Comparison of Deep Learning Methods in Predicting Diabetes Mellitus based on an On-Line Dataset

Diabetes Mellitus is a complex metabolic syndrome that needs prompt and correct categorization to avoid serious complications. In this paper, a deep learning method has been proposed to predict the development of Diabetes Mellitus through laboratory and demographic data of patients. An improved Temporal Convolutional Network (TCN) model is designed, and its architecture is refined with the additions of batch normalization, dropout regularization and fine-tuning of the convolutional layers. The data, including clinical biomarkers, such as HbA1c, FPG, OGTT, and demographic characteristics, is thoroughly preprocesses (imputation, normalization, encoding, feature engineering, etc.). The suggested model has an accuracy of 99 percent that surpasses the traditional CNN, RNN, and DNN architectures, and it has good generalization and explanation capabilities. Its robustness is confirmed by comparative analysis and such performance measures as the confusion matrix, precision-recall curve, and ROC-AUC. The effectiveness of deep learning, and especially TCNs, in temporal pattern recognition and risk stratification is highlighted in this work as a clinically relevant way to solve the problem of diabetes management and early intervention.