This study uses a hybrid CNN-LSTM architecture to create a unique medical decision model for diabetes prediction. Using deep learning algorithms has several advantages, such as improved decision-making skills, flexibility, and accurate outcomes. The recommended methodology consists of three primary processes: initial processing, selection of features, and classifier. During preprocessing phase, we address common challenges in clinical data, such as missing information and inconsistencies, to ensure data quality and integrity. Secondly, wrapper-based methods are utilised to find the most pertinent characteristics from the diabetes dataset during the feature selection phase, guaranteeing that only noteworthy features are included in the final model. Finally, accurate diagnosis of diabetes is achieved in the classification step by using the combined CNN-LSTM architecture. The approach makes use of long short-term memory (LSTM) networks’ capacity for temporal sequence acquisition as well as convolutional neural networks’ (CNN) capacity for spatial recognition of features. By obtaining an accuracy rate of 96.75%, a precision of 0.94, and a recall of 0.92, the suggested CNN-LSTM model greatly surpasses conventional machine learning models, according to comparative analysis. This enhancement leads to lower labour costs, faster processing times, and better decision accuracy. The algorithm performs exceptionally well in diabetes predictions when tested against the Pima Indians Diabetes (PID) dataset. Python is used for the entire implementation.