Hostname: page-component-745bb68f8f-cphqk Total loading time: 0 Render date: 2025-02-11T08:24:15.120Z Has data issue: false hasContentIssue false
  • English
  • Français

Machine learning multivariate pattern analysis predicts classification of posttraumatic stress disorder and its dissociative subtype: a multimodal neuroimaging approach

Published online by Cambridge University Press:  11 October 2018

Andrew A. Nicholson ,
Maria Densmore ,
Margaret C. McKinnon ,
Richard W.J. Neufeld ,
Paul A. Frewen ,
Jean Théberge ,
Rakesh Jetly ,
J. Donald Richardson  and
Ruth A. Lanius
Andrew A. Nicholson*
Affiliation:
Department of Neuroscience, Western University, London, ON, Canada Department of Psychiatry, Western University, London, ON, Canada Department of Psychiatry and Behavioural Neuroscience, McMaster University, Hamilton, ON, Canada Homewood Research Institute, Guelph, ON, Canada Imaging, Lawson Health Research Institute, London, ON, Canada
Maria Densmore
Affiliation:
Department of Psychiatry, Western University, London, ON, Canada Imaging, Lawson Health Research Institute, London, ON, Canada
Margaret C. McKinnon
Affiliation:
Department of Psychiatry and Behavioural Neuroscience, McMaster University, Hamilton, ON, Canada Homewood Research Institute, Guelph, ON, Canada Department of Mood Disorders Program, St. Joseph's Healthcare, Hamilton, ON, Canada
Richard W.J. Neufeld
Affiliation:
Department of Neuroscience, Western University, London, ON, Canada Department of Psychiatry, Western University, London, ON, Canada Department of Psychology, Western University, London, ON, Canada
Paul A. Frewen
Affiliation:
Department of Neuroscience, Western University, London, ON, Canada Department of Psychology, Western University, London, ON, Canada
Jean Théberge
Affiliation:
Department of Psychiatry, Western University, London, ON, Canada Imaging, Lawson Health Research Institute, London, ON, Canada Department of Medical Imaging, Western University, London, ON, Canada Department of Medial Biophysics, Western University, London, ON, Canada Department of Diagnostic Imaging, St. Joseph's Healthcare, London, ON, Canada
Rakesh Jetly
Affiliation:
Canadian Forces, Health Services, Ottawa, Ontario, Canada
J. Donald Richardson
Affiliation:
Department of Psychiatry and Behavioural Neuroscience, McMaster University, Hamilton, ON, Canada Homewood Research Institute, Guelph, ON, Canada Department of Mood Disorders Program, St. Joseph's Healthcare, Hamilton, ON, Canada
Ruth A. Lanius
Affiliation:
Department of Neuroscience, Western University, London, ON, Canada Department of Psychiatry, Western University, London, ON, Canada Imaging, Lawson Health Research Institute, London, ON, Canada
*
Author for correspondence: Andrew A. Nicholson, E-mail: anicho58@gmail.com
Rights & Permissions [Opens in a new window]

Abstract

Background

The field of psychiatry would benefit significantly from developing objective biomarkers that could facilitate the early identification of heterogeneous subtypes of illness. Critically, although machine learning pattern recognition methods have been applied recently to predict many psychiatric disorders, these techniques have not been utilized to predict subtypes of posttraumatic stress disorder (PTSD), including the dissociative subtype of PTSD (PTSD + DS).

Methods

Using Multiclass Gaussian Process Classification within PRoNTo, we examined the classification accuracy of: (i) the mean amplitude of low-frequency fluctuations (mALFF; reflecting spontaneous neural activity during rest); and (ii) seed-based amygdala complex functional connectivity within 181 participants [PTSD (n = 81); PTSD + DS (n = 49); and age-matched healthy trauma-unexposed controls (n = 51)]. We also computed mass-univariate analyses in order to observe regional group differences [false-discovery-rate (FDR)-cluster corrected p < 0.05, k = 20].

Results

We found that extracted features could predict accurately the classification of PTSD, PTSD + DS, and healthy controls, using both resting-state mALFF (91.63% balanced accuracy, p < 0.001) and amygdala complex connectivity maps (85.00% balanced accuracy, p < 0.001). These results were replicated using independent machine learning algorithms/cross-validation procedures. Moreover, areas weighted as being most important for group classification also displayed significant group differences at the univariate level. Here, whereas the PTSD + DS group displayed increased activation within emotion regulation regions, the PTSD group showed increased activation within the amygdala, globus pallidus, and motor/somatosensory regions.

Conclusion

The current study has significant implications for advancing machine learning applications within the field of psychiatry, as well as for developing objective biomarkers indicative of diagnostic heterogeneity.

Keywords

Amygdaladissociative subtypefMRImachine learningPTSDresting-state
Type
Original Articles
Information
Psychological Medicine , Volume 49 , Issue 12 , September 2019 , pp. 2049 - 2059
Copyright
Copyright © Cambridge University Press 2018 

Introduction

Individuals who survive traumatic events are at risk for developing posttraumatic stress disorder (PTSD), a psychiatric illness involving symptoms of vivid re-experiencing of the traumatic event, avoidance of trauma-related stimuli, alterations in cognitions and mood, and hyperarousal (APA, 2013). A dissociative subtype of PTSD (PTSD + DS) has been identified recently, denoting a particular class of patients that exhibit symptoms of depersonalization/derealization, as well as related emotional detachment and hypoemotionality (APA, 2013). Among patients with PTSD as compared to PTSD + DS, dissociation (depersonalization/derealization) is associated with differential patterns of neural activation, documented within multiple cortical and subcortical areas that include the prefrontal cortex (PFC), amygdala, cingulate cortex, insula, thalamus, and temporal cortex (Lanius et al., Reference Lanius, Williamson, Bluhm, Densmore, Boksman, Neufeld, Gati and Menon2005, Reference Lanius, Vermetten, Loewenstein, Brand, Christian, Bremner and Spiegel2010; Frewen and Lanius, Reference Frewen and Lanius2006; Hopper et al., Reference Hopper, Frewen, van der Kolk and Lanius2007; Felmingham et al., Reference Felmingham, Kemp, Williams, Falconer, Olivieri, Peduto and Bryant2008; Mickleborough et al., Reference Mickleborough, Daniels, Coupland, Kao, Williamson, Lanius, Hegadoren, Schore, Densmore, Stevens and Lanius2011; Duerden et al., Reference Duerden, Arsalidou, Lee and Taylor2013). Here, PTSD is associated with decreased regulatory activation from the medial prefrontal cortex (mPFC) and increased bottom-up activation of the amygdala/limbic system, with concomitant hyperarousal and emotion undermodulation (Lanius et al., Reference Lanius, Vermetten, Loewenstein, Brand, Christian, Bremner and Spiegel2010; Hayes et al., Reference Hayes, Hayes and Mikedis2012; Stevens et al., Reference Stevens, Jovanovic, Fani, Ely, Glover, Bradley and Ressler2013; Reinders et al., Reference Reinders, Willemsen, den Boer, Vos, Veltman and Loewenstein2014; Sadeh et al., Reference Sadeh, Spielberg, Warren, Miller and Heller2014; Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017; Pitman, Reference Pitman, Rasmusson, Koenen, Shin, Orr, Gilbertson, Milad and Liberzon2012). By contrast, PTSD + DS is characterized by increased top-down regulatory activation of the mPFC, resulting in hypoactivation of the amygdala/limbic system during symptom provocation, with concomitant detachment, autonomic blunting and emotion overmodulation (Hopper et al., Reference Hopper, Frewen, van der Kolk and Lanius2007; Lanius et al., Reference Lanius, Vermetten, Loewenstein, Brand, Christian, Bremner and Spiegel2010; Mickleborough et al., Reference Mickleborough, Daniels, Coupland, Kao, Williamson, Lanius, Hegadoren, Schore, Densmore, Stevens and Lanius2011; Harricharan et al., Reference Harricharan, Rabellino, Frewen, Densmore, Theberge, McKinnon, Schore and Lanius2016; Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017; Melara et al., Reference Melara, Ruglass, Fertuck and Hien2018). Other biological markers of PTSD + DS include single-nucleotide polymorphisms associated with dissociative symptoms (Wolf et al., Reference Wolf, Rasmusson, Mitchell, Logue, Baldwin and Miller2014) and toxic levels of cortisol that may contribute to exacerbated intergenerational patterns of adverse health outcomes (Seng et al., Reference Seng, Li, Yang, King, Low, Sperlich, Rowe, Lee, Muzik, Ford and Liberzon2017).

In line with the unique neural activation patterns associated with heterogeneous presentations of PTSD, we have also shown previously that PTSD and PTSD + DS patients, as well as healthy controls, display unique patterns of directed connectivity when using resting-state dynamic causal modeling (Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017). Here, PTSD patients showed a pattern of predominantly bottom-up connectivity from the amygdala and PAG to the vmPFC indicative of limbic overmodulation. By contrast, PTSD + DS patients showed predominantly top-down connectivity from the vmPFC to the amygdala and PAG, in keeping with previously described emotion overmodulation in this subtype. Notably, differential patterns of resting-state functional connectivity have also been documented in PTSD, PTSD + DS, and healthy controls, with respect to a multitude of regions spanning across the cortical and subcortical axis, including the basolateral (BLA) and centromedial (CMA) amygdala complexes (Nicholson et al., Reference Nicholson, Densmore, Frewen, Théberge, Neufeld, McKinnon and Lanius2015), insula subregions (Nicholson et al., Reference Nicholson, Sapru, Densmore, Frewen, Neufeld, Theberge, McKinnon and Lanius2016b), the superior colliculus (Olivé et al., Reference Olivé, Densmore, Harricharan, Théberge, McKinnon and Lanius2018), vestibular nuclei (Harricharan et al., Reference Harricharan, Nicholson, Densmore, Théberge, McKinnon, Neufeld and Lanius2017), the bed nucleus of the stria terminalis (BNST) (Rabellino et al., Reference Rabellino, Densmore, Harricharan, Jean, McKinnon and Lanius2017), and the periaqueductal gray (PAG) (Harricharan et al., Reference Harricharan, Rabellino, Frewen, Densmore, Theberge, McKinnon, Schore and Lanius2016).

Machine learning methods are sensitive enough to facilitate inference at the single-subject level, and can identify spatially distributed patterns in the brain that might be undetectable using group comparisons (Orrù et al., Reference Orrù, Pettersson-Yeo, Marquand, Sartori and Mechelli2012; Fu and Costafreda, Reference Fu and Costafreda2013; Wolfers et al., Reference Wolfers, Buitelaar, Beckmann, Franke and Marquand2015). Recently, a growing number of studies have applied machine learning methods to neuroimaging data to predict and characterize psychiatric diseases (Bleich-cohen et al., Reference Bleich-cohen, Jamshy, Sharon, Weizman, Intrator, Poyurovsky and Hendler2014; Mikolas et al., Reference Mikolas, Melicher, Skoch, Matejka, Slovakova, Bakstein, Hajek and Spaniel2016; Rive et al., Reference Rive, Redlich, Schmaal, Marquand, Dannlowski, Grotegerd, Veltman, Schene and Ruhé2016), as well as PTSD (Gong et al., Reference Gong, Li, Du, Pettersson-Yeo, Crossley, Yang, Li, Huang and Mechelli2014; Karstoft et al., Reference Karstoft, Galatzer-levy, Statnikov, Li and Shalev2015; Liu et al., Reference Liu, Xie, Wang and Guo2015; Omurca and Ekinci, Reference Omurca and Ekinci2015; Galatzer-Levy et al., Reference Galatzer-Levy, Ma, Statnikov, Yehuda and Shalev2017; Gradus et al., Reference Gradus, King, Galatzer-levy and Street2017; Jin et al., Reference Jin, Jia, Lanka, Rangaprakash, Li, Liu, Hu and Deshpande2017; Saxe et al., Reference Saxe, Ma, Ren and Aliferis2017). To date, however, no studies have examined the predictive validity of functional magnetic resonance imaging (fMRI) machine learning to classify PTSD and its dissociative subtype. Using a multivariate voxel pattern analysis (MVPA) or supervised machine learning (Orrù et al., Reference Orrù, Pettersson-Yeo, Marquand, Sartori and Mechelli2012; Fu and Costafreda, Reference Fu and Costafreda2013; Wolfers et al., Reference Wolfers, Buitelaar, Beckmann, Franke and Marquand2015), one can classify psychiatric disease from individual neuroimaging data (Orrù et al., Reference Orrù, Pettersson-Yeo, Marquand, Sartori and Mechelli2012). In keeping with this notion, it has been suggested further that the multivariate patterns of brain changes detected by machine learning may be highly sensitive to functional changes in the brain as a result of psychiatric illness, and thus may facilitate the translation of neuroimaging from the bench to the bedside (Davatzikos et al., Reference Davatzikos, Shen, Gur, Wu, Liu, Fan, Hughett, Turetsky and Gur2005; Mikolas et al., Reference Mikolas, Melicher, Skoch, Matejka, Slovakova, Bakstein, Hajek and Spaniel2016).

The objective of the current study was to use MVPA machine learning to predict individual classification of PTSD, PTSD + DS, and healthy individuals, using both resting-state fMRI activation and functional connectivity. Here, our main analyses consisted of an examination of the regional amplitude of low-frequency fluctuations (ALFF), reflecting spontaneous neural activity during the resting-state, in order to determine if resting-state activation could accurately predict PTSD subtypes. In addition, we evaluated the predictive accuracy of amygdala complex functional connectivity using machine learning in a secondary analysis. Given the distinct neural signatures of PTSD and PTSD + DS at both the activation and functional connectivity levels, we hypothesized that medial PFC areas responsible for emotion regulation and executive functioning, as well as regions within the limbic system (amygdala), innate alarm system (PAG) and basal ganglia, would predict accurately PTSD subtype classification when examining resting-state activation.

Methods

Participants

Our sample consisted of 181 participants [PTSD (n = 81); PTSD + DS (n = 49); age-matched healthy trauma-unexposed controls (n = 51); Table 1]. Most PTSD patients (90%) had early aversive experiences, where the majority of the sample was female. Exclusion criteria for patients included alcohol or substance abuse/dependence not in sustained full remission, and diagnosis of bipolar disorder or schizophrenia. Exclusion criteria for the control group included lifetime Axis-I or Axis-II disorders. Exclusion criteria for all participants included: noncompliance with 3 T fMRI safety standards, significant untreated medical illness, pregnancy, a history of neurological or pervasive developmental disorders, and previous head injury with loss of consciousness. All participants were evaluated using the Clinician Administered PTSD Scale (CAPS; IV and 5) (Blake et al., Reference Blake, Weathers, Nagy, Kaloupek, Gusman, Charney and Keane1995) and the DSM-IV Structured Clinical Interview (SCID) (First et al., Reference First, Spitzer, Gibbon and Williams2002). Dissociative subtype patients were identified by scoring ⩾2 for both frequency and intensity on either depersonalization or derealization CAPS symptoms as per standard methods (Nicholson et al., Reference Nicholson, Densmore, Frewen, Théberge, Neufeld, McKinnon and Lanius2015; Harricharan et al., Reference Harricharan, Rabellino, Frewen, Densmore, Theberge, McKinnon, Schore and Lanius2016). A battery of questionnaires was also administered [Childhood Trauma Questionnaire (CTQ) (Bernstein et al., Reference Bernstein, Stein, Newcomb, Walker, Pogge, Ahluvalia, Stokes, Handelsman, Medrano, Desmond and Zule2003), Beck's Depression Inventory (BDI) (Beck et al., Reference Beck, Guth, Steer and Ball1997), and Multiscale Dissociation Inventory (MDI) (Briere, Reference Briere2002); see Table 1 and online Supplemental Material for group comparisons on clinical variables and motion outliers]. Participants took part in a 6-min eyes-closed resting-state scan following standard methods (Nicholson et al., Reference Nicholson, Densmore, Frewen, Théberge, Neufeld, McKinnon and Lanius2015; Harricharan et al., Reference Harricharan, Rabellino, Frewen, Densmore, Theberge, McKinnon, Schore and Lanius2016) and were administered the Responses to Script-Driven Imagery Scale (RSDI; Hopper et al., Reference Hopper, Frewen, van der Kolk and Lanius2007) after the scan to assess state-based clinical symptoms experienced during the scan. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation. Among PTSD participants, 48 (PTSD, n = 29; PTSD + DS, n = 19) were receiving psychotropic treatment at the time of the study. Medications included antidepressants, atypical antipsychotics, sedatives, and anticonvulsants (see online Supplemental Material for details).

Table 1. Demographic and clinical information

PTSD, posttraumatic stress disorder; PTSD + DS, dissociative subtype posttraumatic stress disorder patients; CAPS, Clinician Administered PTSD Scale; CTQ, Childhood Trauma Questionnaire (none or minimal childhood trauma = 25–36, moderate = 56–68, extreme trauma > 72); BDI, Beck's Depression Inventory; MDI, Multiscale Dissociation Inventory; Dep/Dereal, depersonalization and derealization average; MDD, major depressive disorder; OCD, obsessive compulsive disorder; GAD, generalized anxiety disorder; s.d., standard deviation. *indicates the clinical variables on which all groups differed significantly from one another (p < 0.05).

a Indicates significantly higher clinical measures within a group as compared with the control group.

b Indicates significantly higher clinical measures as compared with the PTSD group. Here, the PTSD + DS group exhibited the highest scores among CAPS, CTQ, BDI, MDI-total, MDI depersonalization/derealization scales, and MDD diagnoses, as compared with the PTSD and control groups.

fMRI image acquisition and preprocessing

We acquired and preprocessed the neuroimaging data according to standard procedures in several of our manuscripts (Nicholson et al., Reference Nicholson, Densmore, Frewen, Théberge, Neufeld, McKinnon and Lanius2015, Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017; Harricharan et al., Reference Harricharan, Rabellino, Frewen, Densmore, Theberge, McKinnon, Schore and Lanius2016), and we refer the reader to the online Supplemental Material for a detailed description of the fMRI scanner and preprocessing protocols.

Data analysis

Resting-state data extraction

Here, we prepared individual subject images that would serve subsequently as inputs to the multivariate machine learning analyses, in addition to second-level mass-univariate analyses within SPM. We extracted two modalities of data: (a) the mean amplitude of low-frequency fluctuations (mALFF), which served as the main analysis for the study; and (b) functional connectivity maps of separate amygdala complexes for use in a supplementary analysis. Whereas the mALFF denotes spontaneous resting-state brain activation across the whole-brain, the functional connectivity maps denote areas displaying high-temporal correlations with the amygdala complexes across the whole-brain, thus serving as both data-driven and hypothesis-driven inputs, the mALFF measure has been used previously as an input to multivariate machine learning analyses, which predict accurately PTSD symptoms (Gong et al., Reference Gong, Li, Du, Pettersson-Yeo, Crossley, Yang, Li, Huang and Mechelli2014; Liu et al., Reference Liu, Xie, Wang and Guo2015). We used the REST toolbox (http://www.restfmri.net/forum) within Matlab2012a and SPM8 as a means to de-trend and extract individual ALFF maps from the preprocessed fMRI data (120 volumes) for each participant across the frequency band from 0.01 to 0.08 Hz. In order to obtain mALFF maps, the ALFF spatial maps were then normalized with each voxel divided by the whole-brain ALFF mean. To obtain amygdala complex resting-state connectivity measures, we extracted a mean signal intensity time course for each participant using SPMs Anatomy Toolbox for each of the six seed regions (bilateral BLA, CMA, and SFA) according to standard procedures within Matlab2017 and SPM12 (Nicholson et al., Reference Nicholson, Densmore, Frewen, Théberge, Neufeld, McKinnon and Lanius2015). Signal time courses were then used as regressors in a correlation analysis with whole-brain resting-state volumes for each participant, respectively.

Machine learning: multivariate pattern analysis

Our primary machine learning analysis concerned the mALFF data, where we also computed the same machine learning analysis with amygdala complex functional connectivity maps. Initially, in order to classify PTSD, PTSD + DS, and healthy controls based on mALFF patterns of neural activation, we implemented Multiclass Gaussian Process Classification algorithms within the PRoNTo toolbox (http://www.mlnl.cs.ucl.ac.uk/pronto/) (Schrouff et al., Reference Schrouff, Rosa, Rondina, Marquand, Chu, Ashburner, Phillips, Richiardi and Mourão-Miranda2013) running under Matlab2017 (Mathworks, 2018). Here, individual participant's mALFF maps served as inputs for the machine learning algorithm, in which a resting-state design was modeled with no conditions, and the DARTEL gray matter mask was applied. A feature set was prepared on whole brain mALFF data. Features were mean-centered, and a Multiclass Gaussian Process Classifier (MGPC) (Rasmussen and Willams, Reference Rasmussen and Willams2006) was used to test if whole brain mALFF activation could accurately predict the three groups. MGPC are probabilistic kernel methods, which can be used for multi-class problems (Rasmussen and Willams, Reference Rasmussen and Willams2006). MGPC is a supervised machine learning technique, similar to Support Vector Machines (SVMs) and Gaussian Process Classification (GPC), which can inform predictive probabilities of multiclass membership (Rasmussen and Willams, Reference Rasmussen and Willams2006; Wegrzyn et al., Reference Wegrzyn, Riehle, Labudda, Woermann, Baumgartner, Pollmann, Bien and Kissler2015).

In order to supplement the main analysis based on whole-brain resting-state activation, we evaluated the diagnostic predictive accuracy of a machine learning analysis with amygdala complex functional connectivity maps as inputs. Here, we observed classification accuracy for PTSD, PTSD + DS, and healthy controls using the same Multiclass GPC algorithm within PRoNTo (http://www.mlnl.cs.ucl.ac.uk/pronto/) (Schrouff et al., Reference Schrouff, Rosa, Rondina, Marquand, Chu, Ashburner, Phillips, Richiardi and Mourão-Miranda2013) running under Matlab (Mathworks, 2018). Individual amygdala complex functional connectivity maps (corresponding to the bilateral BLA, CMA, and SFA) for each participant served as inputs for the machine learning algorithm, in which a resting-state design was modeled with no conditions, and each amygdala seed was treated as a separate modality. Again, the DARTEL gray matter mask was applied to the individual connectivity maps. A feature set was prepared on whole brain amygdala functional connectivity maps, and multiple kernels were built for each amygdala complex. Features were mean-centered, and a MGPC was computed to evaluate if amygdala complex functional connectivity could predict accuracy diagnosis among the three groups.

For both the main mALFF and auxiliary amygdala complex seed-based functional connectivity machine learning applications, we used a leave-one-subject-out (LOSO) cross-validation strategy to estimate the generalization ability of our classifiers. We also computed a leave-one-subject-out-per-group (LOSOPG, see online Supplemental Results) cross-validation procedure in order to provide an exhaustive approach to this novel study. Subsequently, MGPC provided probabilistic predictions for each diagnostic category, and, in order to account for the group size difference, balanced accuracy measures were computed to assess the overall categorical performance of each classifier. Balanced accuracy takes the number of samples in each class into account and gives an equal weight to the accuracies obtained on test samples of each class. Statistical significance of these accuracy measures was determined by permutation testing (1000 permutations). Anatomical atlas weights were computed to visualize the relative importance of each region in the decision function of the MGPC, displaying the regional pattern of mALFF and amygdala complex functional connectivity used by the algorithm to categorize each group. Finally, in order to validate the results with an independent algorithm, similar to Rive et al. (Reference Rive, Redlich, Schmaal, Marquand, Dannlowski, Grotegerd, Veltman, Schene and Ruhé2016), we repeated the above analyses using binary GPC, classifying only PTSD v. PTSD + DS (see online Supplemental Material). This added a measure of construct validity to the study by replicating the current results with an independent algorithm, which is crucial given that machine learning has never been used previously to categorize PTSD, PTSD + DS, and healthy controls.

Mass-univariate analysis

In order to compliment the multivariate machine learning analysis, we conducted a mass-univariate analysis thereby informing regional differences in mALFF activation and amygdala complex functional connectivity between groups. Here, second-level analysis of variances (ANOVAs) within SPM12 were computed utilizing the FDR-cluster corrected threshold (p < 0.05, k = 20) in order to control conservatively for multiple comparisons (Eklund et al., Reference Eklund, Nichols and Knutsson2016). First, using mALFF maps, we conducted a three (group) one-way ANOVA. After observing a significant main effect, we then conducted follow-up t tests on the mALFF main analysis. Here, we compared mALFF activation between each group directly, maintaining the same error protection rate (FDR-cluster corrected, p < 0.05, k = 20). Second, for the amygdala complex functional connectivity supplementary analysis, we computed a 3 (group) × 6 (amygdala complex) full-factorial ANOVA. In an effort to mitigate multiple comparisons, we limited our analyses to the group × amygdala complex interaction at the FDR-cluster corrected level (p < 0.05, k = 20).

Clinical data and motion statistical analyses

We computed one-way ANOVAs in order to observe any potential between group differences in CAPS-IV and CAPS-5 total scores, CTQ, BDI, MDI-total, and MDI depersonalization/derealization average scores. We also examined self-report questionnaires of state depersonalization and derealization symptoms during the scan [Response to Script Driven Imagery (RSDI) adapted for resting-state scans (Hopper et al., Reference Hopper, Frewen, van der Kolk and Lanius2007)]. In addition, we examined potential correlations between trait MDI depersonalization/derealization and state RSDI depersonalization/derealization symptoms of dissociation. Additional motion outlier regressors were created through ART software, which were used to calculate extra regressors for motion outliers and movement and were included in each participant's first-level GLM. Here, we computed chi-squared statistics on the number of motion outlier parameters generated by ART across each group, as well as MDD diagnosis in a separate analysis. Some of the PTSD patients were taking SSRIs, SNRIs, benzodiazepines, and atypical antipsychotics at the time of study; therefore, we conducted a chi-squared analysis examining potential differences in the frequency of medication use between groups.

Results

Pattern classification and supporting univariate analyses

MALFF pattern classification

The main MGPC machine learning analysis was able to predict diagnosis of PTSD, PTSD + DS, and healthy individuals using subject mALFF maps with 91.63% balanced accuracy (p < 0.0001 during permutation testing). Here, the class accuracy for healthy individuals was 96.08%, for PTSD patients it was 89.02%, and for PTSD + DS patients it was 89.80%. Moreover, the predictive class value for healthy individuals was 87.50%, for PTSD patients it was 94.81%, and for PTSD + DS patients it was 89.90%. Interestingly, regions weighted as most important in classifying the three groups were the bilateral mid orbitofrontal cortex (BA 11), the bilateral dmPFC (BA 10, 9, 8, 6), the bilateral vmPFC/subgenual ACC (BA 11, 12, 25), and the bilateral superior parietal lobe (BA 5, 7) (see Fig. 1). The confusion matrix in Fig. 1 also shows a diagonal pattern of classification, representing optimal classification. An ideal confusion matrix is diagonal, where all predicted class labels would correspond to the truth. It is evident by this plot that no classes are sacrificed in order to gain accuracy in other classes. Indeed, regions-of-interest (ROIs) retained expected ranking positions, indicating stable ranking of the regions across folds. A critical concept here, however, is that all voxels within the mALFF maps will contribute to the decision function, since the analysis is multivariate. Contrary to common practice in Statistical Parametric Mapping (a mass-univariate approach), current implementation of PRoNTo software cannot isolate part of the pattern and instead reports only on the peaks of the distribution of the decision function's weight map. Finally, similar results were obtained using both an independent LOSOPG cross-validation procedure, and when implementing an independent binary GPC machine learning algorithm (see online Supplemental Results).

Fig. 1. The multiclass GPC machine learning analysis was able to predict diagnosis of PTSD, PTSD + DS, and healthy individuals, using subject mALFF activation maps with 91.63% balanced accuracy, p < 0.0001 during permutation testing. Predictive class value for healthy individuals was 87.50%, for PTSD patients it was 94.81%, and for PTSD + DS patients it was 89.90%. Regions that were weighted as most important in classifying the three groups were the bilateral mid-orbitofrontal cortex, the bilateral dmPFC, the bilateral vmPFC/subgenual ACC, and the bilateral superior parietal lobe. Top left corner: represents the confusion matrix, where an ideal confusion matrix is diagonal and all predicted class labels correspond to the truth. Here, no classes are sacrificed in order to gain accuracy in other classes. For the confusion matrix, group 1 = healthy controls, group 2 = PTSD patients, and group 3 = PTSD + DS patients. Numbers on top of the bars in graph correspond to the number of correctly classified individuals in each group.

Univariate group difference in mALFF resting-state activation

Mass-univariate analyses conducted within SPM for the main effect of group and relevant follow-up comparisons demonstrated intensity differences between groups in areas broadly corresponding to those with the highest weights in the MVPA analysis (Tables 2, 3 and Fig. 2). Here, the main effect of group revealed FDR cluster-corrected peaks within the left dmPFC/pre-central gyrus (BA 6), the left middle temporal gyrus (BA 37), bilateral globus pallidus, left amygdala, right post-central gyrus/superior parietal lobe (BA 4/7), right fusiform gyrus, right temporal pole, and right cerebellum (lobule V/VI). Follow-up comparisons revealed that the PTSD group had increased resting-state mALFF activation within the left post-/pre-central gyrus (BA 43), bilateral globus pallidus, and thalamus, the left putamen, the left amygdala/parahippocampal gyrus, and the right cerebellum (lobule V/VI), as compared with the PTSD + DS group (Tables 2, 3 and Fig. 2a). By contrast, as compared with the PTSD group, the PTSD + DS group displayed increased resting-state mALFF activation within the bilateral cerebellum (crus I/lobule VIIB), the right vmPFC, orbitofrontal cortex, and frontal pole (BA 10) (Tables 2, 3 and Fig. 2b). With regard to control group comparisons, the PTSD group displayed increased resting-state mALFF activation in the bilateral globus pallidus and thalamus, in addition to the right anterior insula, and right hippocampus/amygdala as compared with the healthy control group (Tables 2, 3 and Fig. 2c). The healthy control group showed increased activation within the bilateral superior parietal lobe (BA 5) as compared with the PTSD group (Tables 2, 3 and Fig. 2d). Finally, the PTSD + DS group evidenced increased resting-state mALFF activation in the right dmPFC (BA 9/10), and the right temporal pole (BA 38) as compared with the control group (Tables 2, 3 and Fig. 2e). Here, the control group showed increased activation in the bilateral pre-central gyrus motor strip and post-central gyrus somatosensory cortex (BA 4/6), as compared to the PTSD + DS group (Tables 2, 3 and Fig. 2f). When MDD diagnosis was used as a covariate, regional clusters identified via the main effect of group did not change, albeit magnitude of statistical significance decreased marginally. Critically, medication use was not found to significantly affect ANOVA results when used as a covariate.

Fig. 2. Whole-brain corrected follow-up t tests to the one-way ANOVA (FDR-cluster level p < 0.05, k = 20) main analysis for mALFF resting-state activation, comparing PTSD patients, PTSD + DS patients, and healthy controls, as indicated by contrast notation (< or >). PTSD, posttraumatic stress disorder, PTSD + DS, dissociative subtype posttraumatic stress disorder patients, FDR, false-discovery-rate cluster corrected, mALFF, mean amplitude of low-frequency fluctuations. Coordinates are given in MNI space and images were produced using MRIcron.

Table 2. mALFF one-way ANOVA whole-brain corrected

Whole-brain corrected (FDR-cluster level, p < 0.05, k = 20) one way ANOVA for the mALFF resting-state activation data.

dmPFC, dorsomedial prefrontal cortex; BA, Brodmann area; MNI, Montreal Neurological Institute; FDR, false-discovery-rate cluster corrected; H, hemisphere; mALFF, mean amplitude of low-frequency fluctuations.

Table 3. mALFF follow-up two-sample t tests, whole-brain corrected

Whole-brain corrected follow-up t tests to the one-way ANOVA (FDR-cluster level, p < 0.05, k = 20) mALFF resting-state activation data.

PTSD, posttraumatic stress disorder; PTSD + DS, dissociative subtype posttraumatic stress disorder patients; vmPFC, ventromedial prefrontal cortex; dmPFC, dorsomedial prefrontal cortex; BA, Brodmann area; MNI, Montreal Neurological Institute; FDR, false-discovery-rate cluster corrected; H, hemisphere; mALFF, mean amplitude of low-frequency fluctuations.

Amygdala complex functional connectivity pattern classification and univariate analysis

Our auxiliary MGPC machine learning analysis was able to predict diagnosis of PTSD, PTSD + DS, and healthy individuals using individual amygdala complex functional connectivity maps with 85.00% balanced accuracy (p < 0.0001 during permutation testing). The class accuracy for healthy individuals was 84.31%, for PTSD patients it was 85.37%, and for PTSD + DS patients it was 83.67%. Moreover, the predictive class value for healthy individuals was 93.48%, for PTSD patients it was 85.37%, and for PTSD + DS patients it was 75.93%.

Interestingly, regions weighted as most important in classifying the three groups were the bilateral orbitofrontal cortex (BA 11), the bilateral dmPFC and dlPFC (BA 46, 10, 9, 8, 6), the bilateral anterior cingulate cortex (ACC), the left superior parietal lobe, the bilateral supramarginal gyrus, and the bilateral vmPFC (BA 11, 12, 25). Similar results were obtained using LOSOPG cross-validation procedures and an independent binary GPC machine learning algorithm (see online Supplemental Results). In line with the auxiliary amygdala complex machine learning analysis, we observed a significant univariate amygdala complex × group interaction (FDR-corrected) within the left dlPFC, dmPFC, and orbitofrontal cortex (BA 10, 46, 9), the right precuneus, and the superior parietal lobe (BA 40) (see online Supplementary Table S2).

Clinical data and motion artifacts

As expected, we found that all three groups (PTSD, PTSD + DS, and controls) differed significantly in terms of CAPS (IV and 5), CTQ, BDI, MDI-total, MDI trait depersonalization/derealization average scores, and RSDI state dissociation scores; the PTSD + DS exhibited the highest scores among these clinical variables as compared with the PTSD and control groups (see online Supplemental Table S1, p < 0.05). Additionally, these scores were higher in the PTSD group as compared to the control group. Age did not differ significantly between the three groups. The relation between medication use frequency and group was non-significant. MDD diagnosis was significantly higher in the PTSD + DS group as compared to the PTSD group (p < 0.05), where some studies suggest that high scores of depression are associated with the PTSD + DS, in addition to higher PTSD severity scores (for review see Hansen et al., Reference Hansen, Ross and Armour2017). Furthermore, we found that trait (MDI depersonalization/derealization averages) and state (RSDI depersonalization/derealization averages) symptoms of dissociation were correlated significantly in our sample (r = 0.678, p < 0.001).

Discussion

Here, we present the first machine learning study to predict the diagnosis of PTSD, PTSD + DS, and healthy controls using two different modalities of neuroimaging data: (1) mALFF resting-state activation and (2) amygdala complex resting-state functional connectivity. Critically, high balanced accuracy and convergence was observed between both modalities, despite using different cross-validation procedures for the same MVPA protocol, and when conducting independent binary GPC MVPA protocols. These findings support the emerging validity of machine learning as an adjunctive diagnostic tool for use with neuroimaging resting-state data, and point toward the sensitivity of machine learning in classifying subtypes of PTSD.

Machine learning MVPA

In the current study, by analyzing mALFF maps of resting-state activation with MGPC machine learning computational methods, we were able to predict diagnosis of PTSD, PTSD + DS, and healthy individuals with 91.63% balanced accuracy. Critically, regions identified as being heavily weighted for group classification were also identified as significantly different between groups when conducting a whole-brain corrected mass-univariate analysis. Recently, a growing number of studies have applied machine learning methods to neuroimaging data to predict and classify psychiatric illness (Bleich-cohen et al., Reference Bleich-cohen, Jamshy, Sharon, Weizman, Intrator, Poyurovsky and Hendler2014; Mikolas et al., Reference Mikolas, Melicher, Skoch, Matejka, Slovakova, Bakstein, Hajek and Spaniel2016; Rive et al., Reference Rive, Redlich, Schmaal, Marquand, Dannlowski, Grotegerd, Veltman, Schene and Ruhé2016), including PTSD, typically based on the predictive value of current PTSD symptoms, cortisol levels, and pre-trauma risk factors (Gong et al., Reference Gong, Li, Du, Pettersson-Yeo, Crossley, Yang, Li, Huang and Mechelli2014; Karstoft et al., Reference Karstoft, Galatzer-levy, Statnikov, Li and Shalev2015; Liu et al., Reference Liu, Xie, Wang and Guo2015; Omurca and Ekinci, Reference Omurca and Ekinci2015; Galatzer-Levy et al., Reference Galatzer-Levy, Ma, Statnikov, Yehuda and Shalev2017; Gradus et al., Reference Gradus, King, Galatzer-levy and Street2017; Jin et al., Reference Jin, Jia, Lanka, Rangaprakash, Li, Liu, Hu and Deshpande2017; Saxe et al., Reference Saxe, Ma, Ren and Aliferis2017). Resting-state activation has also been utilized to predict PTSD symptom presentation through machine learning (Gong et al., Reference Gong, Li, Du, Pettersson-Yeo, Crossley, Yang, Li, Huang and Mechelli2014; Liu et al., Reference Liu, Xie, Wang and Guo2015), where functional connectivity maps differentiate PTSD patients from controls (Liu et al., Reference Liu, Xie, Wang and Guo2015; Jin et al., Reference Jin, Jia, Lanka, Rangaprakash, Li, Liu, Hu and Deshpande2017). Here, we provide a novel demonstration of the validity and sensitivity of machine learning analyses for use in classifying subtypes of PTSD across several machine learning algorithms and fMRI neuroimaging inputs, including both resting-state activation and amygdala complex functional connectivity.

Given that the amygdala plays a central role in the manifestation and maintenance of PTSD psychopathology (Etkin and Wager, Reference Etkin and Wager2007; Aupperle et al., Reference Aupperle, Melrose, Stein and Paulus2012; Yehuda et al., Reference Yehuda, Hoge, McFarlane, Vermetten, Lanius, Nievergelt, Hobfoll, Koenen, Neylan and Hyman2015; Shalev et al., Reference Shalev, Liberzon and Marmar2017), with altered functional connectivity to multiple cortical and subcortical structures in PTSD and PTSD + DS (Sripada et al., Reference Sripada, King, Garfinkel, Wang, Sripada, Welsh and Liberzon2012; Brown et al., Reference Brown, LaBar, Haswell, Gold, Beall, Van Voorhees, Marx, Calhoun, Fairbank, Green, Tupler, Weiner, Beckham, Brancu, Hoerle, Pender, Kudler, Swinkels, Nieuwsma, Runnals, Youssef, McDonald, Davison, Yoash-Gantz, Taber, Hurley, McCarthy and Morey2014; Nicholson et al., Reference Nicholson, Densmore, Frewen, Théberge, Neufeld, McKinnon and Lanius2015, Reference Nicholson, Rabellino, Densmore, Frewen, Paret, Kluetsch, Schmahl, Théberge, Neufeld, McKinnon, Reiss, Jetly and Lanius2016a, Reference Nicholson, Sapru, Densmore, Frewen, Neufeld, Theberge, McKinnon and Lanius2016b, Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017), we also investigated the application of machine leaning on neuroimaging data inputs in the form of amygdala complex functional connectivity. Interestingly, our MGPC analysis was able to predict diagnosis of PTSD, PTSD + DS, and healthy individuals using individual amygdala complex functional connectivity maps with 85.00% balanced accuracy.

Univariate results

Providing substantial validity for this novel protocol, both our multivariate machine learning analyses and our mass-univariate analyses within SPM yielded highly converging results. Our main analysis of mALFF resting-state activation revealed that the following regions were ranked as most important in classifying groups: the bilateral mid orbitofrontal cortex, dmPFC, vmPFC/subgenual ACC, and the superior parietal lobe (it is noteworthy that all regions within the brain are considered part of the decision function for MVPA).

Our mass-univariate analyses revealed specifically that the PTSD + DS group evidenced greater activation within the right vmPFC, orbitofrontal cortex, and bilateral cerebellum crus I/lobule VIIB as compared with PTSD patients. Moreover, the PTSD + DS group also displayed increased activation within the right dmPFC and temporal pole as compared with healthy individuals. Taken together, these findings are in line with the emotion overmodulation model, which posits that PTSD + DS is characterized by exacerbated top-down prefrontal regulation over the limbic system, with associated emotional detachment (Hopper et al., Reference Hopper, Frewen, van der Kolk and Lanius2007; Lanius et al., Reference Lanius, Vermetten, Loewenstein, Brand, Christian, Bremner and Spiegel2010; Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017; Melara et al., Reference Melara, Ruglass, Fertuck and Hien2018). Notably, whereas the vmPFC and dmPFC are involved in executive fear and emotion regulation (Mobbs et al., Reference Mobbs, Marchant, Hassabis, Seymour, Tan, Gray, Petrovic, Dolan and Frith2009, Reference Mobbs, Yu, Rowe, Eich, FeldmanHall and Dalgleish2010; Etkin et al., Reference Etkin, Büchel and Gross2015), the vmPFC displays top-down effective connectivity in PTSD + DS to the amygdala and midbrain (Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017). Indeed, emotion overmodulation associated with dissociative symptoms likely results from the collective integration of top-down emotion regulation signals from multiple areas of the PFC and ACC, thereby blunting bottom-up defense/fear generation from the amygdala and midbrain (Lanius et al., Reference Lanius, Vermetten, Loewenstein, Brand, Christian, Bremner and Spiegel2010; Åhs et al., Reference Åhs, Kragel, Zielinski, Brady and LaBar2015; Etkin et al., Reference Etkin, Büchel and Gross2015; Ledoux, Reference Ledoux2016; McKinnon et al., Reference McKinnon, Boyd, Frewen, Lanius, Jetly, Richardson and Lanius2016; Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017; pitman, Reference Pitman, Rasmusson, Koenen, Shin, Orr, Gilbertson, Milad and Liberzon2012). Finally, the orbitofrontal cortex, which displayed increased activation in the PTSD + DS group, is thought to serve as a dynamic interface site between consciousness and emotion processing (Tamietto and De Gelder, Reference Tamietto and De Gelder2010). Interestingly, the function of this region aligns closely with PTSD + DS symptoms of altered states of consciousness, depersonalization, and derealization, which may emerge due to aberrant emotion/consciousness processing within the orbitofrontal cortex as a result of trauma (Tamietto and De Gelder, Reference Tamietto and De Gelder2010; APA, 2013). Moreover, the orbitofrontal cortex has direct modulatory connections with the amygdala, where the amygdala is thought to be overmodulated in PTSD + DS (Hopper et al., Reference Hopper, Frewen, van der Kolk and Lanius2007; Lanius et al., Reference Lanius, Vermetten, Loewenstein, Brand, Christian, Bremner and Spiegel2010; Mickleborough et al., Reference Mickleborough, Daniels, Coupland, Kao, Williamson, Lanius, Hegadoren, Schore, Densmore, Stevens and Lanius2011; Reinders et al., Reference Reinders, Willemsen, den Boer, Vos, Veltman and Loewenstein2014; Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017; Melara et al., Reference Melara, Ruglass, Fertuck and Hien2018) and is known to be implicated heavily in PTSD symptom presentation (Ghashghaei et al., Reference Ghashghaei, Hilgetag and Barbas2007; Yehuda et al., Reference Yehuda, Hoge, McFarlane, Vermetten, Lanius, Nievergelt, Hobfoll, Koenen, Neylan and Hyman2015; Aghajani et al., Reference Aghajani, Veer, van Hoof, Rombouts, van der Wee and Vermeiren2016).

The observed group differences with regard to the cerebellum are particularly noteworthy when comparing results from the PTSD and PTSD + DS groups. Here, the PTSD + DS displayed increased crus I and lobule VIIB activation as compared with the PTSD group, where these regions have been identified as key executive functioning areas linked with the PFC in cerebro-cerebellar loops (Stoodley and Schmahmann, Reference Stoodley and Schmahmann2009). Interestingly, this in contrast to the increased cerebellar activation observed within lobules V/VI in the PTSD group as compared to PTSD + DS, which are primarily motor and somatosensory areas within the cerebellum (Stoodley and Schmahmann, Reference Stoodley and Schmahmann2009). This neuroanatomical differentiation may reflect the unique symptoms profiles of emotion under- and overmodulation observed in PTSD and PTSD + DS, respectively.

By contrast, the PTSD group showed higher activation within the bilateral globus pallidus, left post-/pre-central gyrus, amygdala, parahippocampus, and the right cerebellar lobule V/VI as compared with the PTSD + DS group. PTSD patients also displayed increased activation in the bilateral globus pallidus, right anterior insula, and right amygdala/hippocampus as compared with control participants. Under the emotion-modulation model, PTSD patients are characterized by increased amygdala, insula, and limbic/midbrain activation, as well as somatosensory/motor readiness and exacerbated defense circuit priming, which together are associated with observed hypervigilance, hyperarousal, and active defensive responding in PTSD patients as compared with PTSD + DS and controls (Lanius et al., Reference Lanius, Vermetten, Loewenstein, Brand, Christian, Bremner and Spiegel2010, Reference Lanius, Rabellino, Boyd, Harricharan, Frewen and McKinnon2017; Mickleborough et al., Reference Mickleborough, Daniels, Coupland, Kao, Williamson, Lanius, Hegadoren, Schore, Densmore, Stevens and Lanius2011; Hayes et al., Reference Hayes, Hayes and Mikedis2012; Stevens et al., Reference Stevens, Jovanovic, Fani, Ely, Glover, Bradley and Ressler2013; Reinders et al., Reference Reinders, Willemsen, den Boer, Vos, Veltman and Loewenstein2014; Sadeh et al., Reference Sadeh, Spielberg, Warren, Miller and Heller2014; Harricharan et al., Reference Harricharan, Rabellino, Frewen, Densmore, Theberge, McKinnon, Schore and Lanius2016; Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017; Rabellino et al., Reference Rabellino, Densmore, Harricharan, Jean, McKinnon and Lanius2017; Olivé et al., Reference Olivé, Densmore, Harricharan, Théberge, McKinnon and Lanius2018). Previous research indicates that whereas activation in the globus pallidus and thalamus relate directly to emotionally arousing stimuli and signaling of motor planning, globus pallidus activation is also related to emotion-response learning (Frank et al., Reference Frank, Dewitt, Hudgens-Haney, Schaeffer, Ball, Schwarz, Hussein, Smart and Sabatinelli2014) and displays increased connectivity to the amygdala in PTSD (Stevens et al., Reference Stevens, Jovanovic, Fani, Ely, Glover, Bradley and Ressler2013). Furthermore, we found previously that PTSD and PTSD + DS patients displayed aberrant connectivity between insula subregions and the amygdala, which correlated positively with dissociative symptoms and is hypothesized to be related to a disruption of normal interoceptive processing (Nicholson et al., Reference Nicholson, Sapru, Densmore, Frewen, Neufeld, Theberge, McKinnon and Lanius2016b), where altered insula activation has also been found between PTSD patients and controls (Etkin and Wager, Reference Etkin and Wager2007; Aupperle et al., Reference Aupperle, Melrose, Stein and Paulus2012; Pitman et al., Reference Pitman, Rasmusson, Koenen, Shin, Orr, Gilbertson, Milad and Liberzon2012; Mazza et al., Reference Mazza, Tempesta, Pino, Catalucci, Gallucci and Ferrara2013). On balance, the current findings of increased amygdala, insula, and pre-/post-central as well as cerebellar motor and somatosensory areas, likely correspond to increased innate alarm system activation (fight-or-flight), chronic interoceptive monitoring, and reinforced limbic priming, related to hypervigilance and motor readiness as a result of attenuated PFC inhibition in PTSD (Hopper et al., Reference Hopper, Frewen, van der Kolk and Lanius2007; Weston, Reference Weston2014; Kozlowska et al., Reference Kozlowska, Walker, McLean and Carrive2015; Yehuda et al., Reference Yehuda, Hoge, McFarlane, Vermetten, Lanius, Nievergelt, Hobfoll, Koenen, Neylan and Hyman2015; Nicholson et al., Reference Nicholson, Sapru, Densmore, Frewen, Neufeld, Theberge, McKinnon and Lanius2016b, Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017; Lanius et al., Reference Lanius, Rabellino, Boyd, Harricharan, Frewen and McKinnon2017; Melara et al., Reference Melara, Ruglass, Fertuck and Hien2018).

Finally, the healthy control group displayed increased activation within the superior parietal lobe as compared to the PTSD group, as well as increased pre-/post-central gyrus activation as compared to the PTSD + DS group. The superior parietal lobe is a heteromodal association cortex that integrates sensory information to construct awareness and is also involved in exerting top-down control of attention to enhance sensory representation (Wolbers et al., Reference Wolbers, Weiller and Büchel2003; Behrmann et al., Reference Behrmann, Geng and Shomstein2004; Wang et al., Reference Wang, Yang, Fan, Xu, Li, Liu, Fox, Eickhoff, Yu and Jiang2015). Notably, reduced connectivity to the superior parietal lobe has been reported among PTSD patients as compared to controls and PTSD + DS patients (Nicholson et al., Reference Nicholson, Densmore, Frewen, Théberge, Neufeld, McKinnon and Lanius2015; DiGangi et al., Reference DiGangi, Tadayyon, Fitzgerald, Rabinak, Kennedy, Klumpp, Rauch and Phan2016), which may be related to altered sensory/awareness processing disrupted by hyperactive defense circuits in PTSD. Moreover, with regard to PTSD + DS control group comparisons, increased activation in pre-/post-central gyrus motor cortex and somatosensory cortex within controls may parallel passive defense responses in PTSD + DS patients related to symptoms of emotional detachment (Kozlowska et al., Reference Kozlowska, Walker, McLean and Carrive2015; Harricharan et al., Reference Harricharan, Rabellino, Frewen, Densmore, Theberge, McKinnon, Schore and Lanius2016; McKinnon et al., Reference McKinnon, Boyd, Frewen, Lanius, Jetly, Richardson and Lanius2016; Lanius et al., Reference Lanius, Rabellino, Boyd, Harricharan, Frewen and McKinnon2017; Nicholson et al., Reference Nicholson, Friston, Zeidman, Harricharan, Mckinnon, Densmore, Neufeld, Th, Jetly, Spiegel and Lanius2017).

Several limitations of the current study are worth noting. First, although the currently implemented machine learning algorithms accounted for group size imbalance, future studies would benefit from increased sample sizes that are more equivalent, as well as with a longitudinal experimental design examining the effects of medication. Interestingly, the PTSD + DS group was characterized by higher MDD comorbidity and increased PTSD symptom severity. As higher PTSD severity and depression scores are associated with the PTSD + DS (for review see: Hansen et al., Reference Hansen, Ross and Armour2017), future investigation of these two variables is warranted with regard to predicting psychiatric diagnosis with machine learning tools.

Conclusion

We present the first study to utilize machine learning computational analyses to predict accurately subtypes of PTSD, including PTSD and PTSD + DS from healthy controls. This was accomplished using both spontaneous resting-state activation measures (mALFF) and amygdala complex functional connectivity measures. Moreover, areas weighted as most important for classifying the three groups also displayed significant group differences at the univariate level. Here, the PTSD + DS group displayed increased activation within emotion regulation regions as compared to the PTSD group, which showed increased activation within the amygdala, globus pallidus, and motor/somatosensory regions. In order to optimize care, the field of psychiatry would benefit significantly from the early identification of objective biomarkers that could predict heterogeneous subtypes of illness as well as individual patient symptom trajectories. The results of the current study have significant implications for advancing machine learning applications within the field of psychiatry, as well as for developing objective biomarkers indicative of diagnostic heterogeneity.

Supplementary material

The supplementary material for this article can be found at https://doi.org/10.1017/S0033291718002866

Acknowledgements

Our research group would like to acknowledge the following funding sources in which we are extremely grateful for their support: the Canadian Institutes of Health Research (CIHR), General Dynamics Land Systems, and the Canadian Institute for Veteran Health Research (CIMVHR).

Conflict of interest

All authors declare no financial interests or potential conflicts of interest with regard to the current study.

References

Aghajani, M, Veer, IM, van Hoof, MJ, Rombouts, SARB, van der Wee, NJ and Vermeiren, RRJM (2016) Abnormal functional architecture of amygdala-centered networks in adolescent posttraumatic stress disorder. Human Brain Mapping 25, S637–S638.Google Scholar
Åhs, F, Kragel, PA, Zielinski, DJ, Brady, R and LaBar, KS (2015) Medial prefrontal pathways for the contextual regulation of extinguished fear in humans. NeuroImage 122, 262–271.Google Scholar
APA (2013) Diagnostic and Statistical Manual of Mental Disorders. American Journal of Psychiatry, 5th Edn. VAP Publishing. Washington DC: Arlington, VA: American Psychiatric Publishing.Google Scholar
Aupperle, RL, Melrose, AJ, Stein, MB and Paulus, MP (2012) Executive function and PTSD: disengaging from trauma. Neuropharmacology 62, 686–694.Google Scholar
Beck, AT, Guth, D, Steer, RA and Ball, R (1997) Screening for major depression disorders in medical inpatients with the beck depression inventory for primary care. Behaviour Research and Therapy 35, 785–791.Google Scholar
Behrmann, M, Geng, JJ and Shomstein, S (2004) Parietal cortex and attention. Current Opinion in Neurobiology 14, 212–217.Google Scholar
Bernstein, DP, Stein, JA, Newcomb, MD, Walker, E, Pogge, D, Ahluvalia, T, Stokes, J, Handelsman, L, Medrano, M, Desmond, D and Zule, W (2003) Development and validation of a brief screening version of the Childhood Trauma Questionnaire. Child Abuse and Neglect 27, 169–190.Google Scholar
Blake, DD, Weathers, FW, Nagy, LM, Kaloupek, DG, Gusman, FD, Charney, DS and Keane, TM (1995) The development of a Clinician-Administered PTSD scale. Journal of Traumatic Stress 8, 75–90.Google Scholar
Bleich-cohen, M, Jamshy, S, Sharon, H, Weizman, R, Intrator, N, Poyurovsky, M and Hendler, T (2014) Machine learning fMRI classifier delineates subgroups of schizophrenia patients. Schizophrenia Research 160, 196–200.Google Scholar
Briere, J (2002) Multiscale Dissociation Inventory Professional Manual. Odessa, Florida: Psychological Assessment Resources.Google Scholar
Brown, VM, LaBar, KS, Haswell, CC, Gold, AL, Beall, SK, Van Voorhees, E, Marx, CE, Calhoun, PS, Fairbank, JA, Green, KT, Tupler, LA, Weiner, RD, Beckham, JC, Brancu, M, Hoerle, JM, Pender, M, Kudler, H, Swinkels, CM, Nieuwsma, JA, Runnals, JJ, Youssef, NA, McDonald, SD, Davison, R, Yoash-Gantz, R, Taber, KH, Hurley, R, McCarthy, G and Morey, RA (2014) Altered resting-state functional connectivity of basolateral and centromedial amygdala complexes in posttraumatic stress disorder. Neuropsychopharmacology 39, 361–369.Google Scholar
Davatzikos, C, Shen, D, Gur, RC, Wu, X, Liu, D, Fan, Y, Hughett, P, Turetsky, BI and Gur, RE (2005) Whole-brain morphometric study of schizophrenia revealing a spatially complex set of focal abnormalities. Archives of General Psychiatry 62, 1218–1227.Google Scholar
DiGangi, JA, Tadayyon, A, Fitzgerald, DA, Rabinak, CA, Kennedy, A, Klumpp, H, Rauch, SAM and Phan, KL (2016) Reduced default mode network connectivity following combat trauma. Neuroscience Letters 615, 37–43.Google Scholar
Duerden, EG, Arsalidou, M, Lee, M and Taylor, MJ (2013) Lateralization of affective processing in the insula. NeuroImage 78, 159–175.Google Scholar
Eklund, A, Nichols, TE and Knutsson, H (2016) Cluster failure: why fMRI inferences for spatial extent have inflated false-positive rates. Proceedings of the National Academy of Sciences 113, 201602413.Google Scholar
Etkin, A and Wager, TD (2007) Functional neuroimaging of anxiety: a meta-analysis of emotional processing in PTSD, social anxiety disorder, and specific phobia. The American Journal of Psychiatry 164, 1476–1488.Google Scholar
Etkin, A, Büchel, C and Gross, JJ (2015) The neural bases of emotion regulation. Nature Reviews Neuroscience 16, 693–700.Google Scholar
Felmingham, K, Kemp, AH, Williams, L, Falconer, E, Olivieri, G, Peduto, A and Bryant, R (2008) Dissociative responses to conscious and non-conscious fear impact underlying brain function in post-traumatic stress disorder. Psychological Medicine 38, 1771–1780.Google Scholar
First, MB, Spitzer, RL, Gibbon, M and Williams, JBW (2002) Structured Clinical Interview for DSM-IV-TR Axis I Disorders, Patient Edition (SCID-I/P, 11/2002 Revision) for DSMIV. New York: Biometrics Research, New York State Psychiatric Institute.Google Scholar
Frank, DW, Dewitt, M, Hudgens-Haney, M, Schaeffer, DJ, Ball, BH, Schwarz, NF, Hussein, AA, Smart, LM and Sabatinelli, D (2014) Emotion regulation: quantitative meta-analysis of functional activation and deactivation. Neuroscience and Biobehavioral Reviews 45, 202–211.Google Scholar
Frewen, Pa and Lanius, Ra (2006) Toward a psychobiology of posttraumatic self-dysregulation: reexperiencing, hyperarousal, dissociation, and emotional numbing. Annals of the New York Academy of Sciences 1071, 110–124.Google Scholar
Fu, CHY and Costafreda, SG (2013) Neuroimaging-based biomarkers in psychiatry: clinical opportunities of a paradigm shift. Canadian Journal of Psychiatry 58, 499–508.Google Scholar
Galatzer-Levy, IR, Ma, S, Statnikov, A, Yehuda, R and Shalev, AY (2017) Utilization of machine learning for prediction of post-traumatic stress: a re-examination of cortisol in the prediction and pathways to non-remitting PTSD. Translational Psychiatry 7, e1070; doi:10.1038/tp.2017.38.Google Scholar
Ghashghaei, HT, Hilgetag, CC and Barbas, H (2007) Sequence of information processing for emotions based on the anatomic dialogue between prefrontal cortex and amygdala. NeuroImage 34, 905–923.Google Scholar
Gong, Q, Li, L, Du, M, Pettersson-Yeo, W, Crossley, N, Yang, X, Li, J, Huang, X and Mechelli, A (2014) Quantitative prediction of individual psychopathology in trauma survivors using resting-state fMRI. Neuropsychopharmacology 39, 681–687.Google Scholar
Gradus, JL, King, MW, Galatzer-levy, I and Street, AE (2017) Gender differences in machine learning models of trauma and suicidal ideation in veterans of the Iraq and Afghanistan wars. Journal of Traumatic Stress 30, 362–371.Google Scholar
Hansen, M, Ross, J and Armour, C (2017) Evidence of the dissociative PTSD subtype: a systematic literature review of latent class and profile analytic studies of PTSD. Journal of Affective Disorders 213, 59–69.Google Scholar
Harricharan, S, Rabellino, D, Frewen, P, Densmore, M, Theberge, J, McKinnon, M, Schore, A and Lanius, R (2016) fMRI functional connectivity of the periaqueductal gray in PTSD and its dissociative subtype. Human Brain Mapping 6, e00579. doi: 10.1002/brb3.579.Google Scholar
Harricharan, S, Nicholson, AA, Densmore, M, Théberge, J, McKinnon, MC, Neufeld, RWJ and Lanius, RA (2017) Sensory overload and imbalance: resting-state vestibular connectivity in PTSD and its dissociative subtype. Neuropsychologia 106, 169–178.Google Scholar
Hayes, JP, Hayes, SM and Mikedis, AM (2012) Quantitative meta-analysis of neural activity in posttraumatic stress disorder. Biology of Mood & Anxiety Disorders 2, 9.Google Scholar
Hopper, JW, Frewen, PA, van der Kolk, BA and Lanius, RA (2007) Neural correlates of reexperiencing, avoidance, and dissociation in PTSD: symptom dimensions and emotion dysregulation in responses to script-driven trauma imagery. Journal of Traumatic Stress 20, 713–725.Google Scholar
Jin, C, Jia, H, Lanka, P, Rangaprakash, D, Li, L, Liu, T, Hu, X and Deshpande, G (2017) Dynamic brain connectivity is a better predictor of PTSD than static connectivity. Human Brain Mapping 38, 4479–4496.Google Scholar
Karstoft, K, Galatzer-levy, IR, Statnikov, A, Li, Z and Shalev, AY (2015) Bridging a translational gap: using machine learning to improve the prediction of PTSD. BMC Psychiatry 15, 30.Google Scholar
Kozlowska, K, Walker, P, McLean, L and Carrive, P (2015) Fear and the defense cascade. Harvard Review of Psychiatry 5, 1.Google Scholar
Lanius, Ra, Williamson, PC, Bluhm, RL, Densmore, M, Boksman, K, Neufeld, RWJ, Gati, JS and Menon, RS (2005) Functional connectivity of dissociative responses in posttraumatic stress disorder: a functional magnetic resonance imaging investigation. Biological Psychiatry 57, 873–884.Google Scholar
Lanius, Ra, Vermetten, E, Loewenstein, RJ, Brand, B, Christian, S, Bremner, JD and Spiegel, D (2010) Emotion modulation in PTSD: clinical and neurobiological evidence for a dissociative subtype. American Journal of Psychiatry 167, 640–647.Google Scholar
Lanius, RA, Rabellino, D, Boyd, JE, Harricharan, S, Frewen, PA and McKinnon, MC (2017) The innate alarm system in PTSD: conscious and subconscious processing of threat. Current Opinion in Psychology 14, 109–115.Google Scholar
Ledoux, JE (2016) Using neuroscience to help understand fear and anxiety: a two-system framework. International Journal of Climatology 173, 1083–1093.Google Scholar
Liu, F, Xie, B, Wang, Y and Guo, W (2015) Characterization of post-traumatic stress disorder using resting-state fMRI with a multi-level parametric classification approach. Brain Topography 28, 221–237.Google Scholar
Mazza, M, Tempesta, D, Pino, MC, Catalucci, A, Gallucci, M and Ferrara, M (2013) Regional cerebral changes and functional connectivity during the observation of negative emotional stimuli in subjects with post-traumatic stress disorder. European Archives of Psychiatry and Clinical Neuroscience 263, 575–583.Google Scholar
McKinnon, MC, Boyd, JE, Frewen, PA, Lanius, UF, Jetly, R, Richardson, JD and Lanius, RA (2016) A review of the relation between dissociation, memory, executive functioning and social cognition in military members and civilians with neuropsychiatric conditions. Neuropsychologia 90, 210–234.Google Scholar
Melara, RD, Ruglass, LM, Fertuck, EA and Hien, DA (2018) Regulation of threat in post-traumatic stress disorder: associations between inhibitory control and dissociative symptoms. Biological Psychology 133, 89–98.Google Scholar
Mickleborough, MJS, Daniels, JK, Coupland, NJ, Kao, R, Williamson, PC, Lanius, UF, Hegadoren, K, Schore, A, Densmore, M, Stevens, T and Lanius, Ra (2011) Effects of trauma-related cues on pain processing in posttraumatic stress disorder: an fMRI investigation. Journal of Psychiatry and Neuroscience 36, 6–14.Google Scholar
Mikolas, P, Melicher, T, Skoch, A, Matejka, M, Slovakova, A, Bakstein, E, Hajek, T and Spaniel, F (2016) Connectivity of the anterior insula differentiates participants with first-episode schizophrenia spectrum disorders from controls: a machine-learning study. Psychological Medicine 46, 2695–2704.Google Scholar
Mobbs, D, Marchant, JL, Hassabis, D, Seymour, B, Tan, G, Gray, M, Petrovic, P, Dolan, RJ and Frith, CD (2009) From threat to fear: the neural organization of defensive fear systems in humans. Journal of Neuroscience 29, 12236–12243.Google Scholar
Mobbs, D, Yu, R, Rowe, JB, Eich, H, FeldmanHall, O and Dalgleish, T (2010) Neural activity associated with monitoring the oscillating threat value of a tarantula. Proceedings of the National Academy of Sciences of the United States of America 107, 20582–20586.Google Scholar
Nicholson, A, Densmore, M, Frewen, PA, Théberge, J, Neufeld, RWJ, McKinnon, MC and Lanius, RA (2015) The dissociative subtype of posttraumatic stress disorder: unique resting-state functional connectivity of basolateral and centromedial amygdala complexes. Neuropsychopharmacology 40, 2317–2326.Google Scholar
Nicholson, AA, Rabellino, D, Densmore, M, Frewen, PA, Paret, C, Kluetsch, R, Schmahl, C, Théberge, J, Neufeld, RWJ, McKinnon, MC, Reiss, J, Jetly, R and Lanius, RA (2016 a) The neurobiology of emotion regulation in posttraumatic stress disorder: amygdala downregulation via real-time fMRI neurofeedback. Human Brain Mapping 38, 541–560.Google Scholar
Nicholson, AA, Sapru, I, Densmore, M, Frewen, PA, Neufeld, RWJ, Theberge, J, McKinnon, MC and Lanius, RA (2016 b) Unique insula subregion resting-state functional connectivity with amygdala complexes in posttraumatic stress disorder and its dissociative subtype. Psychiatry Research: Neuroimaging 250, 61–72.Google Scholar
Nicholson, AA, Friston, KJ, Zeidman, P, Harricharan, S, Mckinnon, MC, Densmore, M, Neufeld, RWJ, Th, J, Jetly, R, Spiegel, D and Lanius, RA (2017) Dynamic causal modeling in PTSD and Its dissociative subtype: bottom-up versus top-down processing within fear and emotion regulation circuitry. Human Brain Mapping 38, 5551–5561.Google Scholar
Olivé, I, Densmore, M, Harricharan, S, Théberge, J, McKinnon, MC and Lanius, R (2018) Superior colliculus resting state networks in post-traumatic stress disorder and its dissociative subtype. Human Brain Mapping 39, 563–574.Google Scholar
Omurca, SI and Ekinci, E (2015) An alternative evaluation of post traumatic stress disorder with machine learning methods. 2015 International Symposium on Innovations in Intelligent SysTems and Applications (INISTA), Madrid, Spain, pp. 1–7.Google Scholar
Orrù, G, Pettersson-Yeo, W, Marquand, AF, Sartori, G and Mechelli, A (2012) Using support vector machine to identify imaging biomarkers of neurological and psychiatric disease: a critical review. Neuroscience and Biobehavioral Reviews 36, 1140–1152.Google Scholar
Pitman, RK, Rasmusson, AM, Koenen, KC, Shin, LM, Orr, SP, Gilbertson, MW, Milad, MR and Liberzon, I (2012) Biological studies of post-traumatic stress disorder. Nature Reviews Neuroscience 13, 769–787.Google Scholar
Rabellino, D, Densmore, M, Harricharan, S, Jean, T, McKinnon, MC and Lanius, RA (2017) Resting-state functional connectivity of the bed nucleus of the stria terminalis in post-traumatic stress disorder and its dissociative subtype. Human Brain Mapping 39, 1367–1379. doi: 10.1002/hbm.23925Google Scholar
Rasmussen, C and Willams, CKI (2006) Gaussian Processes for Machine Learning. Cambridge, MA: The MIT Press.Google Scholar
Reinders, AATS, Willemsen, ATM, den Boer, JA, Vos, HPJ, Veltman, DJ and Loewenstein, RJ (2014) Opposite brain emotion-regulation patterns in identity states of dissociative identity disorder: a PET study and neurobiological model. Psychiatry Research: Neuroimaging 223, 236–243.Google Scholar
Rive, MM, Redlich, R, Schmaal, L, Marquand, AF, Dannlowski, U, Grotegerd, D, Veltman, DJ, Schene, AH and Ruhé, HG (2016) Distinguishing medication-free subjects with unipolar disorder from subjects with bipolar disorder: state matters. Bipolar Disorders 18, 612–623.Google Scholar
Sadeh, N, Spielberg, JM, Warren, SL, Miller, GA and Heller, W (2014) Aberrant neural connectivity during emotional processing associated with posttraumatic stress. Clinical Psychological Science 2, 748–755.Google Scholar
Saxe, GN, Ma, S, Ren, J and Aliferis, C (2017) Machine learning methods to predict child posttraumatic stress: a proof of concept study. BMC Psychiatry 17, 223. https://doi.org/10.1186/s12888-017-1384-1.Google Scholar
Schrouff, J, Rosa, MJ, Rondina, JM, Marquand, AF, Chu, C, Ashburner, J, Phillips, C, Richiardi, J and Mourão-Miranda, J (2013) PRoNTo: pattern recognition for neuroimaging toolbox. Neuroinformatics 11, 319–337.Google Scholar
Seng, JS, Li, Y, Yang, JJ, King, AP, Low, LMK, Sperlich, M, Rowe, H, Lee, H, Muzik, M, Ford, JD and Liberzon, I (2017) Gestational and postnatal cortisol profiles of women with posttraumatic stress disorder and the dissociative subtype. Journal of Obstetric, Gynecologic & Neonatal Nursing 47, 12–22. doi: 10.1016/j.jogn.2017.10.008.Google Scholar
Shalev, A, Liberzon, I and Marmar, C (2017) Post-traumatic stress disorder. New England Journal of Medicine 376, 2459–2469.Google Scholar
Sripada, RK, King, AP, Garfinkel, SN, Wang, X, Sripada, CS, Welsh, RC and Liberzon, I (2012) Altered resting-state amygdala functional connectivity in men with posttraumatic stress disorder. Journal of Psychiatry and Neuroscience 37, 241–249.Google Scholar
Stevens, JS, Jovanovic, T, Fani, N, Ely, TD, Glover, EM, Bradley, B and Ressler, KJ (2013) Disrupted amygdala-prefrontal functional connectivity in civilian women with posttraumatic stress disorder. Journal of Psychiatric Research 47, 1469–1478.Google Scholar
Stoodley, CJ and Schmahmann, JD (2009) Functional topography in the human cerebellum: a meta-analysis of neuroimaging studies. NeuroImage 44, 489–501.Google Scholar
Tamietto, M and De Gelder, B (2010) Neural bases of the non-conscious perception of emotional signals. Nature Reviews Neuroscience 11, 697–709.Google Scholar
Wang, J, Yang, Y, Fan, L, Xu, J, Li, C, Liu, Y, Fox, PT, Eickhoff, SB, Yu, C and Jiang, T (2015) Convergent functional architecture of the superior parietal lobule unraveled with multimodal neuroimaging approaches. Human Brain Mapping 36, 238–257.Google Scholar
Wegrzyn, M, Riehle, M, Labudda, K, Woermann, F, Baumgartner, F, Pollmann, S, Bien, CG and Kissler, J (2015) Investigating the brain basis of facial expression perception using multi-voxel pattern analysis. Cortex 69, 131–140.Google Scholar
Weston, CSE (2014) Posttraumatic stress disorder: a theoretical model of the hyperarousal subtype. Frontiers in Psychiatry 5, 1–20.Google Scholar
Wolbers, T, Weiller, C and Büchel, C (2003) Contralateral coding of imagined body parts in the superior parietal lobe. Cerebral Cortex 13, 392–399.Google Scholar
Wolf, EJ, Rasmusson, AM, Mitchell, KS, Logue, MW, Baldwin, CT and Miller, MW (2014) A genome-wide association study of clinical symptoms of dissociation in a trauma-exposed sample. Depression and Anxiety 31, 352–360.Google Scholar
Wolfers, T, Buitelaar, JK, Beckmann, CF, Franke, B and Marquand, AF (2015) From estimating activation locality to predicting disorder: a review of pattern recognition for neuroimaging-based psychiatric diagnostics. Neuroscience and Biobehavioral Reviews 57, 328–349.Google Scholar
Yehuda, R, Hoge, CW, McFarlane, AC, Vermetten, E, Lanius, RA, Nievergelt, CM, Hobfoll, SE, Koenen, KC, Neylan, TC and Hyman, SE (2015) Post-traumatic stress disorder. Nature Reviews Disease Primers 1, 15057.Google Scholar
Figure 0

Table 1. Demographic and clinical information

Figure 1

Fig. 1. The multiclass GPC machine learning analysis was able to predict diagnosis of PTSD, PTSD + DS, and healthy individuals, using subject mALFF activation maps with 91.63% balanced accuracy, p < 0.0001 during permutation testing. Predictive class value for healthy individuals was 87.50%, for PTSD patients it was 94.81%, and for PTSD + DS patients it was 89.90%. Regions that were weighted as most important in classifying the three groups were the bilateral mid-orbitofrontal cortex, the bilateral dmPFC, the bilateral vmPFC/subgenual ACC, and the bilateral superior parietal lobe. Top left corner: represents the confusion matrix, where an ideal confusion matrix is diagonal and all predicted class labels correspond to the truth. Here, no classes are sacrificed in order to gain accuracy in other classes. For the confusion matrix, group 1 = healthy controls, group 2 = PTSD patients, and group 3 = PTSD + DS patients. Numbers on top of the bars in graph correspond to the number of correctly classified individuals in each group.

Figure 2

Fig. 2. Whole-brain corrected follow-up t tests to the one-way ANOVA (FDR-cluster level p < 0.05, k = 20) main analysis for mALFF resting-state activation, comparing PTSD patients, PTSD + DS patients, and healthy controls, as indicated by contrast notation (< or >). PTSD, posttraumatic stress disorder, PTSD + DS, dissociative subtype posttraumatic stress disorder patients, FDR, false-discovery-rate cluster corrected, mALFF, mean amplitude of low-frequency fluctuations. Coordinates are given in MNI space and images were produced using MRIcron.

Figure 3

Table 2. mALFF one-way ANOVA whole-brain corrected

Figure 4

Table 3. mALFF follow-up two-sample t tests, whole-brain corrected

Supplementary material: File

Nicholson et al. supplementary material

Nicholson et al. supplementary material 1

Download Nicholson et al. supplementary material(File)
File 26.5 KB