Modelling the impact of dietary diversity on child nutrition in Pakistan: a machine learning analysis with Shapley...

Pakistan faces a serious challenge in malnutrition, as its core nutrition indicators and minimum dietary diversity remain below the recommended level. Here, we investigate the association between the food/dietary diversity index and the nutritional status of children under five years of age and identify influential factors associated with child nutritional status. We hypothesise that poor dietary diversity, i.e. the consumption of fewer than five food groups, is a risk factor, while the consumption of five or more groups is a protective factor for child nutrition.

We used national representative survey data on 4499 children from the 2018 Pakistan Demographic and Health Survey and, through the application of a hybrid machine learning framework for mixed mechanisms, analysed the relationship between the predictive model's performance and its indicators using machine learning-based logistic regression (ML-LR). We used Shapley Additive exPlanations (SHAP) to assess the model's characteristics and selected key risk factors through a Boruta algorithm.

We observed that stunting (38.13%), underweight (23.04%), and wasting (8.05%) remain widespread in Pakistan. The ML-LR model identified living in poor areas (Sindh, Balochistan, and federally administered tribes), high birth order, older age, low dietary diversity, recent diarrhoea cases, and maternal unemployment as important risk factors. The SHAP analysis identified the marginal effects of each predictor and confirmed that child age, maternal underweight, and unimproved water were the main risk drivers, while adequate dietary diversity (over five food groups), higher maternal education level and male gender were protective factors. The Boruta algorithm identified low dietary intake, the child's higher age, and the mother's low nutritional status as the most important determinants among 13 selected factors.stunting (38.13%), underweight (23.04%), and wasting (8.05%) remain widespread in Pakistan. The ML-LR model identified living in poor areas (Sindh, Balochistan, and federally administered tribes), high birth order, older age, low dietary diversity, recent diarrhoea cases, and maternal unemployment as important risk factors for child nutrition. In this sense, the SHAP analysis identified the marginal effects of each predictor and confirmed that child age, maternal underweight, and unimproved water as the main risk drivers for child nutrition, while adequate dietary diversity (over five food groups), higher maternal education level, and male gender were protective factors. The Boruta algorithm identified low dietary intake, the child's higher age, and the mother's low nutritional status as the most important determinants among 13 selected factors.

We found that dietary diversity is a key threshold in shaping child nutritional status. Children consuming fewer than five food groups were shown to be at higher risk of malnutrition, while those consuming five or more food groups had improved nutritional outcomes. Public health interventions should prioritise strategies to ensure that this food diversity threshold is met, while also addressing potential socioeconomic constraints. As a secondary finding, we suggest that a combination of ML-LR, SHAP, and Boruta provides a robust, explanatory, and replicatory analytical framework for research in nutrition epidemiology and public health through predicting malnutrition, assessing its characteristics, and identifying its key predictive factors.