Email

cnievergelt@ucsd.edu

To read the full PDF, click here

Individualized Diagnosis of PTSD Using Data-Driven Multimodal Approach

Xi Zhu, Ph.D.,1,2

Pavel Goldstein, Ph.D.,3 

Benjamin Suarez-Jimenez, Ph.D.,1,2

Tor D. Wager, Ph.D., 3

Yuval Neria, Ph.D.1,2

1) Columbia University, Department of Psychiatry, New York, NY

2) New York State Psychiatric Institute, New York, NY

3) University of Colorado Boulder, Department of Psychology and Neuroscience, CO

Xi Zhu, Ph.D.

The New York State Psychiatric Institute,

Columbia University Medical Center.

1051 Riverside Drive, New York, NY, 10032.

Tel: 646-774-8010

E-mail: xi.zhu@nyspi.columbia.edu

Yuval Neria, Ph.D.

The New York State Psychiatric Institute,

Columbia University Medical Center.

1051 Riverside Drive, New York, NY, 10032.

Tel: 646-774-8092; Fax: 646 774 8105;

E-mail: ny126@cumc.columbia.edu

Abstract

Post-traumatic stress disorder (PTSD) is a debilitating disorder observed in individuals exposed to traumatic events. While a growing number of neuroimaging studies have investigated the pathophysiology of PTSD, only few studies to date have apply multimodal imaging aiming to clarify neural markers of PTSD. Furthermore, neuroimaging research to date is mostly based on group level univariate analysis approaches, thus far not revealing an unbiased neural-based screening tool or instrument for reliable individualized diagnosis of PTSD, with many cases being undetected or misdiagnosed. The goal of this study is to advance PTSD research by using novel multimodal magnetic resonance imaging (MRI) features including structural MRI, resting state functional MRI (rs-fMRI) and diffusion tensor imaging (DTI), with machine learning methods, in classifying patients with PTSD from trauma-exposed healthy controls (TEHC) individuals using large-scale dataset from ENIGMA PTSD Working Group, and to examine its utility in predicting diagnosis, functional impairment, and symptom severity. This study will lay the groundwork to further utilize multimodal fMRI and machine-learning algorithms to quantify both functional and structural large-scale networks in PTSD, and offer new insights into the identification of potential biomarkers for the clinical diagnosis of PTSD.

Keywords: machine learning, PTSD, resting state functional connectivity, fMRI, volumatric, cortical thickness, DTI

Background and Goal: PTSD is a debilitating disorder observed in individuals exposed to traumatic events. It is characterized by intrusive symptoms, avoidance of trauma reminders, negative alterations in cognitions and mood, and heightened arousal (First, Spitzer, Gibbon, & Williams, 1996){First, 1996 #158}{First, 1996 #158}{First, 1996 #158}{First, 1996 #158}{First, 1996 #158}. {First, 1996 #158}{First, 1996 #158}{First, 1996 #158}{First, 1996 #158}{First, 1996 #158}While a growing number of neuroimaging studies have investigated the pathophysiology of PTSD using single MRI modality, to provide complementary information for disease diagnosis, and to improve classification accuracy, a few studies started to use the application of multimodal imaging in classifying neural markers of PTSD from controls (Li et al., 2014; Zhang et al., 2016). Furthermore, neuroimaging research to date is mostly based on group level univariate analysis approaches, thus far not revealing an unbiased neural-based screening tool or instrument for reliable individualized diagnosis of PTSD, with many cases being undetected or misdiagnosed (Sumpter & McMillan, 2005). However, recent significant advances in computational power and machine learning have shown great promise in their ability to classify psychiatric disorder at an individual level. In recent years, the number of pattern recognition studies in translational neuroimaging has grown dramatically (Woo, Chang, Lindquist, & Wager, 2017). 75% of these studies using multivariate predictive models in neuroimaging have focused on diagnosis, identifying brain signatures that discriminate patients from healthy controls. The aim of these studies is not just replacing existing diagnostic tool but rather to establish objective and meaningful neuro-biomarkers of disease of interest, thus supporting the development of new clinical measures and therapeutics. The majority of diagnosis studies still focused on Alzheimer (42%), psychosis (17%) and depression (12%). Only 1.6% of these studies started examining PTSD. Up to date, only 5 machine learning studies were found in PTSD (Table 1). Three out of the 5 studies used a single modality (Gong et al., 2014; Im et al., 2017; Liu et al., 2015), and others combined data from two MRI modalities (Li et al., 2014; Zhang et al., 2016). The accuracy for classifying patients with PTSD from controls varies from 80% to 93%. The number of study and number of patient with PTSD are far behind other disorders.

Table 1: studies using machine learning to classify PTSD patients from controls.

 PTSDTEHCHCModalityMethodsMLAccuracy
Gong 2014505040StruGMVSVM91%
Li 20144451DTI, RSRS dynamic changesFDDL80%
Liu 20152020RSALFF, seed-basedSVM92.50%
Zhang 2016172020Stru, RSGMV, ALFF, ReHoSVM89%
Im 20173029Stru, DTIGMVLRAUC=0.73

Despite increasing interest in devel­oping clinically useful biomarkers based on multiple neuroimaging modalities using machine learning, many challenges still remain to be solved. First, most of the studies have small sample sizes. Classification models based on small sample sizes tend to have better performance than the studies using larger samples. This may due to over fitting or curse of dimensionality, and is typically taken as an indicator of bias in the meta-analysis literature. Second, none of the PTSD studies evaluated a classifier on a totally independent cohort, which is the ideal way of examining the generalizability and robustness of the classifier. Third, we cannot perform a direct comparison of classification performance across these studies. The variation of the findings may come from the different study populations, different image analysis pipelines or the choice of the machine learning methods, which makes a direct comparison of classification performance difficult. Thus, testing the predicting model in a large sample and to test the model in one or more independent cohort is needed.

These large-scale studies have been made possible by the development of research consortia committee such as Psychiatric Genomic Consortium (PGC)-Enhancing Neuroimaging Genetics Through Meta-Analysis (ENIGMA), in which investigators shared their data across the world. These consortiums greatly promote model development on large samples, which can increase statistical power; development of models based on multisite samples, which are more likely to generalize across scanners and commonly encountered variations in study procedures; and tests on independent data sets with different characteristics.

Our overarching goal of this project is to overcome those challenges by utilizing multimodal brain features including structural magnetic resonance imaging (MRI), resting state functional MRI (rs-fMRI) and diffusion tensor imaging (DTI), with machine-learning, in classifying patients with PTSD from trauma-exposed healthy controls (TEHC) individuals using large-scale dataset from ENIGMA PTSD Working Group (PTSD N~=1500; TEHC N~=1500), and to examine its utility in predicting diagnosis, functional impairment, and symptom severity.

Specific Aims: The specific aims are as following: Specific Aim 1: To use multimodal MRI, including rs-fMRI, structural MRI, and DTI separately, and to quantify how well each type of modality distinguishes PTSD patients from TEHC, and identify the significant features in each modality. Specific Aim 2: To test the generalizability of the predictive models, in addition to the within sample leave-one-out cross-validation, we will test the predictive model in an independent dataset. Specific Aim 3: To test whether combined different MRI modalities will improve the predicting power above and beyond single modality alone. We will implement a multimodal predictive model to investigate whether the combination of the different imaging modalities (MRI, rs-fMRI and DTI) would improve prediction or whether these imaging modalities contain redundant information. Specific Aim 4: To use the machine-learning approach to examine the utility of the multimodal signature in predicting functional impairment and symptom severity among the patients with PTSD. Exploratory aim 1: Totest whether sex plays a role in the predictive model. Recent PTSD literatures showed females and males may have different brain signatures in PTSD patients (Helpman et al., 2017; Shvil et al., 2014), so we will test whether sex plays a role in the predictive models. More specifically, we will build two classifiers, one for male and one for female, and compare the classification results across different sex. Exploratory aim 2: To use machine-learning techniques to characterize the similarities and differences between patients with PTSD and patients with comorbid PTSD and major depressive disorder (PTSD-MDD) in order to better understand each’s pathology. This work will provide insight into the brain mechanisms underlying PTSD and MDD. We will develop new machine learning algorithms that combine fMRI, demographic and clinical information to better predict PTSD and PTSD-MDD at individual level.

Important innovative aspects to this proposal: this is the first study aiming to identify and quantify the multimodal neural signature of PTSD, and it is the first to utilize machine learning in advancing an individualized diagnosis of PTSD in a large-scale dataset. This study will lay the groundwork to further utilize multimodal fMRI and machine-learning algorithms to quantify both functional and structural large-scale networks, and offer new insights into the identification of potential biomarkers for the clinical diagnosis of PTSD.

Preliminary Work: Utilizing imaging data from NIMH RO1 MH072833 (PI, Yuval Neria PhD), we have developed multimodal imaging quality control and processing pipelines for structural MRI, rs-fMRI and DTI. We also applied Dr. Zhu’s previous expertise in the realm of pattern recognition to design the machine-learning algorithms differentiating PTSD patients from trauma-exposed healthy controls (TEHC). In the first step, analyzing data from 53 PTSD patients and 36 TEHCs, network level neural signatures of PTSD were probed using rsFC. PTSD was found to be associated with decreased connectivity of basolateral amygdala (BLA)-orbitalfrontal cortex (OFC) and centromedial amygdala (CMA)-thalamus (THA) pathways, key to fear processing and fear expression, respectively (Figure 1 and 2) (Zhu et al., 2016). In the second step, a machine-learning approach was applied to classify PTSD from TEHC at an individual level. Two levels of measures were extracted as classification features: 1) gray matter volume (GMV) from structural MRI, and 2) temporal seed-based resting state FC matrix from rs-fMRI. Then, random forest was applied for pattern classification at individual level. The performance of the classifier was evaluated using the leave-one-out cross-validation method. Preliminary results showed that rsFC features were able to differentiate patients with PTSD from TEHCs at an 80.0% accuracy level and the GMV features achieved a 68% accuracy rate. The regions and networks providing the strongest contribution to this classification were consistent with the group-level results. Our preliminary data demonstrated the feasibility and utility of using multimodal MRI methods to identify neural signatures, and as a potential diagnostic aid for PTSD.

Figure 1: presents the decreased BLA-OFC, CMA-THA connectivity in PTSD, when compared with TEHC.

Figure 2: presents the mean resting state functional connectivity in fear network in PTSD and TEHC.

*BLA: basolateral amygdala; CMA: centromedial amygdala; OFC: orbitalfrontal cortex; THA: thalamus

Methods, Subjects and Settings:

We will analyze neuroimaging and clinical data from ~3000 subjects (~1500 PTSD patients, and ~1500 TEHC subjects) contributed by over 10 cohorts, representing the largest neuroimaging study of PTSD to date. We will exclude participants with 1) history of Axis I psychiatric diagnosis, e.g., psychotic disorder, bipolar disorder, tic disorder, or eating disorder (comorbid current major depressive disorder will be allowed). 2) Depression which is antecedent to PTSD; score of > 25 on the Hamilton Rating Scale for Depression (HAM-D-17-item); significant depression and /or depression related impairment that is judged to warrant pharmacotherapy or combined medication and psychotherapy. 3) For PTSD, history of substance/alcohol dependence within the past six months, and abuse within past two months History. 4) Patients who are receiving effective medication for their PTSD, and/or depression. Antipsychotic, antidepressant, or mood stabilizer medications in the last 4 weeks prior to the study (6 weeks for fluoxetine). Standing daily dosing of benzodiazepine class of medication in the 2 weeks prior to the study (as needed use of benzodiazepines is not an exclusion, but must be clinically judged to tolerate no benzodiazepines for the 72-hour period before each of the fMRI days). Triptan anti-migraine medications. Other medications that may interfere with fear circuitry and fear memory such as blood-brain-barrierpenetrating β-blockers.

Neurocognitive Battery: CAPS will be used to assess PTSD symptom. Hamilton Depression Scale (HAMD, 17 item) will be used to assess depressive severity. Also, the SF-36, a 36-item measure of generic health status (Ware, 1996), will be collected across sites to assess functional impairment of patient with PTSD or TEHC.

MRI Data Analysis: A standardized image-analysis and quality-control pipeline established by the ENIGMA consortium will be used.

Structural MRI: Cortical thickness and volume values for anatomical regions will be obtained for each participant by processing the T1 images using Freesurfer 5.1.0 (http://surfer.nmr.mgh.harvard.edu) in conjuncation with standardized ENIGMA protocols. We will assess the ROI-based cortical thickness and volumes of eight subcortical structures (nucleus accumbens, amygdala subregions, caudate, hippocampus subregions, pallidum, putamen, thalamus, and lateral ventricle).

Resting state fMRI: Seed-based analysis will be carried out following the ENIGMA resting-state preprocessing pipeline using AFNI software. ROI-to-ROI seed-based functional connectivity analyses will be carried out using the CONN-fMRI Functional Connectivity toolbox v13. Between ROI connectivity analysis was performed based on ROIs defined in the FSL Harvard-Oxford Atlas maximum likelihood cortical and subcortical atlases.

Independent Component Analysis (ICA): We will apply probabilistic independent component analysis (PICA) by using the Multivariate Exploratory Linear Decomposition into Independent Components (MELODIC) toolbox of the FMRIB Software Library (FSL) package. A temporal concatenation tool in MELODIC will be used to derive group level components across all subjects. The independent component (IC) maps associated with motion or which were localized primarily in the white matter or CSF will be excluded from further study. The between-subject analysis was carried out using dual regression, a regression technique that back-reconstructs each group level component map at the individual subject level. Statistic inferences were then tested using FSL’s randomize permutation testing tool.

DTI: FMRIB’s Diffusion Toolbox (FDT) from FMRIB’s Software Library (FSL) (http://www.fmrib.ox.ac.uk/fsl/) will be applied to process DTI Data in order to assess white matter integrity, such as Fractional Anisotropy (FA), Mean Diffusivity (MD), Axial Diffusivity (AD), and Radial Diffusivity (RD).

Machine learning: We will obtain the features from different perspectives, such as volumetric, cortical thickness, resting state correlation matrix, ICA network features, and DTI features.

In order to effectively combine different feature categories, the multi-kernel combination strategy will be used (Liu, Wee, Chen, & Shen, 2014). Classification methods such as Random forest, SVM, Logistic regression will be performed using scikit-learn library in Python. Both filter-based and wrapper-based approaches will be used for feature elimination. The leave-one-out cross validation (LOOCV) strategy will be used to validate the performance of our proposed approach. Accuracy, sensitivity and specificity are defined based on the prediction results of LOOCV to quantify the performance of all approaches.

References

First, M. B., Spitzer, R. L., Gibbon, M., & Williams, J. B. (1996). Structured Clinical Interview for the DSM-IV Axis I Disorders. Washington, D.C.: American Psychiatric Press, Inc.,.

Gong, Q., Li, L., Tognin, S., Wu, Q., Pettersson-Yeo, W., Lui, S., . . . Mechelli, A. (2014). Using structural neuroanatomy to identify trauma survivors with and without post-traumatic stress disorder at the individual level. Psychol Med, 44(1), 195-203. doi:10.1017/S0033291713000561

Helpman, L., Zhu, X., Suarez-Jimenez, B., Lazarov, A., Monk, C., & Neria, Y. (2017). Sex Differences in Trauma-Related Psychopathology: a Critical Review of Neuroimaging Literature (2014-2017). Curr Psychiatry Rep, 19(12), 104. doi:10.1007/s11920-017-0854-y

Im, J. J., Kim, B., Hwang, J., Kim, J. E., Kim, J. Y., Rhie, S. J., . . . Yoon, S. (2017). Diagnostic potential of multimodal neuroimaging in posttraumatic stress disorder. PLoS One, 12(5), e0177847. doi:10.1371/journal.pone.0177847

Li, X., Zhu, D., Jiang, X., Jin, C., Zhang, X., Guo, L., . . . Liu, T. (2014). Dynamic functional connectomics signatures for characterization and differentiation of PTSD patients. Hum Brain Mapp, 35(4), 1761-1778. doi:10.1002/hbm.22290

Liu, F., Wee, C. Y., Chen, H., & Shen, D. (2014). Inter-modality relationship constrained multi-modality multi-task feature selection for Alzheimer’s Disease and mild cognitive impairment identification. Neuroimage, 84, 466-475. doi:10.1016/j.neuroimage.2013.09.015

Liu, F., Xie, B., Wang, Y., Guo, W., Fouche, J. P., Long, Z., . . . Chen, H. (2015). Characterization of post-traumatic stress disorder using resting-state fMRI with a multi-level parametric classification approach. Brain Topogr, 28(2), 221-237. doi:10.1007/s10548-014-0386-2

Shvil, E., Sullivan, G. M., Schafer, S., Markowitz, J. C., Campeas, M., Wager, T. D., . . . Neria, Y. (2014). Sex differences in extinction recall in posttraumatic stress disorder: a pilot fMRI study. Neurobiol Learn Mem, 113, 101-108. doi:10.1016/j.nlm.2014.02.003

Sumpter, R. E., & McMillan, T. M. (2005). Misdiagnosis of post-traumatic stress disorder following severe traumatic brain injury. Br J Psychiatry, 186, 423-426. doi:10.1192/bjp.186.5.423

Ware, J. (1996). The MOS 36-Item Short Form Health Survey (SF-36). In: Sederer LI, Dickey B, eds. Outcomes assessment in clinical practice.: Baltimore: Williams & Wilkins.

Woo, C. W., Chang, L. J., Lindquist, M. A., & Wager, T. D. (2017). Building better biomarkers: brain models in translational neuroimaging. Nat Neurosci, 20(3), 365-377. doi:10.1038/nn.4478

Zhang, Q., Wu, Q., Zhu, H., He, L., Huang, H., Zhang, J., & Zhang, W. (2016). Multimodal MRI-Based Classification of Trauma Survivors with and without Post-Traumatic Stress Disorder. Front Neurosci, 10, 292. doi:10.3389/fnins.2016.00292

Zhu, X., Helpman, L., Papini, S., Schneier, F., Markowitz, J. C., Van Meter, P. E., . . . Neria, Y. (2016). Altered resting state functional connectivity of fear and reward circuitry in comorbid PTSD and major depression. Depress Anxiety. doi:10.1002/da.22594

Recommended Articles

Leave A Comment

Your email address will not be published.