A Machine-Learning Approach to Identify Candidates for HIV Pre-Exposure Prophylaxis

World AIDS Day
To facilitate increased use of PrEP, researchers from France developed a machine-learning algorithm to identify individuals at increased risk for HIV infection.

Despite HIV pre-exposure prophylaxis (PrEP) being available in France and reimbursed through the country’s social security system, the treatment remains underutilized among men who have sex with men (MSM). To facilitate increased use, researchers from France developed a machine-learning algorithm to identify individuals at increased risk for HIV infection who should be encouraged to use PrEP. These findings were published in Studies in Health Technology and Informatics.

Electronic health records obtained from the eHOP database at Rennes University comprising 624,708 patients hospitalized between 2013 and 2019 were used to identify risk factors for HIV infection.

During the study period, 0.07% of patients (n=422) were diagnosed with incident HIV infection. Among these patients, 156 were included in the machine-learning model because they had sufficient data available.

The HIV cohort comprised 66.1% men and 33.9% women. Of the men, 9.7% were MSM.

The strongest predictors of incident HIV infection were MSM status (odds ratio [OR], 1.285), history of syphilis (OR, 1.206), and history of schizophrenia (OR, 1.176).

Among the 621,370 individuals without incident HIV infection, 39.8% had at least 1 risk factor, such as MSM status (1.8%).

Predictors of decreased risk for incident diagnosis of HIV infection included recent hepatitis C virus testing (OR, 0.460), previous sexually transmitted infection testing (OR, 0.560), and previous sexual disorders (OR, 0.852).

The best performing model for incident HIV infection had an area under the curve of 0.888 (95% CI, 0.814-0.962), sensitivity of 0.733, and specificity of 0.887.  This model was able to identify 73% of patients with incident HIV infection.

Because data were sourced from a single center, these findings may not be generalizable to other populations.

The study authors concluded this model may be used to target individuals at increased risk for incident HIV infection in order to encourage PrEP use and decrease rates of HIV infection.


Duthe JC, Bouzille G, Sylvestre E, Chazard E, Arvieux C, Cuggia M. How to identify potential candidates for HIV pre-exposure prophylaxis: an AI algorithm reusing real-world hospital data. Stud Health Technol Inform. 2021;281:714-718. doi:10.3233/SHTI210265