Page 343 - Invited Paper Session (IPS) - Volume 2

IPS273 Tomoki Tokuda et al.

labelled ‘Treatment Effect’ in Fig.3. A higher value in these features implies that
a patient is not remitted. Therefore, we can characterize D1, D2, and D3 as
resistant, responsive, and responsive clusters for SSRI treatment. Furthermore,
features related to CATS are included in the same feature cluster F1. These
results suggest that the subject clusters D1, D2 and D3 for depressive subjects
may be related to after-treatment status of depression, which might be further
related to stress experiences during childhood. Lastly, feature clusters F2 and
F4 are related to specific functional connectivity in fMRI image data (i.e.,
angular-gyrus related FC), which suggests a possible association between the
subject clusters and neural substrates.
This result of cluster analysis raises the possibility of prediction of treatment
outcome prior to SSRI treatment. In this regard, we explore several important
implications drawn from the result. First, subject clusters can be represented
by a small number of relevant features. Since in our clustering method each
feature cluster consists of similar (i.e., highly correlated) features, a feature
cluster can be represented by a reduced number of these features. It turns out
that the subject clusters D1, D2 and D3 are represented by CATS scores
(associated to feature cluster F1), and the first principal scores of angulargyrus
related FC (associated to F2 and F4; we simply refer to it ‘AG related FC score’).
These features can indeed explain the resultant subject cluster membership
properly (Fig.4A). In the scatter plot of Fig.4A, AG-related FC scores
discriminate between subjects in D3 and other subjects. On the other hand,
CATS scores do not discriminate a single class by itself, but they discriminate
between subjects in D1 and D2, once subjects in D3 are sorted out. This
observation motivated us to consider a classifier that consists of the following
steps. First, we classify subjects into either D3 or non-D3 based on AGrelated
FC scores. A subject with low scores in AG-related FC is classified into D3,
otherwise into non-D3. Subsequently, the non-D3 subjects are classified into
either D1 or D2 based on CATS scores: A subject with low scores in CATS is
classified into D2, otherwise into D1. This procedure of classification is
summarized in Fig.4B. Since these subject clusters correspond to degrees of
remission of SSRI treatment as well, this classifier leads to predictions of
whether SSRI treatment may be effective, prior to the onset of treatment. We
can interpret this classification as follows. For subjects in D2 and D3, SSRI
treatment may be appropriate (low after-six-week BDI scores), while it may not
for those in D1.

330 | I S I W S C 2 0 1 9