This article is more than 5 years old. Information may no longer be current.
Automated analysis of EHR data could identify candidates for PrEP
Click Here to Manage Email Alerts
NEW ORLEANS — Using EHR data from a large group of ambulatory practices in Massachusetts, researchers developed an automated algorithm to identify patients at an increased risk for acquiring HIV. Their research raised the question of whether this type of analysis could be used to identify potential candidates for pre-exposure prophylaxis, or PrEP.
“We think that the answer is yes,” Douglas S. Krakower, MD, assistant professor of medicine at Harvard Medical School and staff physician in the division of infectious diseases at Beth Israel Deaconess Medical Center in Boston, said during a presentation here at IDWeek 2016.
“But we have to be humble here. Yogi Berra famously was quoted as saying, ‘It’s tough to make predictions, especially about the future,’ ” Krakower joked. “I think that will hold true in terms of algorithms for prediction in this sort of exercise where it can be improved to some degree. And we have to make sure that they are applied in a way that’s useful to frontline clinicians.”
About 80,000 people are receiving PrEP in the United States, according to Krakower — far below the 1.2 million Americans who are eligible for the HIV prevention drug, according to CDC estimates.
Krakower and colleagues used the EHR repository of Atrius Health, which serves approximately 800,000 patients in the greater Boston area. The first step in developing their algorithm involved extracting potentially relevant data such as age, gender, race and certain diagnoses, prescriptions, laboratory tests and procedures from the repository for patients with incident HIV and patients without HIV, using the latter group as a control.
Then, each case of incident HIV infection between 2006 and 2015 (n = 138) was matched with up to 100 control cases with similar gender and who were affiliated with Atrius for a comparable duration. Third, they developed logistical regression models and machine learning algorithms to predict incident HIV infection in cases vs. controls.
According to Krakower and colleagues, strong predictors of incident HIV infection included prior diagnosis of anorectal ulcer, total number of positive gonorrhea tests and acute HIV testing in the past 2 years, among others.
Their results showed that some 8,414 patients were potential new candidates for PrEP. These patients also were candidates for HIV testing. While that might seem like a daunting number, Krakower said, it represented just 1.1% of the total number of patients in the Atrius system.
“I think that this is a clinically reasonable and manageable group of the population for this more intensive screening and to identify people with alerts,” he said.
However, this approach has limitations. For one, Krakower said the reliance on EHRs means that certain behavioral data will never be part of the algorithm.
Next, Krakower and colleagues want to validate their algorithm by testing its performance in a population where PrEP is more normative than it is within the Atrius system, and then modify it accordingly. They also plan to seek clinicians’ perspectives to make sure the analysis is delivered in a useful and appropriate way.
“Finally,” Krakower said, “we would like to pilot test this in a real-world clinical setting with frontline clinicians and see if it leads to increased appropriate utilization of PrEP.” – by Gerard Gallagher.
Reference:
Krakower D, et al. Abstract 860. Presented at: IDWeek; Oct. 26-30, 2016; New Orleans.
Disclosures: Krakower reports receiving unrestricted research support from Gilead Sciences. Please see the full study for a list of all other authors’ relevant financial disclosures.