Optimizing Diabetes Detection with ML and DL: Performance Evaluation on Large-Scale Patient Data

This study presents a comprehensive approach to diabetes prediction by leveraging both Machine Learning (ML) and Deep Learning (DL) techniques on a large-scale clinical and demographic dataset consisting of $\mathbf{1 0 0, 0 0 0}$ patient records. The dataset includes key attributes such as age, gender, BMI, blood glucose level, HbA1c level, smoking history, hypertension, and geographical location. A wide range of models, including XGBoost, Random Forest, LightGBM, KNN, Gradient Boosting, Deep Neural Network (DNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), were developed and rigorously evaluated using performance metrics such as accuracy, precision, recall, and F1-score. Among these, LightGBM emerged as the most balanced ML model in terms of overall predictive power, while GRU demonstrated the highest performance among DL models in detecting diabetic cases, particularly excelling in recall and F1-score. In addition to performance analysis, this study also explores correlation analysis, feature importance using multiple selection techniques, and cross-validation to ensure robustness and generalizability of the models. The results highlight the critical impact of features like BMI, blood glucose, HbA1c, and age on diabetes prediction. Furthermore, detailed preprocessing steps such as label encoding, one-hot encoding, and handling categorical variables were applied to optimize model input. This research not only validates the efficacy of current ML/DL models in healthcare data analysis but also provides insights into future enhancements through feature engineering, hyperparameter tuning, and temporal pattern analysis. The findings reinforce the potential of data-driven models in supporting early diagnosis and intervention strategies for diabetes at a large scale.