Gaussian process modelling of infectious diseases using the Greta software package and GPUs.

Area of Science:

• Computational statistics
• Epidemiology
• Bioinformatics

Background:

• Gaussian processes are essential for non-parametric inference in applied sciences.
• Existing software packages facilitate Gaussian process fitting and prediction.
• Spatio-temporal modeling is crucial for understanding infectious disease dynamics.

Purpose of the Study:

• To evaluate the Greta software for Bayesian inference of Gaussian process regression models.
• To apply these models to spatio-temporal infectious disease outbreak data.
• To assess the computational efficiency and predictive performance of Greta on GPUs.

Main Methods:

• Utilized Greta software, built on Tensorflow, for Bayesian inference.
• Applied Gaussian process regression to spatio-temporal tuberculosis incidence data.
• Compared computational time on Graphics Processing Units (GPUs) versus Central Processing Units (CPUs).
• Investigated the influence of covariance kernel choice on model inference and extrapolation.

Main Results:

• Achieved up to a 70% reduction in computational time using GPUs compared to CPUs for complex spatio-temporal models.
• Demonstrated the impact of different covariance kernels on inferring disease spread and extrapolating predictions.
• Successfully applied the inference pipeline to real-world tuberculosis incidence data.

Conclusions:

• Greta software offers significant computational speed-ups for spatio-temporal Gaussian process modeling via GPUs.
• Covariance kernel selection is critical for accurate inference and extrapolation in disease modeling.
• The approach provides a viable tool for predicting infectious disease outbreaks using complex spatio-temporal data.