Enhancing Diabetic Prediction with Hybrid Supervised and Semi-Supervised Models

Predicting diabetic conditions is critical for early intervention and management, driving interest in advanced machine learning techniques. This research enhances diabetic prediction accuracy by combining supervised and semi-supervised learning methods. A diverse set of machine learning algorithms—Decision Tree, Naive Bayes, AdaBoost, Random Forest, and XGBoost—were employed within a semi-supervised learning framework to capture different aspects of the dataset's characteristics. Pseudo-labeling was applied to extend the training set by predicting labels for unlabeled instances, thereby augmenting the informativeness of the dataset. Additionally, feature engineering techniques were utilized to enhance the quality of input features. Two distinct semi-supervised learning models, Self-Training and Co-Training, were subsequently utilized to iteratively refine predictions. The findings demonstrated the effectiveness of each model and provided a nuanced understanding of their performance metrics. The Random Forest model, both in supervised and Self-Training frameworks, exhibited high accuracy and balanced precision, recall, and F1 Score metrics. Specifically, Self-Training with Random Forest and XGBoost achieved accuracies of 0.97, maintaining high precision, recall, and F1 Score metrics. Co-Training, while slightly reducing accuracy, demonstrated robust performance enhancements, with models like Random Forest and XGBoost achieving accuracies of 0.98. Interestingly, the Decision Tree model showcased notable performance improvements through pseudo-labeling, with Self-Training yielding results comparable to the supervised model. Naive Bayes and AdaBoost models, known for their distinctive characteristics, showed consistent accuracy across supervised and semi-supervised approaches, reflecting their stability. This study highlights the potential for integrating domain-specific knowledge into the machine learning pipeline to further enhance predictive performance. The results underscore the utility of leveraging additional unlabeled data in refining predictive models for medical diagnostics, particularly for diabetic condition assessment.