Page 342 - Invited Paper Session (IPS) - Volume 2
P. 342
IPS273 Tomoki Tokuda et al.
Figure 3: Heatmap of view 10. MDD subjects and healthy subjects are clearly
separated: Subject cluster C1-2 consist of only healthy subjects, while D1-3 consist of
only depression subjects. Further, feature cluster F1 includes features related to
treatment effect, which characterize D1-D3 as treatment resistant, responsive and
responsive.
Pre-processing of data:
For the numerical features, we standardized each feature using its mean and
standard deviation, based on all subjects (ignoring missing entries). In
contrast, for categorical and integer type of features, we did not carry out pre
processing. Note that we left missing entries as they were, because our
clustering method can handle them in a Bayesian manner without explicit
substitutions.
Results of multiple co-clustering: The multiple co-clustering method
yielded 15 views in which the view index is based on the number of features
in a view in the descending order. Specifically, we analyzed view 10 more in
detail (Fig.3). This view consists of five subject clusters, which match the label
of control/depression well: Two clusters for control subjects (subject clusters
C1, C2 in Fig.3) and three clusters for depressive subjects (subject clusters D1,
D2, D3). View 10 contained several non-FC features that discriminate well
among subject clusters. Further, the view had a high proportion of depression-
related features (60%), including 39 numerical features and 19 categorical
features, but none of the integer features. Since the categorical features do
not clearly distinguish between subject clusters, we focus only on numerical
features. These numerical features are clustered into five feature clusters F1-5.
It is noteworthy that the view contains a feature cluster (F1) that is related to
the after-treatment status of depression, namely, the features in a red box
329 | I S I W S C 2 0 1 9