Predictive Analysis of Diabetes-Risk with Class Imbalance

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Computational Biology

Background:

• Type 2 diabetes (T2DM) is a prevalent chronic condition with significant health risks due to hyperglycemia.
• Early and accurate diagnosis of T2DM is crucial for managing complications and improving patient outcomes.
• Imbalanced datasets pose a challenge for developing effective predictive models for T2DM.

Purpose of the Study:

• To develop a high-classification-accuracy machine learning model for predicting type 2 diabetes.
• To evaluate the effectiveness of data balancing techniques in improving T2DM prediction.
• To compare the performance of different machine learning algorithms for T2DM classification.

Main Methods:

• Utilized advanced machine learning techniques for T2DM prediction.
• Compared oversampling (SMOTE) and undersampling (Tomek-links) strategies on imbalanced datasets.
• Evaluated an Artificial Neural Network (ANN) model against Support Vector Machine (SVM), Random Forest (RF), and Decision Tree (DT) classifiers.

Main Results:

• The tuned ANN model achieved a peak accuracy of 92.2%.
• The Synthetic Minority Oversampling (SMOTE) technique significantly improved accuracy, AUC, and F1-measure compared to baseline and undersampling methods.
• The proposed model demonstrated superior performance in classifying the minority class of T2DM.

Conclusions:

• Machine learning, particularly with SMOTE, offers a powerful approach for accurate T2DM prediction.
• Data balancing techniques are essential for building robust predictive models for imbalanced health datasets.
• Dynamic hyperparameter optimization is recommended for further enhancing model accuracy in T2DM prediction.