Gaussian process emulation to improve efficiency of computationally intensive multidisease models: a practical...

Area of Science:

• Public Health
• Epidemiology
• Health Informatics

Background:

• The rising burden of non-communicable diseases (NCDs) in people with HIV (PLWH) in sub-Saharan Africa (SSA) necessitates accurate predictive models for care planning.
• Existing multidisease simulation models are computationally intensive, requiring extensive parameters and run times, limiting their use in resource-constrained environments.

Purpose of the Study:

• To introduce a novel, efficient, and user-friendly emulator for approximating complex simulators of long-term HIV and NCD outcomes in Africa.
• To provide a tutorial for implementing the emulator using publicly available data from Kenya.

Main Methods:

• Developed a Gaussian process-based emulator to approximate agent-based simulation models.
• Derived emulator parameters from existing simulation models and published literature on HIV, hypertension, and depression.
• Validated the emulator's accuracy using Bayesian posterior predictive checks and leave-one-out cross-validation.

Main Results:

• The emulator achieved a 13-fold improvement in computing time compared to traditional simulation models.
• A single emulator run completed in seconds on a standard laptop, versus hours on a high-performance computing cluster.
• Sufficient predictive accuracy was demonstrated with Pareto k estimates below 0.70.

Conclusions:

• The developed emulator is a practical and flexible tool for informing health policy in regions facing both HIV and NCD burdens.
• Future applications include forecasting disease burdens, estimating prevalence of co-occurring conditions, and projecting intervention impacts.