Data mining approaches could also be used to establish
and validate guidelines for clinical practice for home
healthcare, which have been lacking [25]. Next steps
could be to include larger sample sets of data, or testing
those finding on other medical conditions. Other future
research could investigate further differences between US
regions or according to patient socioeconomic class, for
example rural and/or low income patients.
Problems with CART still include the fact that a cluster
tree does not always show the "best" picture of a problem
domain – a specific node or class might hide another, as
overtraining may occur. Another potential issue with
CART is that the trees are not always perfectly stable: for
example, a new variable can change the whole branches
and nodes. A derived consequence is that although trees
are optimal at each node, the optimality does not apply to
the overall tree. Taking these problems in consideration,
other statistical techniques could be used in combination
of CART to try to improve on these results. For example,
the bagging technique (a technique to reduce variance)
might be considered as this technique calculates a
"model" tree averaging the number of trees derived from
boot-strapped samples included in the original training
set of data. Another technique to improve results might be
the use of boosting which grows the tree several times, but
each time the trees are grown again, the data (i.e. patient
records) which was misclassified are assigned a greater
weight in order to improve the predictive nature of the
tree. Another future research area could be to improve validation
results by analyzing optimum size of trees for better
results for example.