Unlocking the Power of Machine Learning for Diabetes Prediction: A Comparative Evaluation

Abstract

Diabetes is a chronic metabolic condition with severe long-term health consequences, making early and accurate detection crucial to preventing complications. While traditional diagnostic methods are reliable, they can be time-consuming and less effective in detecting diabetes at early stages. This study applies machine learning (ML) techniques to develop a predictive model for diabetes using the PIMA Indian Diabetes Dataset, which includes critical clinical and lifestyle factors like glucose levels, BMI, and age. The data was cleaned to address missing values, normalized using feature scaling, and divided into training (80%) and testing (20%) sets. A Logistic Regression model was selected for its interpretability and efficiency in binary classification tasks. The model achieved an accuracy of 79.3%, with precision, recall, and F1-scores of 76.5%, 74.1%, and 75.3%, respectively. A comparative analysis with other ML models (such as Random Forest and XGBoost) showed a solid balance between performance and simplicity. Key predictors included glucose levels and BMI, which align with clinical insights. Although the model demonstrates potential for early diabetes detection, challenges like class imbalance and false negatives highlight the need for improvements, potentially through oversampling techniques or more advanced algorithms. This work showcases the potential of ML to improve diabetes diagnostic tools, offering a scalable, data-driven solution to enhance healthcare decision-making. Future research could focus on exploring deep learning approaches and larger datasets to further boost accuracy and generalization.