Automatic Classification of Brain Connectivity Matrices - a toolbox for supporting neuropsychiatric diagnosis
Ricardo Jorge Maximiano1, Tiago Constantino1,2,3, André Santos-Ribeiro1,4, and Hugo Alexandre Ferreira1

1Institute of Biophysics and Biomedical Engineering, Faculty of Sciences of the University of Lisbon, Lisbon, Portugal, 2Spitalzentrum Biel, Bienne, Switzerland, 3Lisbon School of Health Technology - ESTeSL, Lisbon, Portugal, 4Centre for Neuropsychopharmacology, Imperial College London, London, United Kingdom

Synopsis

In this work, a user-friendly toolbox that aims to classify automatically brain connectivity matrices is described. To test this tool, we used the Parkinson’s Progression Markers Initiative (PPMI) data which includes structural and functional Magnetic Resonance Imaging data of healthy subjects, patients with “scans without evidence for dopaminergic deficit” (SWEDD) and patients diagnosed with Parkinson’s Disease (PD). Using default parameters, this tool was able to achieve a maximum accuracy of 85.4% in classifying the 3 groups of subjects by selecting features that were related to the rostral middle frontal gyrus and splenium, which are in agreement with PD literature.

Purpose

Parkinson’s Progression Markers Initiative (PPMI) is a multi-center study that aims to identify the progression of clinical features, imaging and biologic biomarkers in Parkinson’s Disease (PD) patients and also in PD subtypes. One of the PPMI’s main objectives is then to identify what features are relevant for the differential diagnosis of patients with “scans without evidence of dopaminergic deficit” (SWEDD) and PD patients, as this nosological entity is still not well understood.1

Within this context, we propose the Automatic Classification of Brain Connectivity Matrices (ACBC) extension for the Multimodal Imaging Brain Connectivity Analysis (MIBCA) toolbox2, making it easier to study ML in Connectomics via a user-friendly interface.

In this work, PPMI data was explored using ACBC in order to identify features from Diffusion Tensor Imaging (DTI) based structural connectivity data that have the highest relevance for the distinction between healthy controls (HC), PD and SWEDD patients.

Methods

Data was gathered from the publicly available PPMI database at LONI (http://www.loni.usc.edu) which includes T1-weighted and DTI images of healthy controls (N=30), SWEDD patients (N=29) and patients diagnosed with PD (N=30), and they were used to test the ACBC toolbox.

Using the MIBCA toolbox (http://www.mibca.com), 89 DTI-based structural connectivity matrices (one for each subject) of 94x94 regions-of-interest (ROI) were computed from fiber tracts connecting each pair of ROIs (derived from Freesurfer atlases). For this study, it was decided to use the thresholded brain connectivity matrices, automatically generated by the MIBCA, hence binary undirected matrices.

Inputting the brain connectivity matrices and labels, ACBC (developed in Matlab and connecting different tools - Weka3 and Brain Connectivity Toolbox, BCT4 - in a single user-interface as shown in Fig. 1) is able to estimate automatically the type of brain connectivity matrix, allowing the extraction of up to 20 different metrics such as node degree, transitivity, and edge betweenness centrality for each ROI and observation (Fig. 2).

After feature extraction, a series of pre-processing steps that comprise normalization, removal of features with very low variance and very low Fisher score is performed, leading to an increased classification accuracy and speed. Users are further given the possibility of choosing what feature selection (FS) and classifiers methods ACBC should perform on the feature matrix previously pre-processed. When all available FS and classifiers are selected, 20 possible combinations of methods are obtained.

Finally, ACBC performs cross-validation automatically using pre-defined parameters to reduce the amount of required input from the user, resulting in a batch of HTML reports describing all steps and results in terms of different model performance statistics, such as accuracy, Kappa and area under curve. These results are automatically saved and can be easily shared amongst the community without requiring Matlab.

Results

Applying these data and selecting all BCT’s graph theory metrics, 31358 features were obtained. These data were pre-processed automatically, resulting in a selected and more compacted matrix with 2155 features.

Average accuracies (using cross-validation 10-fold) for distinguishing the 3 groups of subjects ranged from 85.4% using Gain-Ratio and Sequential Floating Forward Selection as FS methods and Naïve Bayes Updateable as classifier, to 40.5% without using any FS and using the J.48 classifier. The best three combinations of techniques can be seen in Fig. 3. A confusion matrix when using the best combination can be seen in Fig. 4.

As expected, FS improved accuracy. In addition, ACBC was able to pick the 34 most important features of the dataset, which were different measures (node degree, clustering coefficient, topological overlap and edge betweenness centrality) related to ROIs identified as: splenium of the corpus callosum, Broca's area, rostral middle middle frontal gyrus (rMFG), frontal pole and insula.

Discussion and Conclusion

The results were found to be in agreement with the literature, specifically regarding the selected features, as rMFG and frontal pole are parts of the dopaminergic mesocortical pathway known to be changed in PD.5 Additionally, the corpus callosum or its cortical connections have been associated with cognitive impairment in PD patients.6 With respect to insula, it has been proposed that this is substantially affected by alpha-synuclein deposition in PD.7 Concerning the Broca's area, few studies correlate it directly with PD, although it is known that language impairment is relatively common within those patients.8 In addition, PD patients with gait disorder may have changes in the Broca's area.9

Lastly, the fact that ACBC was able to extract the most important features to distinguish between HC, PD and SWEDD patients with high accuracy, suggests that this tool can be used to support diagnosis of neuropsychiatric disorders.

Acknowledgements

Data used in the preparation of this article were obtained from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/data). For up-to-date information on the study, visit www.ppmi-info.org.

PPMI—a public-private partnership—is funded by the Michael J. Fox Foundation for Parkinson’s Research and its funding partners, which include Abbvie, Avid, Biogen, Bristol-Myers Squibb, Covance, GE Healthcare, Genentech, GalaxoSmithKline, Lilly, Lundbeck, Merck, Meso Scale Discovery, Pfizer, Piramal, Roche, Servier and UCB—for a current list see http://www.ppmi-info.org/about-ppmi/who-we-are/study-sponsors/.

Research supported by the Edmond J. Safra Philanthropic Foundation, and by the Fundação para a Ciência e Tecnologia (FCT) and Ministério da Ciência e Educação (MCE) Portugal (PIDDAC) under grants UID/BIO/00645/2013 and PTDC/SAU-ENB/120718/2010.

References

1. Schwingenschuh P, Ruge D, Edwards M, et al. Distinguishing SWEDDs patients with asymmetric resting tremor from Parkinson’s disease: a clinical and electrophysiological study. Movement Disorders: Official Journal of the Movement Disorder Society. 2010; 25(5), 560–9.

2. Santos-Ribeiro A, Lacerda L, Ferreira H. Multimodal Imaging Brain Connectivity Analysis (MIBCA) toolbox. PeerJ. 2015; 3, e1078.

3. Hall M, Frank E, Holmes G, et al. The WEKA data mining software. ACM SIGKDD Explorations Newsletter. 2009; 11(1), 10.

4. Rubinov M, Sporns O. Complex network measures of brain connectivity: uses and interpretations. NeuroImage. 2010; 52(3), 1059–69.

5. Karagulle K, Lehericy S, Luciana M, et al. Altered diffusion in the frontal lobe in Parkinson disease. American Journal of Neuroradiology. 2008; 29(3), 501–5.

6. Wiltshire K, Concha L, Gee M, et al. Corpus callosum and cingulum tractography in Parkinson’s disease. The Canadian Journal of Neurological Sciences. 2010; 37(5), 595–600.

7. Christopher L, Koshimori Y, Lang A, et al. Uncovering the role of the insula in non-motor symptoms of Parkinson’s disease. Brain. 2014; 137(Pt 8), 2143–54.

8. Liu L, Luo X, Dy C, et al. Characteristics of language impairment in Parkinson’s disease and its influencing factors. Translational Neurodegeneration. 2015; 4(1), 2.

9. Albani G, Künig G, Soelch C, et al. The role of language areas in motor control dysfunction in Parkinson’s disease. Neurological Sciences: Official Journal of the Italian Neurological Society and of the Italian Society of Clinical Neurophysiology. 2001; 1(1), 43–4.

Figures

Main interface of ACBC (Automatic Classification of Brain Connectivity Matrices) Toolbox. A linear sequence of steps makes it easy-to-use even by someone with no experience in ML and/or Connectomics.

Feature Extraction interface. When inputting the brain connectivity matrices, ACBC will automatically detect the type of brain connectivity matrix and recall the interface with compatible metrics with that type of matrix.

Top three combinations of feature selection and classifiers methods that achieved the highest accuracies scores. AUC is calculated for each assumed positive class (GR - Gain Ratio; SFFS - Sequential Floating Forward Selection; AUC - Area Under Curve; HC - Healthy Controls; SWEDD - Scans without evidence for dopaminergic deficit; PD - Parkinson’s Disease).

Confusion matrix when using Gain Ratio (GR) and Sequential Floating Forward Selection (SFFS) as feature selection methods and Naïve-Bayes Updateable as classifier (Class 0: Healthy Control, Class 1: SWEDD, Class 2: Parkinson’s Disease).



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
0255