<![CDATA[Diabetes mellitus is a chronic disease with a high prevalence that requires accurate and efficient early detection. The process of identifying diabetes risk at Hasyim Asy’ari Hospital, which is still carried out conventionally, is considered ineffective in handling large volumes of medical records, thereby potentially slowing down patient care. This study developed a web-based diabetes mellitus prediction system using the Random Forest algorithm with a configuration of 100 decision trees (n_estimators=100, random_state=42). The dataset consists of 900 medical records from patients at Hasyim Asy’ari Hospital (630 positive cases and 270 negative cases of diabetes), which were divided in a 70:30 ratio into 630 training data and 270 test data using a stratified split. The features used include Fasting Blood Glucose, HbA1c, Random Blood Glucose, Systolic Blood Pressure, BMI, Age, Diastolic Blood Pressure, and Gender. The system was implemented using PHP 8.2, Python 3.11.4 (scikit-learn 1.3.0), and Tailwind CSS 3.0. Evaluation results on the test data show an accuracy of 96.67% (95% CI: 94.12–98.89%), precision of 98.85% (95% CI: 96.55–100.00%), recall of 96.63% (95% CI: 92.86–100. 00%), and an F1-score of 97.73% (95% CI: 95.24–99.42%), calculated via bootstrap resampling with 1,000 iterations. This performance significantly outperformed SVM (accuracy 91.00%) and Naive Bayes (precision 89.00%) based on the McNemar test (α=0.05). Feature importance analysis using Mean Decrease in Impurity (MDI), cross-validated with Permutation Importance (n_repeats=30), revealed that Fasting Blood Glucose (19.60%), HbA1c (15.78%), and Random Blood Glucose (15.46%) were the three most influential indicators, consistent with the WHO and ADA diagnostic criteria.]]>
