<![CDATA[Type 2 Diabetes Mellitus (T2DM) is frequently undetected at an early stage because its symptoms develop gradually. Therefore, an accessible risk-screening method is needed to support early awareness and further medical examination. This study developed an Adaptive Neuro-Fuzzy Inference System with a Takagi-Sugeno-Kang structure for three-class T2DM risk prediction using eight non-laboratory health indicators from the Behavioral Risk Factor Surveillance System dataset. After conflict removal, 65,369 records were divided into stratified training, validation, and testing sets. SMOTENC was applied only to the training set to reduce class imbalance. To provide a fair evaluation, ANFIS was compared with Logistic Regression, Decision Tree, Random Forest, XGBoost, Linear Support Vector Machine, Multilayer Perceptron, and a majority-class predictor under the same experimental protocol. Macro-F1 was used as the primary evaluation metric because of the severe class imbalance. A sensitivity analysis of Top-K values of 128, 256, and 512 were also conducted, and permutation importance was used to examine the contribution of each input variable. The final ANFIS model using Top-K = 128 achieved an accuracy of 84.18%, a Macro-F1 of 48.59%, a weighted F1-score of 88.06%, and a balanced accuracy of 62.85%. Random Forest achieved the highest Macro-F1 of 50.98%, indicating that ANFIS did not outperform the strongest baseline in overall class-balanced performance. However, ANFIS achieved the highest balanced accuracy and the highest prediabetes recall among the evaluated models. BMI, GenHlth, and Age were the most influential variables according to permutation importance. The selected model was implemented in a Django-based web expert system as a preliminary risk-screening prototype rather than a medical diagnostic tool. Keywords: ANFIS, Type 2 Diabetes Mellitus, Expert System, Risk Prediction]]>
