Data Science for Decision Support in the HIV Care Cascade
Data Science for Decision Support in the HIV Care Cascade
The goal of this study is to determine whether clinical prediction algorithms derived using statistical machine learning methods can be used to improve patient outcomes in large HIV care programs in sub-Saharan Africa and elsewhere.
There are two main questions to be answered. First, can the prediction algorithms accurately identify those who are at risk for (a) missing scheduled clinic visits and/or (b) treatment failure, evidenced by elevated HIV viral load? And second, can the risk predictions be used in a structured way to (a) improve retention in care and/or (b) reduce the number of patients having elevated viral load? Researchers will develop machine learning prediction algorithms, incorporate the risk prediction information into the electronic health record, provide guidance to clinical health workers on use of the point-of-care interface tools that display risk prediction information, and incorporate feedback from clinic staff to modify and co-develop the protocol for using risk predictions for improving patient outcomes.
They will then compare the proportion of patients having missed visits and longer-term loss to follow up, and the proportion with elevated viral load, between clinics that use the information from the risk prediction algorithms and those that do not.
Clinical decision support systems (CDSS) tailored to the requirements of low- and middle-income countries (LMIC) have been shown to improve compliance with guidelines and quality of care by a range of healthcare staff. To be most effective CDSS should be developed and tested with large clinical data sets from the local region. Use of machine learning algorithms allows the development of prediction models for clinical complications and outcomes, which can guide health care staff in early identification of problems and appropriate interventions. This requires well established electronic health record (EHR) systems acting as both data sources and as platforms for delivering feedback through CDSS (Learning Health System approach). The EHR at the Academic Model Providing Access to Healthcare (AMPATH), a large HIV care program in western Kenya funded in large part by President's Emergency Plan for AIDS Relief (PEPFAR), has used a version of the OpenMRS EHR for nearly two decades and provides a unique environment for this research.
The objective of this proposal is to develop and implement data-driven tools to aid health-related decision-making at patient, clinic and county levels, and evaluate the efficacy of using these methods. The hypothesis is that health facilities utilizing these data driven CDSS will show improvements in the processes and outcomes of care compared to health facilities not utilizing data driven CDSS within their EHRs.
The two primary endpoints for the study are retention in care and viral load suppression.
The motivation is driven by 95-95-95 HIV cascade of care benchmarks established by UNAIDS for eradicating HIV worldwide. In brief, the framework calls for diagnosis of 95% of individuals who have HIV, initiating antiretroviral (ART) treatment for 95% of those who have been diagnosed, and achieving suppression of viral load (VL) for 95% of those who are on treatment. Our project addresses the second and third phases.
Regarding the second 95, retention is a necessary condition for maintaining persons living with HIV (PLWH) on antiretroviral therapy (ART) because global care guidelines now specify that all PLWH initiate ART once engaged in care. Regarding the third 95, in Kenya and many other LMIC, viral load testing for most adult clients is done six months after treatment initiation and annually thereafter. Even after a measured VL indicating suppression, viral failure due to nonadherence or drug resistance can occur well before the next follow up one year later. Hence our models will generate predicted viral load values in the interim and use them to flag individuals who should have a VL measurement prior to the scheduled follow up.
The trial is part of a larger NIH-funded study. The aims related to the trial are as follows:
Aim 1: Develop and validate statistical machine learning models and algorithms for clinical and programmatic decision support.
1a: Develop and validate statistical machine learning algorithms to identify those at high risk for disengagement from care and viral failure, and to generate predicted values of current viral load.
b: Develop representations of statistical uncertainty about the predictions to enable optimal decision making.
Aim 2. Develop, implement and field test decision support and data visualization tools to enhance data driven decision making by physicians and program managers.
a: Create the server architecture to support the prediction models in the OpenMRS user interface (UI).
2b: Develop and refine the specific protocol for using the risk predictions to reduce missed visits and reduce incidence of viral load failure.
Aim 3: Conduct evaluation of the impact and efficacy of the clinical decision support tools in the AMPATH Care Program
3a: Implement the CDSS at the point of care in all clinics using the AMPATH Medical Records System (AMRS) in Uzima and Dumisha catchment areas. These clinics have varying size and geographic location.
3b: Following a pilot phase, conduct a stepped wedge randomized longitudinal comparison of retention rates and viral load suppression rates in 30 clinics at AMPATH
The successful completion of the work will provide effective CDSS tools to improve HIV care in Kenya and other LMICs, as well as a set of tools for the development, updating and evaluation of CDSS for other clinical problems. Previous work by the investigators and colleagues in development and wide deployment OpenMRS in more than 44 countries provides a platform for broad dissemination of this work.
Inclusion Criteria:
Exclusion Criteria: