Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bayesian Federated Cause-of-Death Classification and Quantification Under Distribution Shift

Yu Zhu, Zehang Richard Li

TL;DR

This work tackles distribution shift in verbal autopsy by introducing a Bayesian Federated Learning (BFL) framework that ensembles $p_m(X|Y)$ from multiple pre-trained VA analyses without sharing raw data. The target-domain inference uses a latent indicator to combine domain-specific conditional likelihoods: $p_0(X|Y=c) = \sum_{m=1}^M \lambda_{cm} p_m(X|Y=c)$, enabling both individual-level cause assignment and population-level CSMF quantification with limited local labels. The approach is modular, supports various base VA methods, and is evaluated on PHMRC and CHAMPS neonatal datasets, showing robust gains over single-domain baselines and competitive performance to pooled analyses. The findings demonstrate that privacy-preserving federated ensembling can effectively handle both label shift and conditional distribution shift across diverse populations, guiding practical deployment in real-world mortality surveillance.

Abstract

In regions lacking medically certified causes of death, verbal autopsy (VA) is a critical and widely used tool to ascertain the cause of death through interviews with caregivers. Data collected by VAs are often analyzed using probabilistic algorithms. The performance of these algorithms often degrades due to distributional shift across populations. Most existing VA algorithms rely on centralized training, requiring full access to training data for joint modeling. This is often infeasible due to privacy and logistical constraints. In this paper, we propose a novel Bayesian Federated Learning (BFL) framework that avoids data sharing across multiple training sources. Our method enables reliable individual-level cause-of-death classification and population-level quantification of cause-specific mortality fractions (CSMFs), in a target domain with limited or no local labeled data. The proposed framework is modular, computationally efficient, and compatible with a wide range of existing VA algorithms as candidate models, facilitating flexible deployment in real-world mortality surveillance systems. We validate the performance of BFL through extensive experiments on two real-world VA datasets under varying levels of distribution shift. Our results show that BFL significantly outperforms the base models built on a single domain and achieves comparable or better performance compared to joint modeling.

Bayesian Federated Cause-of-Death Classification and Quantification Under Distribution Shift

TL;DR

This work tackles distribution shift in verbal autopsy by introducing a Bayesian Federated Learning (BFL) framework that ensembles from multiple pre-trained VA analyses without sharing raw data. The target-domain inference uses a latent indicator to combine domain-specific conditional likelihoods: , enabling both individual-level cause assignment and population-level CSMF quantification with limited local labels. The approach is modular, supports various base VA methods, and is evaluated on PHMRC and CHAMPS neonatal datasets, showing robust gains over single-domain baselines and competitive performance to pooled analyses. The findings demonstrate that privacy-preserving federated ensembling can effectively handle both label shift and conditional distribution shift across diverse populations, guiding practical deployment in real-world mortality surveillance.

Abstract

In regions lacking medically certified causes of death, verbal autopsy (VA) is a critical and widely used tool to ascertain the cause of death through interviews with caregivers. Data collected by VAs are often analyzed using probabilistic algorithms. The performance of these algorithms often degrades due to distributional shift across populations. Most existing VA algorithms rely on centralized training, requiring full access to training data for joint modeling. This is often infeasible due to privacy and logistical constraints. In this paper, we propose a novel Bayesian Federated Learning (BFL) framework that avoids data sharing across multiple training sources. Our method enables reliable individual-level cause-of-death classification and population-level quantification of cause-specific mortality fractions (CSMFs), in a target domain with limited or no local labeled data. The proposed framework is modular, computationally efficient, and compatible with a wide range of existing VA algorithms as candidate models, facilitating flexible deployment in real-world mortality surveillance systems. We validate the performance of BFL through extensive experiments on two real-world VA datasets under varying levels of distribution shift. Our results show that BFL significantly outperforms the base models built on a single domain and achieves comparable or better performance compared to joint modeling.
Paper Structure (13 sections, 11 equations, 11 figures, 1 table)

This paper contains 13 sections, 11 equations, 11 figures, 1 table.

Table of Contents

  1. Introduction
  2. Cause-of-death assignment using VA algorithms
  3. Bayesian federated learning for VA data
  4. Joint modeling of VAs from multiple domains
  5. Federated learning and model averaging
  6. The BFL model for VA data
  7. Fine-tuning with local labeled data
  8. Comparing BFL with VA calibration
  9. Simulation analysis using PHMRC gold-standard dataset
  10. The case with no within-target label shift
  11. The cases with within-target label shift
  12. Analysis of CHAMPS neonatal Dataset
  13. Discussion

Figures (11)

  • Figure 1: The BFL workflow for VA analysis. The solid black arrows indicate the information shared from training domains to the global model. The dashed arrow indicates the information sharing from the optional local base model trained on the target domain (in the BFL-domain and BFL-mix models).
  • Figure 2: No within-target label shift: comparison of CSMF accuracy across different methods. The dashed line shows the CSMF accuracy from BFL without local labeled data. The modified GBQL-50 and LCVA rank the top two methods in all sites. BFL-partial slightly outperforms BFL-domain and BFL-mix. BFL models are consistently better than the base models trained on a single domain (local-self and local-avg).
  • Figure 3: No within-target label shift: comparison of top cause accuracy across different methods. The dashed line shows the top cause accuracy from BFL without local labeled data. LCVA with full data pooling achieves the highest top cause accuracy in four sites. In Mexico city and Dar es Salaam, LCVA trained locally on the labeled data (local-self) achieves the highest accuracy. BFL-domain is consistently better than the base models trained on a single domain, except for the local models trained on these two sites.
  • Figure 4: Mild within-target label shift: comparison of CSMF accuracy across different methods. The differences between models are smaller, especially between LCVA and the various BFL models.
  • Figure 5: Mild within-target label shift: comparison of top cause accuracy across different methods. BFL-domain achieves highest average accuracy in five out six sites and slightly below LCVA in Pemba.
  • ...and 6 more figures