<p>Type 2 diabetes mellitus disease is one of the common metabolic disorders, which is caused by the incapability of insulin-sensitive tissues to properly respond to insulin, resulting in blindness, kidney failure, and increased mortality rates. Early screening and diagnosis are crucial for ensuring a healthier lifestyle and minimizing the rates of end-organ damage. Numerous type 2 diabetes mellitus prediction methods have been designed in recent times. Yet, it is not sufficient to continuously capture and differentiate the subtle variation of input data more efficiently, thus enhancing inaccurate outcomes that impact timely diagnosis. Further, it takes a longer training period to deal with the vast amount of diabetes datasets to minimize generalization and improve early decision-making performance, affecting the patient’s quality of life. Moreover, it produces overfitting issues in the validation process while handling the imbalanced nature of diabetes data; thus, it enhances poor performance and delayed treatment. In order to solve the abovementioned complexities, a new multi-class type 2 diabetes mellitus prediction model is developed by employing the deep learning model. Initially, the publicly available Pima Indians Diabetes Database is collected. It helps identify the potential risks for enabling early intervention and taking recommended action. Then, the collected data undergoes preprocessing, which comprises outlier detection and filling, utilized to enhance the quality of data. The preprocessed data is applied to the Adaptive Synthetic Sampling (ADASYN) to resolve the imbalanced datasets. It generates synthetic samples for underrepresented samples. Data balancing performance is improved by the optimization of parameters from ADASYN, which is done with the help of Fitness-based Northern Goshawk Optimization (F-NGO). The balanced data is further passed to the feature extraction stage, where features like <i>t</i>-distributed Stochastic Neighbor Embedding (t-SNE), Principal Component Analysis (PCA), and statistical features are extracted for getting more details from the resultant balanced data. Finally, the multi-class types such as type 2 diabetes-mild, type 2 diabetes-moderate, and type 2 diabetes-severe in type 2 diabetes mellitus prediction is done using the Multi-scale Temporal Convolution Network with Gated Recurrent Unit (MTCN-GRU). It can ultimately analyze and differentiate the different class labels of the type 2 diabetes mellitus stages such as mild, moderate, and severe within a limited duration to improve the patient’s quality of life by minimizing the mortality rates. This integrated deep learning network provides better classification outcomes with higher accuracy over imbalanced datasets. Overall, the performance of the recommended technique shows 86% and 92% in terms of precision and accuracy. The simulation outcome of the statistical analysis of the developed model achieves 0.03%, 0.53%, 1.46%, and 0.15% better than AOA, WHI, WPA, and NGO, respectively, in terms of best measure. Predicting type 2 diabetes mellitus in an earlier stage can significantly impact healthcare by allowing better treatment plans for the individuals. Utilizing the deep learning model to analyze the difficult relationship between the data can ensure a better understanding of the disease and mitigate risk occurrence.</p>