Tree-informed Bayesian multi-source domain adaptation: cross-population probabilistic cause-of-death assignment using...

Area of Science:

• Biostatistics
• Public Health
• Epidemiology

Background:

• Determining causes of death (COD) outside vital statistics systems is difficult.
• Verbal autopsy (VA) is a common method, but requires methods adaptable to new populations (domains).
• Existing statistical methods for cause-specific mortality fractions (CSMFs) may not fully utilize between-domain similarities.

Purpose of the Study:

• To propose a domain-adaptive method for VA that integrates external information on between-domain similarities.
• To improve the accuracy of CSMF estimation and individual COD assignment in diverse populations.
• To provide a scalable and data-driven approach for analyzing VA data across different domains.

Main Methods:

• Developed a domain-adaptive method using a prespecified rooted weighted tree to encode between-domain similarity.
• Employed latent class models to characterize domain-specific response distributions.
• Utilized a logistic stick-breaking Gaussian diffusion process prior and spike-and-slab priors for information pooling.
• Conducted posterior inference using a scalable variational Bayes algorithm.

Main Results:

• Simulation studies demonstrated that the domain adaptation method improves CSMF estimation.
• The proposed method enhances the accuracy of individual COD assignment.
• Validation using a real-world dataset confirmed the method's effectiveness.

Conclusions:

• The proposed domain-adaptive method offers a significant advancement for verbal autopsy analysis.
• This approach effectively balances domain-specific characteristics with shared information across populations.
• The method has the potential to improve global health surveillance by providing more accurate mortality data.