Proposal

Accurate and credible cause-of-death (COD) data are lacking for 65% of the global population. In many low-and-middle-income countries (LMICs), a nonclinical approach called verbal autopsy (VA) is often adopted for determining COD in communities. VA is a standardized interview of caregivers to gather information on the decedent’s health and symptoms, which is then passed through artificial intelligence (AI) algorithms like InSilicoVA, InterVA, EAVA, and smartVA that are custom-made for VA data and produce estimated COD outputs. This algorithmic automation has made COD determination from VA scalable, driving its increased uptake for obtaining national-level cause-specific mortality fraction (CSMF) estimates, the proportion of deaths in a population attributed to a certain cause. CSMFs are critical metrics to understand the burden of diseases and guide public health policymaking.

The AI algorithms used to predict a COD from VA data systematically misclassify COD for many deaths, introducing significant algorithmic bias in the CSMF estimates obtained from VA data. A Bayesian transfer learning method, VA calibration, has been proposed to mitigate this algorithmic bias. VA calibration leverages a unique multi-country dataset from the CHAMPS project, containing paired information on COD from both VA and a more comprehensive (gold standard) diagnostic method, Minimally Invasive Tissue Sampling (MITS). The paired data is used to estimate the algorithmic bias of VA which is then adjusted for in studies with only VA data. VA calibration has improved CSMF estimates for child and neonatal deaths in Mozambique using VA-only data from the COMSA-Mozambique study.

This proposal aims to expand the scope of VA calibration to be applicable to calibrating VA studies in other LMICs. Specifically, we propose developing and applying statistical and machine learning methods that accounts for country-specific heterogeneity and uncertainty in the algorithmic biases of the AI algorithms used for obtaining COD from VA data.

The proposed approach will be used to calibrate VA data from over 300 studies spanning over 80 high-mortality countries. Beyond leading to improved national, regional, and global child and neonatal cause specific estimates, and enabling health policy, the work would also offer a deeper understanding of the underlying causes of death among children. These insights are vital for developing targeted interventions that can save children lives and guide global health policies.

The proposal aims to accomplish the following:

Aim 1: Unidirectional uncertainty propagation when using AI algorithm outputs in downstream analysis

Aim 2: Generating bias- and uncertainty-corrected estimates of child cause-specific mortality in high-mortality countries.