Machine learning fMRI classifier delineates subgroups of schizophrenia patients

Abstract

Background

The search for a validated neuroimaging-based brain marker in psychiatry has thus far been fraught with both clinical and methodological difficulties. The present study aimed to apply a novel data-driven machine-learning approach to functional Magnetic Resonance Imaging (fMRI) data obtained during a cognitive task in order to delineate the neural mechanisms involved in two schizophrenia subgroups: schizophrenia patients with and without Obsessive–Compulsive Disorder (OCD).

Methods

16 schizophrenia patients with OCD (“schizo-obsessive”), 17 pure schizophrenia patients, and 20 healthy controls underwent fMRI while performing a working memory task. A whole brain search for activation clusters of cognitive load was performed using a recently developed data-driven multi-voxel pattern analysis (MVPA) approach, termed Searchlight Based Feature Extraction (SBFE), and which yields a robust fMRI-based classifier.

Results

The SBFE successfully classified the two schizophrenia groups with 91% accuracy based on activations in the right intraparietal sulcus (r-IPS), which further correlated with reduced symptom severity among schizo-obsessive patients.

Conclusions

The results indicate that this novel SBFE approach can successfully delineate between symptom dimensions in the context of complex psychiatric morbidity.

Introduction

The search for neuroimaging based brain markers in psychiatry has faced two major methodological hurdles. The first is that a region-of-interest (ROI) approach requires a neuro-anatomical hypothesis, whose pathophysiological validity isn't always evident and which provides a limited scope of findings. The second is that in whole brain data-driven approach sensitivity is compromised by the large number of multiple comparisons (Hendler et al., 2014).

Further complicating this effort is the fact that current clinical diagnostic systems are category-based and not symptom-domain specific. Thus, schizophrenia and Obsessive–Compulsive Disorder (OCD) have traditionally been viewed as two distinct psychopathologies with differing underlying pathophysiologies (Kurokawa et al., 2000). Nevertheless, there is compelling evidence for the co-occurrence of obsessive–compulsive symptoms (OCS) in a substantial proportion of schizophrenia patients (roughly 25%) (Poyurovsky et al., 2012), termed “schizo-obsessive” when there is frank OCD. Schizo-obsessive patients appear to have a more deteriorative course, with poorer prognosis and treatment responses, as compared to their non-OCD schizophrenia counterparts (Poyurovsky et al., 2012). In terms of neuroimaging findings, while there is vast literature regarding the neural mechanisms (structural and functional) involved in schizophrenia and in OCD separately, there is scarce data regarding the functional neurobiological substrates involved in schizophrenia-OCD co-morbidity.

Hypofrontality, expressed as reduced neural activation in the dorsolateral prefrontal cortex (DLPFC) (Callicott, 2003), is the most replicated functional brain imaging finding in SCH patients. Therefore, in order to characterize some of the abnormalities in functional brain circuits in schizo-obsessive patients, we previously used fMRI during a working memory task, namely, the N-back task, in which schizophrenia (SCH) patients are known to display abnormal brain processing (Bleich-Cohen et al., 2014). Applying region-oriented analyses focusing on the right DLPFC and the right Caudate we demonstrated that schizo-obsessive patients exhibit remarkably similar brain activation patterns as their pure schizophrenia counterparts. Moreover, whole-brain analyses did not produce significant differences between the groups as well.

To further examine possible dissimilarities between these two schizophrenia groups in a data-driven manner, we chose to further apply a whole brain classification based on multi-voxel pattern analysis (MVPA) during the N-back task. MVPA approaches (Mur et al., 2009, Pereira et al., 2009, Orrù et al., 2012) utilize machine-learning algorithms that combine several different markers for classification and robust statistical analysis of the results. Current MVPA methods rely on either a small number of ROIs (Eger et al., 2007) or whole-brain activity (Mourão-Miranda et al., 2005, Fu et al., 2008). However, both approaches have major drawbacks: regional MVPA may achieve better classification accuracy than whole brain MVPA, but remains limited to hypothesis-driven assumptions, while whole brain MVPA may have difficulty achieving significant accuracy rates (Chu et al., 2012) and does not necessarily produce neuroscientifically interpretable results. To tackle this issue of neurological interpretability, some whole-brain studies use weights assigned by the classifier to the different markers in an attempt to produce an interpretation (Mourão-Miranda et al., 2005, Fu et al., 2008). However, this approach can lead to two equally important neural markers receiving substantially different weights, resulting in misleading interpretations (Lee et al., 2010). To overcome these limitations, we used a modified MVPA approach termed as Searchlight Based Feature Extraction (SBFE), which combines a data-driven ROI search with whole-brain classification (Jamshy et al., 2012). The goal of the present fMRI study was to evaluate whether SBFE can distinguish between these two schizophrenia subgroups and delineate the neurobiological substrates involved.