Prediction of Illness severity rating scores from structural brain scans in Major Depressive Disorder.

Background

Correlations between illness severity ratings and brain grey matter volume reductions have been reported in patients with Major Depressive Disorder (MDD)(Ebmeier, et al. 2006; Heinzel, et al. 2009). However, until recently it has not been possible to predict illness severity rating scores from individual structural brain scans. Recently, multivariate image analysis techniques, such as Support Vector Machines (SVM) (Vapnik 1998) have been explored in neuroimaging with the aim of categorical predictive classification of individual structural brain scans (e.g. Control vs Patient). A further extension of this multivariate approach is the prediction of continuous non-categorical measures such as illness severity ratings, using individual structural brain scans. Here, we use Relevance Vector Regression (RVR)(Bishop 2007; Tipping 2001) to predict clinical rating scores such as Beck Depression Inventory (BDI) and Hamilton Depression Rating Scale (HRSD) from structural MRI scans obtained from patients with MDD.

Methods

Thirty patients with MDD were recruited in two centres. All patients met DSM IV diagnostic criteria for unipolar Major Depression, moderate to severe without significant comorbidity. T₁weighted images were acquired using 1.5 Tesla scanners and self-rated (BDI-II) and clinician-rated (HRDS), syndrome-specific illness severity ratings obtained just before scanning. Using Statistical Parametric Mapping (SPM5) software (Friston, et al. 2007), images were segmented to obtain grey matter maps which involved normalisation to MNI stereotactic anatomic space, with modulation to control rescaling artifacts and smoothed with a 6mm full-width half maximum Gaussian kernel. Multiple linear regression was done using SPM5 with three covariates (Centre, BDI and HRSD) to test the null hypothesis of no difference between the two scanner-patient groups with Centre as the covariate of interest. The threshold of significance was defined as p<0.05 corrected at a whole brain level with a cluster extent threshold of 144 voxels and brain regions which differed between centres were masked out from the later analyses. Smoothed whole brain raw voxels were used as input feature vectors for RVR learning with a linear kernel. The result was a model able to predict an individual clinical score (e.g. BDI or HRSD) given an unknown T₁ weighted MRI scan, using the leave-one-out cross-validation method to test the trained RVR’s ability to generalise from unseen data.

Results

Individual subject’s BDI scores were predicted well (correlation between actual score and RVR predicted scores r=0.694, p<0.0001). However, HRSD scores were less well predicted (r=0.34, p=0.068). A weighted image of brain regions that significantly contributed to the successful prediction of BDI scores was derived. Specifically, these regions include the hippocampus, medial brain structures and the superior temporal gyrus.

Conclusions

These results suggest that T₁ weighted MRI scans contain adequate information about neurobiological changes in patients with MDD to allow predictions about illness severity. The results contribute to the understanding of MDD and may facilitate discovery of biomarkers for the diagnosis and clinical management of individual patients. To our knowledge, no previous work has reported prediction of MDD severity measures from individual scans.

References

Bishop C. 2007. Pattern Recognition and Machine Learning. London: Springer-Verlag.

Ebmeier K, Donaghey C, Steele J. (2006): Recent developments and current controversies in depression. The Lancent 367:153-167.

Friston K, Ashburner J, Kiebel S, Nichols T, Penny. W. 2007. Statistical Parametric Mapping: The Analysis of Functional Brain Images. London: Academic Press.

Heinzel A, Grimm S, Beck J, Schuepback D, Hell D, Boesiger P, Boeker H, Northoff G. (2009): Segregated neural representation of psychological and somatic-vegetative symptoms in severe major depression. Neuroscience Letters. 456:49-53.

Tipping M. (2001): Sparse Bayesian learning and the relevance vector machine. Journal of Machine Learning Research 1:211-244.

Vapnik V. 1998. Statistical Learning Theory. New York: John Wiley and Sons.