Information Technology Reference
In-Depth Information
I3: IF BMI
<
26
.
8 AND HTN
=
no AND gender
=
female
yes
These two rules provide a new KB insight. The induced decision tree divides fe-
male patients with normal blood pressure (HTN = no) and normal BMI (or slightly
overweight) into two age groups: age
AND age
>
56 THEN OSA
=
56. This specific age-based
division could be associated with an increased risk of OSA among postmenopausal
women. Thus, the existing rules for female patients can be modified to reflect the
higher postmenopausal risk for OSA.
≤
56 and age
>
14.5
Conclusions
In this paper, we examined the KE methods and their roles in the KM process in
medicine. We focused on the CDSS and its most important component - knowledge
base (KB). We discussed verification and validation of the medical KB in the context
of the user requirements. The CDSS users expect the KB component to be trans-
parent (human-readable), updatable (users should be able to modify KB), adaptable
(user should have support to adapt KB to their local clinical requirements), and
learnable (KB should be able to learn from experience). To address these issues,
we used a rule-based representation for KB. We built two highly specialized KBs
for small sets of diagnostic rules, which are used in two domains: (1) psychiatry,
for the diagnosis of clinical depression and (2) sleep medicine for the diagnosis of
obstructive sleep apnea. To represent the vagueness of medical concepts and data,
we used a combination of crisp and fuzzy rules. To represent the inherent uncer-
tainty of clinical prediction rules, we used the certainty factors. Furthermore, we
created two data warehouses, one for the data from the psychiatric clinic and the
other one for the data from the sleep disorders clinics. We used these warehouses
for the machine learning (data mining) tasks. We experimented with two main data
mining techniques: the "black-box" approach (artificial neural networks) and the
“transparent-box” approach (classification rule and decision tree induction).
In this paper, we argued that KB should include (1) an explicit mechanism for
self-verification, (2) an automated mechanism for learning from experience, and (3)
a built-in data warehouse (repository) of diagnosed cases. We demonstrated that
the data warehouse of diagnosed cases could be used for the validation of KB. Fur-
thermore, we used a machine-learning approach to demonstrate how the CDSS can
learn from experience, which is represented by diagnosed cases stored in a data
warehouse. We presented two examples of OSA diagnostic rules to illustrate how
the induced rules could be used for KB modification. We are planning to integrate
and formalize the proposed framework for the verification and validation of KB. We
will expand the fuzzy-logic based rules and the fuzzy inference mechanism. Further-
more, we intend to create KB rules for other well-known predictors of obstructive
sleep apnea such as diabetes, large neck, and excessive daytime sleepiness. Also,
we will expand the rule system for the diagnostic criteria for clinical depression.
