Classification of CSE, SGGS IE & T, Nanded Prof.

Classification of Schizophrenia
Avinash M. Hanwate
M.Tech Student,
Dept. of CSE,
SGGS IE & T, Nanded
Prof. B. R. Bombade
Associate Professor
Dept. of CSE,
SGGS IE & T, Nanded
Abstract—The goal of the MLSP 2014 Classification Challenge
was to automatically detect subjects with schizophrenia and
schizoaffective disorder based on multimodal features derived
from the magnetic resonance imaging (MRI) data. The patients
with age range of 18-65 years were diagnosed according to
DSM-IV criteria. The training data consisted of 46 patients and
40 healthy controls. The test set included 119748 subjects with
unknown labels. In the present solution, we implemented so-called
feature trimming, consisting of: 1) introducing a random vector
into the feature set, 2) calculating feature importance based on
mean decrease of the Gini-index derived by running Random
Forest classification, and 3) removing the features with importance
below the “dummy variable”. Support Vector Machine with
Gaussian Kernel was used to run final classification with reduced
feature set achieving test set AUC of 0.923. Lorem ipsum dolor
sit amet, consectetuer adipiscing elit. Etiam lobortis facilisis sem.
Nullam nec mi et neque pharetra sollicitudin. Praesent imperdiet
mi nec ante. Donec ullamcorper, felis non sodales commodo,
lectus velit ultrices augue, a dignissim nibh lectus placerat pede.
Vivamus nunc nunc, molestie ut, ultricies vel, semper in, velit.
Ut porttitor. Praesent in sapien. Lorem ipsum dolor sit amet,
consectetuer adipiscing elit. Duis fringilla tristique neque. Sed
interdum libero ut metus. Pellentesque placerat. Nam rutrum
augue a leo. Morbi sed elit sit amet ante lobortis sollicitudin.
Praesent blandit blandit mauris. Praesent lectus tellus, aliquet
aliquam, luctus a, egestas a, turpis. Mauris lacinia lorem sit amet
ipsum. Nunc quis urna dictum turpis accumsan semper.
Keywords—Schizophrenia, MRI, Random Forest, Support Vector
Machines, Feature Trimming
Schizophrenia is a characterised by abnormal social behavior
and failure to grasp reality. Common symptoms embrace
false beliefs, unclear or confused thinking, hearing voices,
reduced social engagement and emotional expression, and an
absence of motivation. individuals usually have extra psychological
state issues like anxiety disorders, major depressive
health problem or substance use disorder. Symptoms usually
come back on step by step, begin in young, adulthood, and last
an extended time. The reason behind schizophrenic disorder is
believed to be a mixture of environmental and genetic factors.
doable environmental factors embrace bound infections,
parental age, and poor nutrition throughout maternity. Males
are a lot of usually affected than females. regarding 200th of
individuals do well and a couple of recover fully if this illness
found early.
Magnetic resonance imaging (MRI) is a best technique of examining
the physical body. MRI is employed to elucidate and
characterize the neural design of the human brain. MRI scanner
employs a magnetic field and radio waves to generate complete
pictures of the human brain. MRI is a medical imaging techniques
accustomed investigate the general structure of healthy
and schizophrenic brain. MRI information is most relevant
within the studies of a head, specifically, for trailing disease of
brain and alternative brain connected issues. Our objective is
to diagnose schizophrenic disorder using multimodal features
of MRI scans. Now-a-days, there aren’t any automated tools
developed for identification of schizophrenic disorder. in this
paper, we tend to surveyed several papers to induce the best
technique or formula. we are combining the features of MRI
that’s functional MRI and Structural MRI to induce the most
effective method.
The whole sample enclosed the coaching set (46 patients
with schizophrenic disorder and schizoaffective disorder and
40 healthy controls HCs), and therefore the testing data
(119748 subjects with unknown labels). The patients were
diagnosed in step with DSM-IV criteria for schizophrenic disorder
and schizoaffective disorder throughout structured interview.
Structural magnetic resonance images (sMRI) and resting
state functional MRI (rs-fMRI) data were acquired on a 3T
MRI scanner at the Mind research Network (Albuquerque,New
Mexico). Image preprocessing was performed using statistical
parametric Mapping software, version 5 (SPM5):
( additional feature extraction
was done using independent part analysis, as enforced within
the GIFT toolbox (, yielding
source-based morphometric (SBM) loadings and functional
Network connectivity (FNC) features for sMRI and rs-fMRI,
correspondingly. For acquisition and preprocessing details of
structural and functional imaging data, together with feature
extraction protocol. For this competition, SBM and FNC
features were already obtainable for the participants.
Feature Trimming At the first step, after feature concatenation
we performed the procedure, which will be further
called feature trimming. Its steps are straightforward and
in very simple words can be described as: 1) introducing
a random vector into the feature set, 2) calculating feature
importance based on mean decrease of the Gini-index derived
by running Random Forest (RF) classification, and 3) removing
the features with importance below the “dummy variable”.
More detailed description follows below. The Random Forest
algorithm is formally defined as a collection of tree-structured
f(x, ?k), k = 1, 2, …., K, (1)
where ?k are random i.i.d. vectors (independent and identically
distributed) 11 and each tree provides a vote for the
most popular label at input x. For classification problems, the
forest prediction output is the majority vote. The algorithm
converges with large number of trees. Here we are given a set
of training data D = (v
, ci
i=1,2,…,n (to be defined below).
The classification task is then to learn a general mapping from
previously unseen test data to their corresponding diagnostic
labels, i.e. c : Rd ? C more specifically, adopting the notation
in 13, let v = (x1, …., xd) ? Rd denote the input data feature
vector (predictor), and let c ? C denote the output diagnostic
label (response). In our case, xi
is a measure (SBM or FNC
feature) derived from the ICA analysis briefly mentioned above
(d = number of features; e.g., 41 for volumetric data), and
C =

Sch, HC
. The RF algorithm incorporate a collection
of binary classification trees indexed by t = 1, …, ntree.
Each classification tree is characterised by its input root node,
internal split nodes, and its leaf terminal nodes containing class
In this setting, the RF algorithm can be briefly described as
follows: (i) Draw ntree samples from the original data D, using
random sampling with replacement; (ii) For each bootstrap
sample, grow a classification tree such that for each node:
randomly sample mtry of the predictor variables and chose the
best split according to the Gini criterion defined below from
among those feature variables (1 < mtry  d). The largest tree possible is grown and is not pruned. Using only mtry of the predictor variables selected at random is in contrast to standard tree classification (CART), where each node is split using the best split among all d variables; (iii) the forest consists of ntree trees. Each tree gives a classification for a given data point. Predict new data point x by putting x down each of the ntree trees and make a majority vote for classification across the forest. For a given tree, let S0 denote the set of input predictor data vectors that is fed into the root node, Sj be the subset of data points reaching node j in the binary split tree, and  S L j , SR j denote the subset of data points that reaches the left and right child, respectively, of node j, where S L j ? S R j = Sj and S L j ? S R j = ?. In the off-line tree training, each split node j is associated with a parameter vector ?j that is trained by optimizing an objective function I (defined below), i.e.: ?j = argmax??T I(Sj , ?) (2) In this notation, a binary-valued test function h(v, ?j ) : Rd × T ?  0, 1 is applied at each split node j. Here, 0 and 1 denote false and true, respectively, and the data point v arriving at split node j is sent to its left (0) or right (1) child node, accordingly. T is the set of all possible split function parameters, and Tj ? T is the subset of parameters available at node j. We thus have S L j (Sj , ?) =  (v, c) ? Sj | h(v, ?) = 0 and S R j (Sj , ?) =  (v, c) ? Sj | h(v, ?) = 1 . The objective function used is the Gini index, i.e.: I = i(? ) = 1 ? ?c?C p 2 j , (3) measuring the likelihood that a data point would be incorrectly labeled if it was randomly classified according to the distribution of class labels within the node. The optimal binary split is then the one that maximises the improvement (mean decrease) in the Gini index, which was used as a metric in our approach. To be more specific, at every split node ? one of the mtry variables, say Xk , is used to form the split and there is a resulting decrease in the Gini index. The mean decrease of the Gini index, ?i(? ) was used as a metric, i.e.: ?i(? ) = i(? ) ? (pLi(? L ) + pRi(? R)) (4) where I = i(? ) = 1 ? ?c?C p 2 j is the Gini index at node ? , pL = |S L j | |Sj | and pR = |S R j | |Sj | are the probabilities of sending a data point to the left and the right node, respectively. This metric reflects the contribution of a variable xk to the node homogeneity of ? . Thus, a higher mean decrease of the Gini index for a particular feature means that the variable is present more often in nodes with higher purity among all trees in the forest (overall). The sum of all decreases in the forest due to a given variable xk, normalized by the number of trees, therefore gives an estimate of its Gini importance i.e.: IG(xk) = 1 ntree nXtree t=1 X ? ?ixk (?, t) (5) Therefore, the Gini importance IG(xk) indicates how frequent the particular feature xk was selected in a split node, and how large its overall discriminative value was for the classification task. Finally, if you introduce a random dummy feature and calculate its mean decrease of the Gini index, you can then exclude (trim) everything with importance below this value. Final Model: The reduced feature set was then used to run final classification employing Support Vector Machine (SVM) with Gaussian Kernel 14. The optimization problem: max(X N i=1 ? ? 1 2 X N i,j=1 ?i?jyiyjK(xi , xj , ?) ) (6) with K(xi , xj , ?) = exp ? kxi-xj k 2 2?2  The reason for not using tree-based ensembles (Random Forest and boosted trees) was empirical because SVM resulted in superior cross-validated accuracy (of note, for RF both out-of-bad estimation and cross-validation were assessed, achieving similar values). IV. RESULTS The original dataset contained 410 features (32 for SBM and 378 for FNC). After the feature trimming, we ended up with 122 variables. Next, we estimated hyper parameter ? (sigma, width parameter) for the Gaussian Radial Basis kernel and tuned C parameter (the penalty factor, controlling trade-off between model complexity and proportion of non-separable instances) using leave-one-out cross validation for the final SVM classifier. The resulted test set area under the receiver operating characteristic curve (AUC) was 0.923. Of note, cross-validated performance of various models that had been tested (RF, boosted trees, neural network, SVM) varied around 0.8 and 0.85 (for overall accuracy) and the public scores that we were receiving after the submissions were unstable. Therefore, we decided not to implement more complex solutions (ensembling, hierarchical models) and stopped on a relatively simple model. REFERENCES 1 J. van Os and S. Kapur, "Schizophrenia", Lancet, vol. 374, pp. 635-45, Aug 22 2009. 2 C. A. Ross, R. L. Margolis, S. A. Reading, M. Pletnikov, and J. T. Coyle, "Neurobiology of schizophrenia", Neuron, vol. 52, pp. 139-53, Oct 5 2006. 3 A. Jablensky, "The diagnostic concept of schizophrenia: its history, evolution, and future prospects",Dialogues Clin Neurosci, vol. 12, pp. 271-87, 2010. 4 M. E. Shenton, T. J. Whitford, and M. Kubicki, "Structural neuroimaging in schizophrenia: from methods to insights to treatments", Dialogues Clin Neurosci, vol. 12, pp. 317-32, 2010. 5 W. Pettersson-Yeo, P. Allen, S. Benetti, P. McGuire, and A. Mechelli, "Dysconnectivity in schizophrenia: where are we now?", Neurosci Biobehav Rev, vol. 35, pp. 1110-24, Apr 2011. 6 Y. Takayanagi, Y. Kawasaki, K. Nakamura, T. Takahashi, L. Orikabe, E. Toyoda, et al., "Differentiation of first-episode schizophrenia patients from healthy controls using ROI-based multiple structural brain variables",Prog Neuropsychopharmacol Biol Psychiatry, vol. 34, pp. 10-7, Feb 1 2010. 7 Kaggle, MLSP 2014 Schizophrenia Classification Challenge, 2014. 8 Karolis Koncevicius, MLSP 2014 Schizophrenia Classification Challenge: 3rd position (solution), 2014. 9 National Institude of Mental Health, Schizophrenia, 2016. 10 Richiardi, Jonas, et al.,"Machine learning with Brain Graphics: Predictive Modeling Approaches for Functional Imaging in Systems Neureosciece. IEEE Signal Processing Magazine" , 2013, vol.30, no.3, p. 58-70. 11 M. E. Shenton, T. J. Whitford, and M. Kubicki, "Structural neuroimaging in schizophrenia: from methods to insights to treatments", Dialogues Clin Neurosci, vol. 12, pp. 317-32, 2010. 12 W. Pettersson-Yeo, P. Allen, S. Benetti, P. McGuire, and A. Mechelli, "Dysconnectivity in schizophrenia: where are we now?" , Neurosci Biobehav Rev, vol. 35, pp. 1110-24, Apr 2011.


I'm Clifton!

Would you like to get a custom essay? How about receiving a customized one?

Check it out