Automated versus physician assignment of cause of death for verbal autopsies: randomized trial of 9374 deaths in 117 villages in India

Table 2 Percent population-level concordance in cause of death distribution between automated assignment and standard (physician assignment) verbal autopsies, by algorithms and age groups

Average and standard deviation (SD) of the population-level concordance attained for the automated algorithms when using data from all PHMRC sites as the training data. The concordance compares the cause of death distributions generated by each algorithm on the 4723 deaths in the automated arm (4393 adult, 213 child, and 117 neonatal deaths) to the distribution on the 4651 standard physician-coded deaths (4311 adult, 190 child, and 150 neonatal deaths). When only the Indian sites were used as the training data, the concordance for NBC, King-Lu, and InSilicoVA was 37, 57, and 68 for adult, 48, 59, and 66 for child, and 23, 76, and 80 for neonatal deaths, respectively. The results were similar if we excluded “ill-defined” deaths (see Additional file 1). InSilicoVA-NT and InterVA-4 do not require training data, whereas SmartVA was pre-trained on the PHMRC data; hence, the percent concordance generated by these algorithms is unchanged when changing the training dataset. Dual physician review of the automated assignment arm generated the population-level concordance of 84, 82, and 91 for adults, child, and neonate age groups, respectively (see Additional file 10)

ISSN: 1741-7015