Abstract Body

Background

Despite the progress achieved in HIV case identification through universal testing, shifting to risk-based testing can enhance HIV identification. While machine learning (ML) remains underutilized, it has the potential to improve HIV risk prediction and facilitate targeted HIV testing. This study evaluates the potential of ML in enhancing case identification.

Methods

We analyzed retrospective data from 58 health facilities in Lower Eastern Kenya (January 2024–March 2025) that utilized the Kenya Electronic Medical Record (EMR) system with ML integration. The ML algorithm applied predefined risk indicators such as age, sex, sexually transmitted infection (STI) history, prior HIV testing, number of sexual partners, condom use, and partner HIV status. These indicators were weighted based on their association with HIV risk, and clients were categorized into low, moderate, high, or very high risk using cumulative scores from the ML model. HIV positivity was the primary outcome, with the ML-assigned risk category as the main exposure. Logistic regression was used to identify correlates of HIV positivity.

Results

Among the 39,995 individuals tested, HIV positivity rate was 1.6% (n=620). Positivity increased across ML-assigned risks, with low risk at 0.5% (84/17,519), moderate risk at 1.2% (164/13,155), high risk at 2.5 % (156/6,362), and very high risk at 7.4% (216/2,919). The median age was 32.0 (IQR=17). Age was associated with increased odds of HIV positivity (adjusted odds ratio [aOR = 1.01, 95% confidence interval [CI]: 1.01–1.02, p < 0.001). Compared to the low-risk group, individuals in the moderate, high, and very high-risk groups had 2.52 (95% CI: 1.95–3.32), 6.61 (95% CI: 5.04–8.67), and 22.23 (95% CI: 17.03–29.01) increased odds of testing positive. Population at Elevated Risk had 0.13 reduced odds of HIV positivity (aOR = 0.13, 95% CI: 0.09–0.18, p < 0.001) compared to the general population. Clients tested at the vertical transmission prevention (VTP) service delivery point had lower odds of HIV positivity (aOR = 0.48, 95% CI: 0.38–0.60) compared to individuals with other testing modalities. Repeat testers, compared to new testers, had lower odds of HIV positivity (aOR = 0.77, 95% CI: 0.65–0.93, p < 0.001).

Conclusions

Machine learning predicted HIV case identification among high-risk individuals, yielding higher positivity rates, implying its potential to optimise targeted testing strategies and enhance case finding in high-burden HIV settings.