Early warning signs: targeting neonatal and infant mortality using machine learning

Abstract

ABSTRACT This article uses a nation-wide household survey data from India and identifies important predictors of neonatal and infant mortality using multiple machine learning (ML) techniques. The consensus on the leading predictors from the interpretable ML algorithms (that we use) serve as early warning signs of neonatal and infant mortality. This enables us to identify a ‘high-mortality risk’ group of mothers and infants – an important goal of India’s ‘India Newborn Action Plan’. This high-risk group comprises firstborns, mothers with prior deaths or several previous births, newborns suffering from complicated deliveries, small size at birth and unvaccinated infants. We identify early newborn care, folic acid supplements and conditional cash transfer (Janani Suraksha Yojana) as the most effective policy interventions. Given the imbalanced nature of the dependent variable (‘events’ being rarer than ‘non-events’) we use additional ML methods (along with the commonly used ones) that are tailor-made for ‘rare-event’ prediction for robustness checks. We also use an evaluation measure called Area under Precision Recall Curve that is tailored for measuring prediction accuracy with imbalanced data. Our analysis sheds light on policy relevance and suggests some new policy prescriptions such as close monitoring of at-risks babies including females and those with small birth-size.