Application of Bayesian Networks (BN) Technology in Predicting Major Factors Behind Poor ART Adherence Trends in Ethiopia - the case of SNNPR region
No Thumbnail Available
Date
2009-03-12
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
AAU
Abstract
This research presented the concepts of knowledge breakthrough based on the Bayesian
Networks technology to extract valid models of knowledge. The application domain of
the research is the health sector, one of the potential sectors to apply Bayesian Network
Technology. The research generally aimed at investigating the potential applicability of
Bayesian Network technology in developing a model that can support the prediction of
ART clients’ adherence tends in Ethiopia. The research was conducted in selected
hospitals of SNNP Regional Health Bureau.
The methodology used to conduct this research consists of four major steps: Data Source
Identification, Data Collection, Data Preprocessing and Model Building/Testing. A total
of 1561 records, having 15 attributes each, were used for building, training and testing a
Bayesian Network model.
The Bayesian Network learning process was done for complete data (the data for which
the training data containing no missing values) which again applied constraint-based
approach. This approach performs conditional independence tests on the data. Then it
will search for a network that is consistent with the observed dependencies and
independencies (applying d-separation concept). Conditional independence relationships
among the attributes can serve as constraints to construct a Bayesian Network.
The belief network modelling software employed for the purpose of training and testing
the BN model was the Belief Network PowerSoft, which applies a constraint-based
XII
approach. Three-Phase-Dependency-Analysis (TPDA) in BN PowerConstructor was
employed in developing the model. The BN PowerPredictor, on the other hand, was
used to evaluate the prediction accuracy of the model. BN PreProcessor was also there
to preprocess the data so as to make it ready for model building purpose.
Model testing was implemented in two phases- the first phase without involvement of
expert knowledge (i.e., without node ordering, in which, the algorithm learns both
structure and parameters), and the second phase by eliciting domain expert knowledge
(i.e., involving node ordering, in which, the algorithm learns only the parameters). For
both cases, to ensure consistency across the data during the selection of the test and
training sets, experiments were carried out by splitting the data into 10 partitions, i.e., a
percentage split (10-fold) was used to partition the dataset into training and test data.
Each partition, in turn, was used for testing while the remainder was used for training.
This process was repeated ten times for the learning algorithm and, at the end, every
instance was used exactly once for testing. Finally, the average result of the 10-fold cross
validation was considered.
Accordingly, the average predictive accuracy for the model without an expert
intervention (Experiment I) was 72.80% at 95% confidence level. According to this
model, the adherence (to the medication) of an ART client is directly affected by two
factors: Addiction (drug or alcoholic behaviour) and loss of job due to ill-health.
Next, TPDA algorithm (Experiment-II) was implemented by allowing elicitation of
expert knowledge. Accordingly, the predictive accuracy of the modified model was
75.9% which is a better result.
XIII
Significant enhancements in prediction and reduction in error rates in the modified model
was taken as the indication of the significance of a domain experts’ intervention during
model building.
According to the later model (Experiment II), adherence of an ART client is directly
affected by six factors: Addiction (drug or alcoholic behaviour), loss of job due to illhealth,
Residence of a patient, Knowledge the client has concerning HIV,
Employment status of the client, and Family dependence (independence).
From the model developed, it was observed that Bayesian Network is a powerful
predictor even in the absence of a domain expert. With a proper intervention of domain
experts, it was noticed to perform even better.