Prediction models for stunting at 2-years-old from Indonesian newborn population

Stunting in children is a health problem, especially in developing countries, such as Indonesia. The lack of information-based early preventive measures resulted in an insignificant reduction in stunting. This study aimed to develop a prediction model for stunting at 2-years old in an Indonesian newborn population.

Various machine learning algorithms as the core of artificial intelligence technology, under the Cross-Industry Standard Process for Data Mining (CRISP-DM) conceptual framework, were used to build a prediction model using data of 5093 children with 23 predictor variables from the Indonesian Family Life Survey (IFLS) open database. Model prediction performance was evaluated using F1 scores, area under the receiver operating characteristic curve (AUC), sensitivity, and accuracy. Confusion matrices were used to calculate the positive and negative predictive values and evaluate the implications of the final prediction model in public health and clinical practice. Explanatory-risk factor model was developed using multivariable logistic regression.

The best model to predict stunting statistically, which contains the six best predictor variables, is k-nearest neighbor (kNN) with an F1 value of 84.5%, compared to random forest, neural network, decision tree, and naïve Bayes with F1 values of 80.5%, 71.2%, 68.7%, and 65.8%, respectively. The positive and negative predicted values for the final model were 80.4% (78.5-82.3% 95% CI) and 86.8% (85.7-87.9% 95% CI), respectively. Factors associated with stunting are birth weight (AOR = 0.79,p-value < 0.01), large infant size (AOR = 0.24,p-value < 0.05), mothers age (AOR = 0.98, p-value < 0.01), mothers height (AOR = 0.91, p-value < 0.01), fathers height (AOR = 0.95, p-value < 0.01), mothers lower education level(AOR = 1.50, p-value < 0.05), birth at health facilities (AOR = 0.83, p-value < 0.05), toilet standard(AOR = 0.86,p-value < 0.05), and waste disposal standard (AOR = 0.79 ,p-value < 0.01).

A machine learning-based stunting prediction model effectively identifies a high-risk newborn population and enables early targeted interventions. Integration of the prediction with a prescriptive approach that considers causal pathways and evidence-based interventions is required for precise and sustainable prevention of stunting.