Evaluation of neonatal verbal autopsy using physician review versus algorithm-based cause-of-death assignment in rural Nepal

Verbal autopsy (VA) is used to ascertain cause-specific neonatal mortality using parental/familial recall. We sought to compare agreement between causes of death obtained from the VA by physician review vs. computer-based algorithms. Data were drawn from a cluster-randomised trial involving 4130 live-born infants and 167 neonatal deaths in the rural Sarlahi District of Nepal. We examined the agreement between causes ascertained by physician review and algorithm assignment by the kappa (κ) statistic. We also compared responses to identical questions posed posthumously during neonatal VA interviews with those obtained during maternal interviews and clinical examinations regarding condition of newborns soon after birth.

Physician reviewers assigned prematurity or acute lower respiratory infection (ALRI) as causes of 48% of neonatal deaths; 41% were assigned as uncertain. The algorithm approach assigned sepsis (52%), ALRI (31%), birth asphyxia (29%), and prematurity (24%) as the most common causes of neonatal death. Physician review and algorithm assignment of causes of death showed high κ for prematurity (0.73), diarrhoea (0.81) and ALRI (0.68), but was low for congenital malformation (0.44), birth asphyxia (0.17) and sepsis (0.00). Sensitivity and specificity of VA interview questions varied by symptom, with positive predictive values ranging from 50% to 100%, when compared with maternal interviews and examinations of neonates soon after birth. Analysis of the VA data by physician review and computer-based algorithms yielded disparate results for some causes but not for others. We recommend an analysis technique that combines both methods, and further validation studies to improve performance of the VA for assigning causes of neonatal death.