Challenges in Predicting Children’s Anthropometric Outcomes: Addressing Class Imbalance in Machine Learning Models

This prognostic analysis uses the data collected in 2023 for the WHO European Childhood Obesity / Growth Surveillance Initiative (COSI) study. This is an international study of first-grade children coordinated by the WHO. From the collected data set, subjects were selected and assigned to the normal BMI group (n = 2176(77.5%)) and elevated body weight (n = 633(22.5%)). Binary classification was performed using the Weka 3.8.5 machine learning software package. Understanding the relationship between children’s anthropometric measures and food consumption patterns is crucial for public health interventions. 18 different (e.g. vegetables, fruits, fast food, sweets, etc.) types of food products were analyzed, trying to build models for prognosis of elevated body weight (reference category). However, modeling this relationship presented challenges due to severe imbalance of class categories, where the reference category has a little more than 20% of the dataset. Binary logistic regression was initially applied, yielding poor model performance (R2 = 0.0023, precision, recall and F-measure equal to 0%), indicating the model’s inability to capture meaningful patterns. To address class imbalance, we applied Weka’s Class Balancer and evaluated multiple classification algorithms. Logistic regression performance remained low (precision = 53.8%, recall = 48.2%, F-measure = 50.8%), as did Decision Stump (precision = 55.3%, recall = 46.1%, Fmeasure = 50.3%) and Na¨ıve Bayes (precision = 54.7%, recall = 47.6%, F-measure = 50.9%). Similarly, Hoeffding Tree (precision = 54.7%, recall = 47.6%, F-measure = 50.9%) showed no substantial improvement. Notably, the baseline ZeroR classifier outperformed other models in recall (69.7%) and F-measure (58.1%), underscoring the difficulty of extracting meaningful predictive patterns from the dataset. These findings highlight the limitations of conventional machine learning methods in highly imbalanced biomedical datasets and suggest the need for alternative approaches, such as ensemble models, cost-sensitive learning, or deep learning techniques. Future research should focus on advanced resampling strategies and domain-specific feature engineering to enhance predictive accuracy in child growth assessments. The findings on public health perspective could be as follows. Lower consumption of vegetables and fruits is associated with an increased likelihood of having elevated body weight. Higher consumption of flavored milk and sugar-sweeten soft drinks is linked to a greater risk of elevated body weight. Also, modeling doesn’t include other possible important predictors such as physical activity, snacking or socioeconomic factors.