Diabetes prediction using machine learning and explainable AI techniques

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Public Health

Background:

• Diabetes Mellitus is a major global non-communicable disease affecting over 537 million people.
• Early prediction of diabetes is vital to mitigate severe health complications such as heart disease, kidney disease, and retinopathy.
• Risk factors include obesity, abnormal cholesterol, family history, inactivity, and poor diet, with increased urination as a common symptom.

Purpose of the Study:

• To develop an automated diabetes prediction system utilizing machine learning techniques.
• To evaluate the performance of various classification algorithms on a private dataset of female patients in Bangladesh.
• To enhance model generalizability and interpretability for practical application.

Main Methods:

• A hybrid dataset combining the Pima Indian diabetes dataset with new samples from 203 Bangladeshi female factory workers was used.
• Feature selection was performed using the mutual information algorithm.
• Class imbalance was addressed using SMOTE and ADASYN techniques.
• Multiple machine learning classifiers (Decision Tree, SVM, Random Forest, Logistic Regression, KNN) and ensemble methods were trained and tested.
• Extreme Gradient Boosting (XGBoost) with ADASYN was employed for prediction.
• Domain adaptation, LIME, and SHAP frameworks were implemented for model versatility and explainability.

Main Results:

• The XGBoost classifier combined with the ADASYN approach achieved the highest prediction accuracy of 81%.
• The model demonstrated strong performance with an F1 score of 0.81 and an Area Under the Curve (AUC) of 0.84.
• Explainable AI methods (LIME, SHAP) provided insights into the model's prediction mechanisms.
• A website and an Android application were developed for real-time diabetes prediction.

Conclusions:

• The developed machine learning system, particularly the XGBoost model with ADASYN, offers a highly accurate and explainable method for early diabetes prediction.
• The system's versatility is demonstrated through domain adaptation, suggesting potential for broader application.
• The creation of user-friendly interfaces (web and mobile) facilitates accessible diabetes risk assessment.