Applying machine learning to electronic health records can identify patients who might benefit from prophylactic medication that can reduce risk of HIV infection by up to 99 percent
BOSTON – Pre-Exposure Prophylaxis (PrEP) is a combination of antiretroviral medications that, when taken daily, can drastically reduce the risk of HIV infection. Yet, as of 2016, only seven percent of the 1.1 million Americans at highest risk of becoming infected with HIV received prescriptions for the preventative treatment. One reason for that low rate of PrEP use is that providers can have difficulty identifying patients who might benefit most from this highly effective preventative care.
Now, a team of researchers including infectious disease specialist Douglas S. Krakower, MD, of Beth Israel Deaconess Medical Center (BIDMC), has used machine learning to build an HIV risk prediction model to improve prescribing of the HIV-preventative medications among high risk populations. The scientists developed and validated automated algorithms that generate models that efficiently identify patients at increased risk for HIV infection based on electronic health records (EHR) data from a large healthcare system in Massachusetts. The research appeared in the July 5 edition of The Lancet HIV.
"Risk prediction tools are often used in other areas of medicine, such as cardiovascular disease, to help providers identify patients who might benefit from preventive care," said Krakower, an Assistant Professor of Medicine in BIDMC's Division of Infectious Diseases and Harvard Medical School. "However, existing HIV risk prediction tools have limitations and are not routinely used. Integrating our automated prediction models into EHRs to alert providers about patients who may benefit from PrEP could improve PrEP prescribing and, ultimately, prevent new HIV infections."
Using anonymized EHR data from Massachusetts' Atrius Health clinics, Krakower and colleagues – including collaborators from Brigham and Women's Hospital and the Massachusetts Department of Public Health – extracted demographic and clinical data from 1.1 million patients seen between 2007-2016 who were HIV-uninfected but had not yet used PrEP. Using data from 2007-2015, the team built their predictive model based on 180 potential predictors of HIV risk, including diagnoses, laboratory tests and results, and prescriptions for sexually transmitted diseases. Next, they validated the model using data from 2016 and data from Fenway Health, a neighboring independent community health center in Boston specializing in sexual health care, to see how well the models would perform in a new setting with higher rates of new HIV infections.
"Although risk prediction tools are imperfect and cannot replace the clinical judgment of skilled providers, our model may be substantially more efficient than efforts to identify PrEP candidates in current practice," said Krakower.
In the 1990s, antiretroviral therapy (ART) transformed HIV from a death sentence to a chronic disease. After a decade of clinical trials demonstrated that the same drugs could prevent HIV infection in up to 99 percent of people who took the drugs daily – an intervention known as PrEP – the FDA approved PrEP for use in HIV-negative adults at high risk of infection in 2012 and in high-risk adolescents in 2018. Yet, HIV infection rates have remained stable, with close to 40,000 Americans becoming newly infected each year. The federal government's new End the HIV Epidemic Initiative, which seeks to reduce new HIV infections in the US by 75 percent over the next 5 years and by 90 percent over the next 10 years, emphasizes that increased prescribing of PrEP will be critical to bending the curve on the HIV epidemic.
In a paper published in the same issue of The Lancet HIV, researchers in the Department of Population Medicine at Harvard Medical School – where Krakower is also an Assistant Professor – built a similar model using data from a west coast health care system. Taken together, the studies support the notion that effective deployment of the predictive models – which is the next step in their research agenda, with pilot studies set to begin this year – could prevent new HIV infections.
Additional authors include Kenneth H. Mayer, MD, of the Division of Infectious Diseases, Beth Israel Deaconess Medical Center and Fenway Health; Susan Gruber, PhD; J. T. Menchaca; Judith C. Maro, PhD; and Michael Klompas, MD, of the Department of Population Medicine, Harvard Medical School; Katherine Hsu, MD of the Bureau of Infectious Disease and Laboratory Sciences, Massachusetts Department of Public Health and the Department of Pediatrics, Boston Medical Center; Ben A. Kruskal, MD of Atrius Health and the New England Quality Care Alliance; and Ira B. Wilson, MD, of the Department of Health Services, Policy and Practice, Brown University. Dr. Klompas is also affiliated with the Division of Infectious Diseases, Brigham and Women's Hospital.
Dr. Krakower has conducted research supported by Gilead Sciences; has received honoraria for authoring or presenting continuing medical education content for Medscape, MED-IQ, DKBmed, and UptoDate, outside of the submitted work. Dr. Mayer has conducted research supported by Gilead Sciences and ViiV and has received funding for authoring continuing medical education content for UptoDate, outside of the submitted work. All remaining authors declare no competing interests.
This work was supported by the Harvard University Center for AIDS Research (P30 AI060354), the Providence/Boston Center for AIDS Research (P30AI042853), and the Rhode Island IDeA-CTR [U54GM11567], all programs funded by the National Institutes of Health; the National Institute of Mental Health (K23 MH098795); and the US Centers for Disease Control and Prevention through the STD Surveillance Network (SSuN, CDC-RFA-PS13-1306).
