Predictive Analytics for ART: How machine learning can predict treatment failure among HIV positive patients on antiretroviral therapy

Widespread use of Electronic Medical Records (EMRs) for clinical care has increased the availability of longitudinal, individual level data at both facility and national level. At an individual level, data can help to monitor health outcomes. However, when health outcomes are monitored for cohorts of patients with similar interventions, machine learning algorithms can predict health outcomes for other patients. As data proliferates and machine learning algorithms become more sophisticated, health programs are harnessing this emerging technology to predict and prevent adverse outcomes among patients infected with HIV. WHO recommends that clinicians use viral load test results to diagnose treatment failure. These tests are typically conducted six months after ART initiation and annually thereafter. When treatment failure occurs, this long testing interval can delay detection for weeks – or even months – preventing a timely intervention. With machine learning, a patient's attributes can be used to predict and proactively address their risk of treatment failure. Under the CDC-funded Kenya HMIS-II project, Palladium adopted and tested multiple machine learning algorithms to assess the applicability of specific demographic, clinical, and biomarker information to predict the risk of treatment failure. Kenya HMIS-II used machine learning models on a small set of de-identified patient records to explore consistencies in the predictions made by different machine algorithms, using a selected set of patient attributes. The models compared the attributes of patients who exhibited treatment failure with patients who had successful viral suppression. The findings showed consistency across different machine learning models applied in predicting treatment failure for participating patient records. The program is planning explore the likelihood of higher precision if these models are applied to larger datasets. This workshop will demonstrate how machine learning algorithms can be applied to health data to help predict treatment failure and enable health workers to intervene.