A Stacked Meta-Model Framework for Diabetes Prediction: From Effective Feature Engineering to Meta-Learning

Diabetes datasets often suffer from limited sample sizes, imbalance, and inconsistencies, making it challenging to develop accurate predictive models. Most existing works concentrate on improvement of prediction performance but do not directly incorporate the feature engineering into the process of model training. The majority of research efforts also utilize homogeneous algorithms such as all tree-based or all neural network models, restricting the ability to explore a broad range of data patterns. To address these limitations, we propose a stacked meta-model approach that introduces a Random Forest meta-learner. This meta-learner goes beyond basic aggregation techniques by training on meta-features derived from diverse base models. The idea is that, the predictions of these base models are used as inputs for the meta-learner and therefore allowing the meta-learner to learn from more enriched features that embody complex and abstract patterns that provide better prediction results. Furthermore, we performed extensive feature engineering and data balancing to fine-tune the input data which definitely aided a lot in the training process of our models making them robust as well. Our experimental results show that the proposed meta-learner approach outperformed existing classification models, demonstrating the effectiveness of our feature engineering techniques and the higher predictive performance of the stacked meta-model framework