design, the collection analysis, the interpretation of data, the writing of the paper, or in the decision for publication.
F.S., T.H. and P.M. designed the study; F.S., P.M., A.Sl. and M.M. participated in the patient recruitment; they also performed clinical examinations and scales. T.M. analysed the subject cohort demographics; P.M. and A.Sk. performed the fMRI pre-processing; P.M., E.B. and T.H. performed the fMRI data analyses and statistics; P.M., F.S. and T.H. wrote the article, which all of the authors reviewed and approved for publication.
Declaration of Interest
None.
