Comparison of causes of stillbirth and child deaths as determined by verbal autopsy and minimally invasive tissue sampling

Abstract

Background: In resource-limited settings where vital registration and medical death certificates are unavailable or incomplete, verbal autopsy (VA) is often used to attribute causes of death (CoD), identify the distribution and trends of diseases, and prioritize resource allocation and interventions. However, VA findings can be non-specific, as this tool is based on family members’ recall of symptoms rather than objective diagnostic testing. We aimed to compare the CoD diagnoses obtained in stillbirths and children below five years of age (<5s) through two very different approaches; namely: 1) VA; and 2) the results obtained through the use of Minimally Invasive Tissue Sampling (MITS) and rigorous diagnostic testing, as part of the approach proposed by the Child Health and Mortality Prevention Surveillance (CHAMPS). Methods: CHAMPS identified stillbirths and deceased children <5s in real time between 2017 and 2021 in catchment areas in seven low- and middle-income countries (LMICs): Bangladesh, Ethiopia, Kenya, Mali, Mozambique, Sierra Leone, and South Africa. Deaths were eligible for MITS if identified <24 hours after death, legal concerns were not present, burial had not occurred, and parents consented. CHAMPS teams utilized information from MITS and VA to determine the causes of death (CoDs); if not eligible for MITS, the InterVA software utilized only VA information to determine the CoDs. CHAMPS attributed CoD using expert panels that reviewed clinical evidence microbiological, and histopathological results from MITS to derive the CoDs (Determination of Cause of Death [DeCoDe]). The InterVA4 package of OpenVA software automatically assigned the underlying CoDs using the Bayesian probabilistic modeling technique. These automatically assigned CoDs from OpenVA were compared to the gold-standard of the CHAMPS-attributed CoDs to evaluate both systems’ agreement, weaknesses, and strengths using Lin’s concordance correlation coefficient. Results: Data from 2852 deaths that underwent MITS were analysed. The most common age categories were stillbirths (n=1075, 37.7%) and neonatal deaths (n=1077, 37.8%). Overall concordance of InterVA4 and DeCoDe in assigning causes of death across surveillance sites, age groups, and causes of death was poor (0.75 with 95% CI: 0.73 – 0.76) and lacked precision. We found substantial differences in agreement among surveillance sites, with Mali showing the lowest and Mozambique and Ethiopia the highest concordance. Lin’s concordance correlation coefficient for children aged < 1 year was 0.69 (95%CI: 0.65 – 0.71), and for children aged 1-4 years was 0.28 (95%CI: 0.19 – 0.37) Conclusion: The InterVA4 assigned CoD agrees poorly in assigning causes of death for under-fives and stillbirths. Because VA methods are relatively easy to implement, such systems could be more useful if algorithms were improved to more accurately reflect causes of death, for example, by calibrating algorithms to information from programs that used detailed diagnostic testing to improve the accuracy of COD determination.