Machine Learning-Based Prediction of Determinants of Appropriate Complementary Feeding Practices Among Women With Children Aged 6-23 Months in Sub-Saharan Africa.

Globally, children's feeding practices are a major public health concern. Evidence indicates that 90% of children receive less than the bare minimum of dietary content. In developing countries, only one out of five children under 24 months old receives the minimum recommended diet. This study utilized a weighted dataset of 24,235 from the Demographic and Health Survey (DHS) conducted across eight Sub-Saharan African countries. The data preprocessing and analysis were performed using STATA version 17 and Python 3.10. Feature scaling was carried out using MinMax scaling and Standard Scalar, while feature selection was performed through Recursive Feature Elimination (RFE). An 80:20 data split ratio was applied, and class imbalance was addressed using Tomek Links combined with Random Oversampling. Eight models were selected and trained with both balanced and unbalanced datasets. Model performance was evaluated using metrics such as ROC-AUC, accuracy, and the confusion matrix. The overall current feeding practice rate is 9.1%, with a Random Forest model achieving an accuracy of 91% and an AUC of 96%. The predictors for appropriate complementary feeding include current breastfeeding status, maternal education, wealth status, number of household members, sex of the household head, sex of the child, place of delivery, maternal employment status, and distance to the nearest health facility. In conclusion, the Random Forest model was effectively used to identify key determinants, revealing that appropriate complementary feeding practices are low. To improve this, we recommend enhancing community-based nutrition and reproductive health education, providing economic support for the poor, increasing access to healthcare facilities, and creating better opportunities to access smartphones and other social media platforms.