Statistical machine learning to identify traumatic brain injury (TBI) from structural disconnections of white matter networks

Highlights

• Method to identify diffuse axonal injury in mild traumatic brain injury (mTBI) patients from structural connectivity patterns.

• Network-based statistics (NBS) is used to find significant network differences in mTBI and controls.

• Fractional anisotropy (FA) features of the different structural connections obtained from NBS used as features.

• Random forest classifier discriminates between mTBI and controls based on FA features.

• Discriminative and significant network differences obtained from feature importance of random forest.

Abstract

Identifying diffuse axonal injury (DAI) in patients with traumatic brain injury (TBI) presenting with normal appearing radiological MRI presents a significant challenge. Neuroimaging methods such as diffusion MRI and probabilistic tractography, which probe the connectivity of neural networks, show significant promise. We present a machine learning approach to classify TBI participants primarily with mild traumatic brain injury (mTBI) based on altered structural connectivity patterns derived through the network based statistical analysis of structural connectomes generated from TBI and age-matched control groups. In this approach, higher order diffusion models were used to map white matter connections between 116 cortical and subcortical regions. Tracts between these regions were generated using probabilistic tracking and mean fractional anisotropy (FA) measures along these connections were encoded in the connectivity matrices. Network-based statistical analysis of the connectivity matrices was performed to identify the network differences between a representative subset of the two groups. The affected network connections provided the feature vectors for principal component analysis and subsequent classification by random forest. The validity of the approach was tested using data acquired from a total of 179 TBI patients and 146 controls participants. The analysis revealed altered connectivity within a number of intra- and inter-hemispheric white matter pathways associated with DAI, in consensus with existing literature. A mean classification accuracy of 68.16% ± 1.81% and mean sensitivity of 80.0% ± 2.36% were achieved in correctly classifying the TBI patients evaluated on the subset of the participants that was not used for the statistical analysis, in a 10-fold cross-validation framework. These results highlight the potential for statistical machine learning approaches applied to structural connectomes to identify patients with diffusive axonal injury.

Introduction

Traumatic brain injuries (TBIs) are highly heterogeneous, with injury resulting from a unique combination of mechanical forces that interact with each individuals distinctive neuroanatomy ( Meaney and Smith, 2011 ). They are also dynamic, involving a complex cascade of metabolic events that affect important ionic fluxes, neurotransmitter concentrations, cerebral haemodynamic status, edema and neuro-inflammatory responses ( Bigler and Maxwell, 2012, Taber and Hurley, 2013). Numerous studies indicate that a hallmark neuropathological feature of TBI is the presence of diffuse axonal injury (DAI) resulting from damage to axolemma and neuro-filaments within the brain ( Bigler and Maxwell, 2012, Omalu et al., 2005). DAI is estimated to occur in approximately 40%–50% of all persons who sustain a TBI ( Meythaler et al., 2001 ). The pathology of DAI has been investigated in some detail ( Povlishock and Katz, 2005 ). If there is sufficient force to rupture the micro-vasculature, small hemorrhages or microbleeds may also be present, which are reliably detected using CT or MRI based susceptibility-weighted imaging (SWI) ( Benson, 2012 ). Although CT and conventional MRI have been proven to be useful for the clinical management of moderate to severe TBI, conventional neuroimaging does not reliably detect DAI in cases of mild TBI (mTBI) ( Hammoud and Wasserman, 2002, Iverson et al., 2000).

There is now growing evidence to suggest that diffusion tensor imaging (DTI) has the potential to revolutionize how we detect DAI in mTBI ( Arfanakis et al., 2002, Bigler and Maxwell, 2012, Taber and Hurley, 2013, Niogi and Mukherjee, 2010, Shenton et al., 2012, Lipton et al., 2012, Goh et al., 2014, Yeh et al., 2012). As DTI probes the subtle changes in the anisotropic nature of the diffusion of water in white matter (WM), it potentially provides a more sensitive and quantifiable measure of DAI within functionally distinct brain regions, thereby overcoming one of the main limitations of conventional MRI ( Huisman et al., 2004 ). Fractional anisotropy (FA), a widely used quantitative measure of anisotropy, reflects the degree of alignment or organization of axonal pathways and has the potential to identify axonal injury associated with TBI ( Johnson et al., 2013 ). For example, using region of interest (ROI) analyses, significant reduction in FA values in patients with mTBI have been reported within the corpus callosum, internal capsule and centrum semiovale ( Kumar et al., 2009, Inglese et al., 2005, Miles et al., 2008, Matsushita et al., 2011). A common finding was a reduction in FA within the genu or splenium of the corpus callosum for mTBI patients. Voxel wise analyses employing Tract-Based Spatial Statistics (TBSS) have revealed more wide spread injury extending to multiple white matter (WM) pathways projecting through the corpus callosum ( Kinnunen et al., 2010, Wada et al., 2012). In contrast to ROI or voxel wise analyses, diffusion imaging along with probabilistic tractography allows DAI to be investigated within specific WM pathways associated with multiple neural circuits. This approach is appealing as loss in connectivity of neural networks can be readily computed and may potentially guide the planning of personalized treatment and rehabilitation strategies ( Wilde et al., 2006, Xu et al., 2007). Using this approach, number of studies have reported loss of integrity of WM pathways within the corpus callosum, fornix, internal capsule and the peduncular projections in TBI patients compared to normal controls ( Wang et al., 2008, Sugiyama et al., 2013).

In recent years, there has been a trend toward using connectome-based strategies to assess the impact of DAI on network connectivity. For example, Pandit et al. (2013) used a graph-theoretic approach based on functional connectivities to show deviation of small-world characteristics i.e. reduced overall connectivity, longer average path lengths and reduced network efficiency in TBI patients compared to control participants, especially within networks involving the posterior cingulate cortex. Their findings also supported previous findings of reduced WM integrity along the corpus callosum, corticospinal tracts and superior longitudinal fasciculi. Similarly, ( Caeyenberghs et al., 2013 ) combined complementary information from the graph-theoretic analysis of the structural and functional hubs to discriminate between the TBI and the healthy groups, with a specific focus on the switching-motor network. Furthermore, graph-theoretic approach using DTI networks was adopted by Irimia et al., 2012a, Irimia et al., 2012b to introduce a patient-tailored approach to the graphical representation of WM change over time. Van Horn et al. (2012) also employed DTI connectomics to study the white matter damage in the simulated case of severe TBI of Phineas Gage, who was an American railroad construction foreman in the 19^th century. The study revealed that the impact on measures of network connectivity between the area of direct injury and other areas were profound and widespread, which probably contributed to acute and long-term behavioral changes.

Machine learning approaches are being increasingly used to identify discriminative features from the neural connections that best separate a diseased group from healthy cohorts. For example, Aribisala et al. (2010) applied a region-wise support vector machine (SVM) analysis to identify DAI using T1, T2 maps and mean diffusivity measures in both gray and white matter regions in mTBI patients compared to control participants. Recently, Lui et al. (2014) used features from conventional MRI, diffusion MRI and fMRI to build a set of features and employed a feature selection algorithm prior to applying several classifiers like support vector machines (SVMs), Bayesian networks and multilayer perceptrons to separate mTBI from controls.

Many of the studies in TBI research were based on the use of specific regions known to be associated with the loss of WM integrity in DAI. Such approaches may not identify WM injury elsewhere in more remote brain regions other than the point of impact, which impair a wide array of cerebral functions ( Love and Webb, 1992, Gioia et al., 2010). Our limited knowledge of how tract structure relates to cognitive function in normal brain ( Kinnunen et al., 2010 ), leads to the importance of assessment of white matter structure after traumatic brain injury with a highly comprehensive spatial coverage.

Network-based statistical analysis between the TBI and control groups may reveal structural connections based on a stochastic model with the assumption that the mean FA values are different between the groups. These connections being statistically significant in differentiating the two groups, may not however, represent an optimally discriminative feature set in classifying the TBI and control groups due to the following reasons: 1) the exclusive dependence on a stochastic model is an overspecialization when the analysis is performed on a limited number of datasets ( Breiman, 2001b ). Such an analysis therefore may fail in generalization when validated on a larger cohort. 2) The validation of these statistical methods is usually by goodness-of-fit measures such as hypothesis testing used in network-based statistical analysis ( Zalesky et al., 2010 ). The number of connections obtained with such methods, therefore, may be sensitive to the choice of the threshold used in t-tests ( Zalesky et al., 2010 ) and it is often challenging to reach a consensus on how many of these are actually affected due to TBI.

In machine learning, standard feature selection algorithms may find out features or neural connections that are discriminative in classification, while these may not be statistically significant in terms of difference in mean FA values between the TBI and control groups. In other words, the machine learning algorithms do not make any assumption for a data model, instead, a complex, unknown function is learned from the data to predict the responses ( Breiman, 2001b ). The robust cross-validation schemes used in machine learning approaches, i.e. building a classification model on a training set and validating on a test set, where the sets are mutually exclusive, also provides more confidence and generalization to the analysis. Unleashing the potential of both statistical and machine learning methods, therefore, in this work, we have adopted a hybrid strategy combining both the statistical analysis of network connections and then machine learning in order to identify the cortical/sub-cortical connections that differentiate between TBI patients and age-matched controls.

In our method, we used probabilistic tractography to build structural networks and estimate diffusion anisotropy in a population of TBI (mostly with mTBI) and matched controls, with DAI being characterized by the reduction in diffusion anisotropy ( Arfanakis et al., 2002 ). Network-based statistics (NBS) ( Zalesky et al., 2010 ), which are based on generalized linear model (GLM) and t-test statistics, were then applied to extract networks or connections that were different between a subset of the two groups. Tractography data from MRI is very noisy, hence it is very likely that the noise is propagated into pairwise associations measuring FA values over the tracts, thus resulting in a low contrast-to-noise ratio. NBS is known to work better than other generic statistical tests due to its ability to deal with low contrast-to-noise ratio ( Zalesky et al., 2010 ). NBS also involves extensive multiple comparisons correction and reduces the family-wise error rate in a weak sense, which is computationally more expensive in other statistical tests. As discussed before, statistical network separation using t-tests is however, sensitive to thresholds. To avoid a precise choice of the t-threshold that may vary with datasets, we selected the network from NBS comprising of a large number of structural connections and a high spatial coverage, using a reasonably high t-threshold. This ensured that all connections obtained from the NBS network were statistically significant with reduced FA values in TBI than controls.

Retaining a large number of connections may however, present serious challenges to classification methods ( Liu and Motoda, 2007 ), i.e. the curse of dimensionality ( Hastie et al., 2001 ), when many of these obtained from NBS may not be actually predictive in classifying the TBI and healthy groups. With the presence of a large number of features, a classification or prediction model tends to overfit, resulting in performance degradation on unknown test data ( Tang et al., 2014 ). Selecting the most relevant features is therefore, usually suboptimal in building a good predictor ( Guyon and Elisseeff, 2003 ), especially if the features are correlated and redundant. Principal component analysis (PCA) ( Pearson, 1901 ), a common feature selection technique was therefore, applied on the connections obtained from NBS, in order to reduce and re-express the correlated feature space with a set of linearly uncorrelated components.

Random forest (RF) ( Breiman, 2001a ) is known to be robust against overfitting and on noisy data. A RF model was therefore, built using the features obtained from NBS and PCA, which was then used to classify the remaining sets of TBI and normal groups that were not used for feature selection using NBS and PCA. The primary advantage of using RF over any other classifier is its uniqueness in providing the feature importance (Gini importance) within the cross-validation scheme based on the training samples. Unlike other traditional feature selection methods, Gini importance is known to capture the multivariate and non-linear relationships among the structural connections and the disease states ( Langs et al., 2011 ) i.e. TBI or non-TBI in our case. Gini importance has been shown to correlate well with the measures based on feature perturbations ( Breiman, 2001a, Archer and Kimes, 2008) and known to be stable across datasets without a need for explicit regularization, thereby, providing an alternative to the computationally more expensive statistical permutation tests. All these factors led to a conscientious combination of NBS, PCA and RF classifier that was particularly suited to our problem of finding a comprehensive set discriminative connections with a good spatial coverage that were affected in TBI. The novelty of our method, hence, is in the preliminary selection of signature connections using NBS and then refining the signature connections according to the feature importance provided by the Gini measure of random forest.