AI Model Designed to Predict Antiretroviral Resistance in HIV Patients in Cameroon

Researchers in Cameroon have designed an artificial intelligence model that predicts antiretroviral resistance in people living with HIV using routine clinical data, offering a potential alternative to expensive genotypic testing in a country where laboratory capacity falls far short of demand. The study, published in BMC Primary Care, describes the construction and pilot testing of a machine learning system intended to flag patients at risk of drug resistance before treatment decisions reach a dead end.

The motivation is stark. According to the authors, antiretroviral resistance poses a major challenge for HIV patient management in Cameroon, with resistance prevalence reaching 15 percent in treatment-naive patients and up to 93.3 percent among those experiencing therapeutic failure. Detecting resistance currently relies on genotypic tests costing approximately 165 US dollars each, a price described as very expensive relative to the cost of living in Cameroon. These tests are largely inaccessible and are performed only at the transition from second- to third-line therapy.

The gap between need and capacity is wide. Of the 448,911 patients on antiretroviral treatment in the country, approximately 4,500 require annual testing, yet national testing capacity is limited to only 1,000 to 1,500 tests per year. That creates a coverage deficit of 66 to 78 percent, leaving most patients who need resistance information without it at the moment it matters most for their care.

To build the model, the team drew on data from two local sources: the Centre International de Référence Chantal Biya, which contributed records for 1,420 patients, and the Étoug-Ébé Electronic Medical Record, which supplied data for 4,982 patients. After processing, these datasets were used to train nine different machine learning algorithms, each learning to associate patterns in routine clinical information with the presence of antiretroviral resistance.

XGBoost, a gradient-boosted decision tree method widely used in structured medical data, achieved the best overall performance on the held-out test set of 272 patients. The model reached an area under the curve of 0.763, with a 95 percent confidence interval of 0.678 to 0.849, indicating moderate discriminative ability. Precision stood at 96.3 percent, meaning that when the model flagged resistance, it was very often correct. Recall, or sensitivity, was 73.6 percent, so roughly a quarter of resistant cases were missed. Specificity was 68.2 percent, with a confidence interval of 47.3 to 83.7 percent, and the negative predictive value was 18.5 percent, with a confidence interval of 11.6 to 28.3 percent. The F1-score, which balances precision and recall, was 0.834.

Those numbers deserve careful reading. The high precision suggests the model could be valuable for prioritizing patients who genuinely need confirmatory testing or closer clinical follow-up. The low negative predictive value, however, means a negative prediction from the model cannot reliably rule out resistance, a consequence of the underlying prevalence of resistance in the tested population. The authors present the tool as a screening and decision-support aid rather than a replacement for definitive genotypic confirmation.

Beyond the algorithm itself, the team developed a functional pilot interface built with Streamlit, a Python framework for rapid web applications, designed for real-time clinical decision support. The interface allows clinicians to enter patient data and receive a resistance risk prediction during a consultation, turning a batch-trained statistical model into something usable at the point of care.

The authors propose integration into the national electronic medical record system as a future step, with the aim of improving therapeutic management and reducing treatment failures. In their assessment, AI-based prediction of antiretroviral resistance using routine clinical data is a feasible and cost-effective approach in resource-limited settings, one that could support earlier treatment decisions, reduce reliance on expensive genotypic testing, and improve HIV patient management in Cameroon.

As with any predictive model trained on local data, performance in the pilot test set does not guarantee identical results across all Cameroonian treatment sites or in other countries. The study is a design and implementation report, and the model’s real-world value will depend on prospective validation, integration into clinical workflows, and whether earlier resistance predictions translate into better treatment outcomes for patients who today wait years for the genetic evidence that their regimen is failing.

Subject of Research: Machine learning prediction of antiretroviral resistance in HIV patients in Cameroon

Article Title: Design of an AI model for predicting antiretroviral resistance in HIV patients: a case study in cameroon

Article References: Kameni Homté, J. S., Chana, A. M., Batchakui, B., Onguene Zanga, L. D., Anguissa Abessolo, J. P., Temgoua, E., & Fokam, J. (2026). Design of an AI model for predicting antiretroviral resistance in HIV patients: a case study in cameroon. BMC Primary Care. https://doi.org/10.1186/s12875-026-03608-8

Image Credits: AI Generated

DOI: 10.1186/s12875-026-03608-8

Keywords: HIV, antiretroviral resistance, machine learning, XGBoost, Cameroon, clinical decision support, genotypic testing, electronic medical records, resource-limited settings, drug resistance, Design, model