Bayesian Tensor Decomposition for Clustering Latent Symptom Profiles for Verbal Autopsy Data.

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Accurate cause-of-death data is vital for public health but challenging in low- and middle-income countries (LMICs).
• Verbal autopsy (VA) is a key method for estimating mortality in LMICs, relying on caregiver interviews.
• Existing latent class models for VA require many classes, hindering interpretation of symptom profiles.

Purpose of the Study:

• To develop a novel Bayesian tensor decomposition framework for verbal autopsy.
• To improve both predictive accuracy and interpretability of cause-of-death assignment.
• To provide a more parsimonious representation of symptom distributions in VA.

Main Methods:

• Proposed a flexible Bayesian tensor decomposition framework.
• Partitioned symptoms into groups to model joint distributions of sub-profiles.
• Applied the methods to the PHMRC gold-standard VA dataset.

Main Results:

• Achieved better predictive accuracy compared to existing VA methods.
• Provided a more parsimonious representation of symptom distributions.
• Offered new insights into symptom and cause clustering patterns.

Conclusions:

• The proposed Bayesian tensor decomposition offers a superior approach to verbal autopsy.
• This method enhances understanding of population health trends and inequalities.
• It facilitates more effective public health interventions through improved mortality data.