Abstract:
Introduction: Data mining is the process of finding interesting hidden knowledge in large database. Since, antiretroviral therapy service in Ethiopia started in 2003 there is large amount of data gathered and stored in large databases. Due to lack of appropriate data analysis technique, this data were not used to overcome early detection and prevention of unintended outcomes from antiretroviral therapy.
Objective: the aim of this study is to predict outcomes of antiretroviral therapy among HIV/AIDS clients at Adama referral Hospital, using data mining techniques.
Methods: Institution based retrospective follow up study in using Cross-Industry Standard Process for Data Mining study on seven years records of Adama referral Hospital antiretroviral therapy clinic were conducted from April 17 to 24, 2012. In this study, electronic database composed of 8 relational table and 19088 records for both ART and pre ART clients were taken. The database suffers from multiple missing value and outliers, therefore cleaned by substituting mean for numeric and mode for nominal values. For important attributes, complete deletion was utilized. Finally, 10,690 records of adult antiretroviral therapy ever started clients were taken. The selected dataset was transformed into ARFF. Moreover, model building and evaluation was applied in support of Waikato Environment of Knowledge Analysis (WEKA) version 6.6.6 machine learning software.
Result: Among reviewed 10,690 records of ART clients 26.0% were drop out from ART services. Accordingly, twelve experiment were conducted in using both J48 and multilayer perceptron algorithms. Among the experiments, decision tree from J48 algorithm with balanced data set showed 93% accuracy of prediction. In addition, adherence, month on ART and family planning utilization were the top three-predictor attribute selected in the model.
Conclusion and recommendation: in this study, prediction of ART outcome in using data mining techniques, showed the applicability of data mining for early prevention of dropout and lost follow up from the ART. For this reason, this data mining approach could be integrated in the Adama Hospital ART clinic database for future prediction of new clients’ outcome.
