Optimizing diabetes prediction using machine learning: a random forest approach

Abstract

Diabetes, a leading cause of global mortality, is responsible for millions of deaths annually due to complications such as heart disease, kidney failure, and stroke. Projections indicate that 700 million people will be affected by diabetes in 2045, placing immense strain on global healthcare systems. Early detection and accurate prediction of diabetes are essential in mitigating complications and reducing mortality rates. However, existing diabetes prediction frameworks face challenges, including imbalanced datasets, overfitting, inadequate feature selection, insufficient hyperparameter tuning, and lack of comprehensive evaluation metrics. To address these challenges, the proposed random forest diabetes prediction (Random DIP) framework integrates advanced techniques such as hyperparameter tuning, balanced training, and optimized feature selection using a random search cross-validation (RandomizedSearchCV). This framework significantly improves predictive accuracy and ensures reliable clinical applicability. Random DIP achieves 99.4% accuracy, outperforming related works by 7.23%, the area under curve (AUC) of 99.6%, surpassing comparable frameworks by 7.32%, a recall of 100%, exceeding existing models by 9.65%, a precision (97.8%), F1-score (98.9%), and outperformance of 6.69%. These metrics demonstrate Random DIP's excellent capacity to identify diabetes cases while minimizing false negatives (FPs) and providing reliable predictions for clinical use. Future work will focus on integrating real-time clinical data and expanding the framework to accommodate multi-disease prediction for broader healthcare applications.