Identifying the ideal deep learning algorithm to quantify the threshold effect of parental smoking on child...

Accurate identification of modifiable risk factors for child malnutrition is of great importance in the formulation of policies to protect children's health in South Asia. Current research has developed an innovative analytical system that uses deep learning algorithms to accurately assess the nutritional status of children and its relevance to indoor parental smoking pollution, at threshold level in South Asia.

Data on 219,168 under-five children were analyzed from recent (2016-2022) nationally representative Demographic and Health Survey (DHS) for five South Asian countries: Bangladesh, India, Maldives, Nepal, and Pakistan. Our method first applied a comprehensive pre-processing and feature engineering pipeline to detect key risk patterns.

By conducting a thorough benchmark of 16 deep learning models, the Bayesian Neural Network (BNN) was identified as the optimal model for predictive inference and risk quantification. The BNN analysis, supported by Mesh query graphs, showed strong co-association of parental exposure (tobacco smoking) and child undernutrition. Moreover, the association was dose-dependent in that predicted risk for child malnutrition increased substantially when the frequency of cigarette smoking exceeded ten cigarettes per day. This study importantly found that the marginal risk of malnutrition increases by 3.2 times with additional consumption of cigarette after threshold.

There is a strong joint association between child malnutrition and parental smoking. Additionally, less than equal to 10 cigarettes considered the threshold smoking level, greater than 10 cigarettes per day higher the risk of malnutrition by 3.2% with additional cigarette consumption. This evidence provides considerable leverage for policymakers as our results suggest that modern AI methods can effectively inform interventions targeted at increasing the prevalence of smoke-free homes and improving child nutrition in South Asia.

Not applicable. This study is a secondary analysis of publicly available, de-identified data from the Demographic and Health Survey, and does not report the results of a prospective health care intervention.