Effective treatment for type 2 diabetes is dependent on early identification of this hazardous disease. Predicting it is challenging due to the need for improved accuracy and the rapid increase in health data, making traditional methods insufficient. This research addresses these challenges with targeted Deep Learning (DL) based solutions for feature extraction and prediction accuracy. The proposed model is Topological Data Analysis with Convolutional neural network and Attention based Gated recurrent unit (TDA-CAG). Topological Data Analysis (TDA) effectively extracts topological characteristics of health time series data which are more relevant features for predicting type 2 diabetes. Additionally, Convolutional Neural Network (CNN) and attention mechanism based Gated Recurrent Unit (GRU) are used for prediction using extracted feature sets that raises the suggested TDA-CAG model's prediction accuracy above that of benchmark models. Evaluating the Pima Indians Diabetes dataset, the proposed TDA-CAG method improves Precision by 12.97%, Sensitivity by 8.56%, Specificity by 3.99%, and Accuracy by 6.15% over existing Type 2 diabetes prediction techniques. It also achieves up to 18% lower RMSE and 45.78% lower MSE compared to existing methods