Country-specific estimates of misclassification rates of computer-coded verbal autopsy algorithms

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Computer-coded verbal autopsy (CCVA) algorithms are crucial for determining cause of death (COD) and estimating cause-specific mortality fractions (CSMFs).
• Frequent COD misclassification by CCVA algorithms introduces bias into CSMF estimates.
• Existing VA-calibration frameworks may overlook systematic patterns and cross-country variations, limiting accuracy.

Purpose of the Study:

• To develop and validate a country-specific VA-calibration framework to reduce COD misclassification bias.
• To estimate and inventory CCVA algorithm misclassification rates across different countries and age groups.
• To improve the accuracy of population-level mortality estimates derived from verbal autopsy data.

Main Methods:

• Utilized CHAMPS (Child Health and Mortality Prevention Surveillance) data to estimate misclassification rates for three CCVA algorithms (Expert Algorithm VA, InSilicoVA, InterVA).
• Estimated rates for two age groups (neonates and children aged 1-59 months) across eight countries.
• Applied Mozambique-specific rates to calibrate VA-only data from the COMSA project.

Main Results:

• Country-specific models demonstrated a better fit to misclassification rates compared to homogeneous models, reducing average absolute loss significantly for both neonates and children.
• Consistent misclassification patterns were observed across CCVA algorithms, with systematic over- or underestimation of certain causes.
• Calibration of COMSA data led to notable shifts in neonatal and child CSMFs, including increases for sepsis/meningitis/infection and malaria, and decreases for intrapartum-related events, prematurity, and pneumonia.

Conclusions:

• An inventory of VA misclassification rates across age groups, algorithms, and countries is now available, enabling broader calibration of VA-only data.
• The study identified systematic algorithmic biases in CCVA, highlighting areas for future algorithm refinement.
• The integrated VA-calibration workflow, supported by open-source software, represents a significant advancement in improving the accuracy of mortality surveillance systems globally.