AI for prognosis among people living with HIV: Protocol for a systematic review and meta-analysis

BACKGROUND

Advanced HIV disease remains a major global health concern, with nearly 40.8 million people living with HIV as of 2024. Antiretroviral therapy has improved outcomes, but its success depends on timely intervention, adherence, and retention in care. AI, including machine learning and deep learning, offers promising tools for prognostic modeling that could support clinical decision-making and personalized treatment. Existing syntheses, however, have been largely narrative and have not systematically evaluated the performance, risk of bias, reporting quality, or clinical readiness of AI-based prognostic models, leaving a critical gap in understanding their validity and applicability.

OBJECTIVE

This study aims to conduct a systematic review and meta-analysis of AI-based prognostic models predicting treatment and disease outcomes among people living with HIV, with a focused assessment of predictive performance, methodological rigor, reporting transparency, and potential for clinical implementation.

METHODS

A comprehensive search of 5 databases (PubMed, Embase, Scopus, OpenAlex, and IEEE Xplore) covered studies published from January 2015 to December 2025, using a 3-block strategy combining AI, HIV, and clinical outcome terms. Eligible studies are original research using AI to predict individual-level outcomes among people living with HIV. Primary outcomes are virologic and immunologic measures, disease progression, and mortality; secondary outcomes include retention in care, treatment failure, antiretroviral therapy discontinuation, drug resistance, opportunistic infections, hospitalization, and HIV-related comorbidities. Data extraction will use a standardized CHARMS-based form extended with elements from PROBAST+AI, TRIPOD-AI, DECIDE-AI, and the NeurIPS paper checklist; these tools will also inform assessment of risk of bias, reporting transparency, implementation, and reproducibility, with a prespecified framework for integrating overlapping or conflicting judgments. Where studies are clinically and methodologically comparable, accuracy metrics (eg, area under the curve, sensitivity, and specificity) will be synthesized using random-effects models, including bivariate analyses and hierarchical summary receiver operating characteristic curves. Analyses will be conducted in R, with code and study-level data made openly available. Heterogeneity will be explored through subgroup analyses and meta-regression, and the strength of evidence will be graded using an adapted GRADE framework incorporating AI-specific quality dimensions.

RESULTS

The literature search was completed in June 2026, yielding 9194 records across the 5 databases. Following deduplication (3138 records removed), 6056 unique records are pending for title and abstract screening. Full-text eligibility assessment, data extraction, quality appraisal, and quantitative synthesis are expected to begin in August 2026, with final results anticipated for publication by late 2026.

CONCLUSIONS

This protocol will provide a focused and methodologically rigorous synthesis of AI-based prognostic models in HIV care, identifying models with robust predictive performance and highlighting critical gaps in validation, reporting, and clinical readiness to inform best practices for future development and implementation.