Providing Case-Based Retrieval as a Decision Support Strategy in Time Dependent Medical Domains - Computational Intelligence in Healthcare: Advanced Methodologies - page 212

time

D

granularity 4h symbol taxonomy

D W D W

D W D S
D S D W

granularity 2h

SD DS DD

D S D S

granularity 1h

SDSS

DDSS

DSSS

Fig. 6. An example multi-level orthogonal index structure

triggered by the types of queries which have been issued so far. If queries have
often involved a time granularity which is not yet represented in the index(es),
the corresponding level can be created. A proper frequency threshold for count-
ing the queries has to be set to this end. We proceed analogously by creating an
orthogonal index from each node which fits the frequent queries time granularity,
but does not match their symbol taxonomy level.
This policy allows to augment the indexes discriminating power only when
it is needed, while keeping the memory occupancy of the index structures as
limited as possible.
We will now illustrate query answering in our approach, by means of a case
study, taken from the haemodialysis domain. In particular, we will focus on a
single case feature, for the sake of clarity: namely, haematic volume (HV).
In a good session, HV fits a model where, after an initial period of D S (strong
decrease), which lasts for the first half of the session, a D W (weak decrease) of
the volume follows. An example query summarizing the correct HV behaviour is
thus the following: D S D W , where each symbol represents a 2-hours-long episode
(globally covering the overall 4 hours duration of the haemodialysis session).
To answer a query, in order to enter the index structure, we first progressively
generalize the query itself in the symbol taxonomy direction (see figure 7 - step
1), while keeping time granularity fixed. Then, we generalize the query in the
time dimension as well (see figure 7 - step 2).
Then, following the generalization steps backwards, we can enter one of the
indexes in the forest from its root, and descend along it, until we reach the node
which fits the original query time granularity. If an orthogonal index stems from
this node, we can descend along it, always following the query generalization
steps backwards. We stop when we reach the same detail level in the symbol
taxonomy as in the original query. If the query detail level is not represented
in the index, because the index is not complete, we stop at the most detailed
possible level. We then return all the cases indexed by the selected node.
In our example, the output of the generalization process allows to identify a
single index structure in the forest, namely the one whose root is D (i.e. the