Enhancing diabetes risk prediction through focal active learning and machine learning models

To improve the effectiveness of diabetes risk prediction, this study proposes a novel method based on focal active learning strategies combined with machine learning models. Existing machine learning models often suffer from poor performance on imbalanced medical datasets, where minority class instances such as diabetic cases are underrepresented. Our proposed Focal Active Learning method selectively samples informative instances to mitigate this imbalance, leading to better prediction outcomes with fewer labeled samples. The method integrates SHAP (SHapley Additive Explanations) to quantify feature importance and applies attention mechanisms to dynamically adjust feature weights, enhancing model interpretability and performance in predicting diabetes risk. To address the issue of imbalanced classification in diabetes datasets, we employed a clustering-based method to identify representative data points (called foci), and iteratively constructed a smaller labeled dataset (sub-pool) around them using similarity-based sampling. This method aims to overcome common challenges, such as poor performance on minority classes and limited generalization, by enabling more efficient data utilization and reducing labeling costs. The experimental results demonstrated that our approach significantly improved the evaluation metrics for diabetes risk prediction, achieving an accuracy of 97.41% and a recall rate of 94.70%, clearly outperforming traditional models that typically achieve 95% accuracy and 92% recall. Additionally, the model's generalization ability was further validated on the public PIMA Indians Diabetes DataBase, outperforming traditional models in both accuracy and recall. This approach can enhance early diabetes screening in clinical settings, helping healthcare professionals reduce diagnostic errors and optimize resource allocation.