Generating synthetic multi-national longitudinal cohorts for clinically grounded HIV research

Area of Science:

• Artificial Intelligence
• Biostatistics
• Epidemiology

Background:

• Longitudinal cohort data for people with HIV (PWH) are crucial for open science but face privacy challenges.
• Generating realistic synthetic HIV data is difficult due to complex temporal dynamics and missing information.

Purpose of the Study:

• To introduce Medical Longitudinal latent Diffusion (MeLD), a novel generative model for synthesizing realistic, variable-length clinical trajectories.
• To address the challenges of privacy, missingness, and temporal complexity in HIV longitudinal data.

Main Methods:

• Developed MeLD, a generative model utilizing latent diffusion for synthesizing mixed-type clinical data.
• Applied MeLD to the Caribbean, Central, and South America Network for HIV Epidemiology (CCASAnet) cohort data.
• Evaluated MeLD against state-of-the-art methods on data utility, fidelity, and privacy.

Main Results:

• MeLD successfully synthesized variable-length, decades-spanning clinical trajectories with missing data.
• The model demonstrated superior performance in data utility, fidelity, and privacy compared to existing methods.
• MeLD accurately reproduced longitudinal inference, including time-to-death estimates and risk factor effects.

Conclusions:

• MeLD provides a robust, privacy-preserving method for generating synthetic longitudinal HIV data.
• The synthesized CCASAnet cohort is an openly accessible resource for HIV research, hypothesis generation, and methods innovation.
• This work facilitates reproducible research and advances data-driven innovation in HIV epidemiology.