Regularized Bayesian transfer learning for population-level etiological distributions.

Area of Science:

• Epidemiology
• Biostatistics
• Computer Science

Background:

• Computer-coded verbal autopsy (CCVA) algorithms estimate cause-specific mortality fractions from questionnaire data.
• CCVA accuracy can decrease when training data is not representative of the local population, a common transfer learning challenge.
• Existing transfer learning methods often focus on individual-level classification, not population-level epidemiological insights.

Purpose of the Study:

• To develop a novel transfer learning framework for accurate population-level cause-of-death estimation.
• To address the challenge of small sample sizes in epidemiological datasets.
• To improve the reliability of national and regional mortality estimates.

Main Methods:

• A parsimonious hierarchical Bayesian transfer learning framework is proposed.
• A shrinkage prior is introduced to handle small target-domain sample sizes and ensure consistency with baseline predictions.
• The framework is extended to an ensemble of baseline classifiers for a unified estimate.

Main Results:

• The proposed Bayesian framework directly estimates population-level class probabilities.
• The shrinkage prior ensures the model defaults to direct aggregation with no local data or perfect baseline classifiers.
• The ensemble approach effectively favors the most accurate baseline classifier.

Conclusions:

• The developed transfer learning framework offers a robust method for estimating population-level mortality fractions.
• This approach enhances the accuracy of cause-of-death estimates in diverse local populations.
• The methodology is particularly valuable for social and health scientists working with limited data.