Predictive Analytics for Retention in Care and Antiretroviral Therapy Adherence Using Supervised Learning: a Case Study of County Health Facilities in Kenya

Abstract

Developing countries in sub-Saharan Africa are scaling up ART programmes to reduce HIV transmission for patients infected by the diseases (UNAIDS, 2014). In healthcare organizations, a great problem is faced by healthcare providers to know the ART adherence and status of HIV/AIDS patients. In this research, a predictive model using supervised learning is developed to let clinicians and healthcare providers know the ART adherence of PLHIV using features of the patients’ treatment history. The research explores the use of different machine learning methods to be able to detect records of patients defaulting and switching ART treatment. The methodology used was CRISP-DM data mining process. The dataset collected from the Ministry of Health sampling unit illustrates measurable baseline and clinical variables such as body weight, ART regimen, patients enrolled in care, Z-Score and phenotype. Data preprocessing and transformation was done to ensure the dataset collected was clean. Predictive model was designed in the process of data collection, dataset preprocessing like missing data, outlier data, feature selection and feature transformation and normalization of the dataset. Data was split into train and test set i.e., 80% training and 20% for test set and model designing, training and evaluation was performed using Anaconda. The baseline results from the benchmark and performance evaluation showed that ensemble random forest algorithm performed the best with training accuracy of 81% and AUC of 79.3% compared to other binary algorithms and classification error rate of 0.333. The machine learning model that performed poorly was Naïve Bayes with an accuracy score of 20.0%. The researcher retrospectively followed 21551 records of patients who were seeking care at comprehensive health care units and county health facilities.