Work

Uncertainty-aware model triage for HIV-1 INSTIs phenotypic resistance in global populations

Year: 2026

Type: article

Abstract: Background As WHO expands dolutegravir (DTG) globally, non-B HIV-1 subtypes prevalent in low- and middle-income countries (LMICs) introduce cross-resistance risks driven by natural polymorphisms. Traditional rule-based genotypic algorithms falter to resolve high-dimensional epistatic interactions, exacerbating subtype biases and threatening virologic suppression. We developed a transparent AI tool to deal with these health inequalities. Methods Following TRIPOD+AI guidelines, we conducted a retrospective study constructing a multi-task attention network (MTL-Attention) to predict phenotypic cross-resistance to five integrase inhibitors from HIV-1 sequences. Monte Carlo Dropout was integrated to quantify epistemic uncertainty, establishing a “system deferral” mechanism. Epidemiological generalizability, along with fairness, was externally validated on a real-world cohort of 19,350 diverse sequences from the LANL database. Results MTL-Attention achieved an AUC of 0.955 for DTG and 1.000 for cabotegravir (CAB). Decision curve analysis confirmed a positive net clinical benefit across 0% to 80%. In the LANL cohort, epistemic uncertainty variance exhibited robust cross-domain stability between B and LMIC-prevalent non-B subtypes (Mann-Whitney U test, P = 0.999). Crucially, large-scale screening identified an 11.57% cryptic phenotypic drift in DTG and an 18.65% predicted baseline resistance rate against CAB among global non-B strains. Conclusions MTL-Attention provides a low-bias solution for predicting HIV-1 phenotypic cross-resistance. When translated into a Web CDSS, it delivers a low-cost computational triage tool designed for LMICs hospitals. To advance public health, we recommend implementing this human-in-the-loop AI workflow to reduce clinical automation bias, optimize equitable allocation of high-value antiretrovirals, and strengthen worldwide surveillance systems for emerging resistant variants. Key messages • A novel uncertainty-aware AI tool enables low-cost computational triage for HIV-1 cross-resistance, mitigating clinical automation bias and ensuring equitable patient safety in resource-limited LMICs. • Real-world deployment identified an 11.57% cryptic phenotypic drift for dolutegravir and 18.65% baseline resistance to cabotegravir among global non-B subtypes, warning public health systems.

Source: European Journal of Public Health

DOI: 10.1093/eurpub/ckag169.451