Page 218 - Special Topic Session (STS) - Volume 4

STS580 Vassilis P. P.

Figure 5: 2-dimensional visualization using t-SNE along with the retrieved clusters from k-
means, when k is set to 2 (left) and 6 (center), respectively, and 2-dimensional visualization using
t-SNE along with the retrieved clusters of Density Peak, when applied on the 2-dimensional data
(right)

In the last part of our experimental analysis in an attempt to make the
evaluation more accessible (assuming that we trust the visualization procedure
enough), we apply a clustering methodology directly on the 2-dimensional
mapping retrieved by t-SNE. This way we guaranty that the clustering result
will appear in a more suitable manner. For this purpose, we also employ the
Density Peak algorithm [20], which can also automatically estimate the number
of existing clusters. The results are reported in Figure 5(right). It is evident that
there exist clear clusters in our dataset. With this evidence one can claim that
Clinics belonging to different clusters could have different control limits and
bounds. The preliminary results we are reporting here can be an extremely
helpful tool when designing Health Policies.

4. Conclusions
Big Data offers a great opportunity in the healthcare domain to elucidate
biomedical research fields such as the healthcare fraud detection. Frauds in
the health domain constitutes an important issue for both states and citizens
since it holds a significant percentage of the annual healthcare expenditure
globally. Nowadays, where we are in Big Data era and ML approaches have a
recent advent, there is the potential for new computational tools able to
handle the challenges of fraud detection in healthcare.
Our analysis based on clinical data from EOPYY, focusing in investigating
the Clinics behavior with respect to their hospital expenditure. Outcomes
indicates that it is obvious that there are clear patterns regarding the Clinics
found in our dataset. We now have enough evidence to claim that Clinics that
belong to different clusters should be examined under different
circumstances. For example, control limits and bounds for Clinics could be
scaled according to the cluster they belong to.

5. Acknowledgment
This project is funded by the International Research Project,“Collective
wisdom driving public health poli¬cies - CrowdHEALTH”, in terms of the

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