Use of Classification Algorithms in the Determination of Diabetes Level and Drug Recommendation Process

Diabetes mellitus (DM) is a chronic metabolic disorder affecting millions worldwide, necessitating accurate diagnosis and effective treatment strategies. Traditional diagnostic approaches, including fasting blood glucose (FBG) and hemoglobin A1c (HbA1c) tests, have limitations in predicting disease progression and addressing complex drug interactions. In this study, machine learning (ML) and deep learning (DL) models were utilized to enhance both diabetes classification and drug interaction risk assessment. Specifically, the Pima Indian Diabetes Dataset was employed for diabetes diagnosis, utilizing eight clinical features selected through Mutual Information (MI). To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied. Among machine learning models, Random Forest (RF) achieved the highest accuracy of 98.7%, while LSTM networks performed best among deep learning models, reaching 97.2% accuracy. In the drug interaction risk classification task, a custom dataset based on clinical guidelines from the Turkish Endocrinology Society was used. This dataset, comprising 109 records, details pairwise drug interactions, pharmacokinetic mechanisms, and risk levels (moderate or high). An ensemble model combining Random Forest and Decision Tree classifiers, utilizing a weighted voting mechanism, achieved 100% accuracy. However, the small dataset size underscores the need for further validation with larger, more diverse datasets. Comparative analysis demonstrated that ensemble learning and deep learning models outperform conventional classification techniques, reinforcing their potential in clinical decision support systems. Future work will focus on expanding the dataset, integrating additional patient-specific parameters, and optimizing model generalizability for real-world applications.