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

Area of Science:

• Medical Informatics
• Machine Learning
• Data Science

Background:

• Machine learning models struggle with imbalanced medical datasets, leading to underperformance in predicting minority classes like diabetes.
• Effective diabetes risk prediction is crucial for early intervention and patient management.

Purpose of the Study:

• To develop a novel method, Focal Active Learning, to enhance the effectiveness of diabetes risk prediction.
• To address the challenges of imbalanced datasets and improve model performance and interpretability.

Main Methods:

• Implemented a Focal Active Learning strategy combined with machine learning models.
• Utilized SHAP (SHapley Additive Explanations) for feature importance quantification and attention mechanisms for dynamic feature weighting.
• Employed a clustering-based method to identify data foci and construct a smaller, representative labeled dataset via similarity-based sampling.

Main Results:

• Achieved a 97.41% accuracy and 94.70% recall rate, significantly outperforming traditional models (95% accuracy, 92% recall).
• Demonstrated superior generalization ability on the PIMA Indians Diabetes DataBase.
• Validated the effectiveness of Focal Active Learning in mitigating class imbalance and improving prediction outcomes.

Conclusions:

• The proposed Focal Active Learning method offers a significant advancement in diabetes risk prediction.
• This approach enhances early diabetes screening, reduces diagnostic errors, and optimizes resource allocation in clinical settings.
• The method provides a more efficient and interpretable solution for analyzing imbalanced medical data.