Diabetes is a chronic metabolic disorder for which early diagnosis is highly important in preventing serious complications and improving clinical decision-making. In this study, a hybrid deep learning framework, termed ICHOA-CNN, is proposed, in which an improved cheetah optimization algorithm is employed to automatically optimize the hyperparameters of a one-dimensional convolutional neural network (1D-CNN). The proposed approach aims to enhance the search process in the hyperparameter space, effectively extract nonlinear patterns from the tabular PIMA dataset, and improve diabetes diagnosis performance. For a fair evaluation, the dataset was stratified and divided into training/validation and independent test sets; model optimization was conducted using 10-fold cross-validation on the training/validation set, while final evaluation was performed solely on the independent test set. The proposed method was compared with the baseline CNN, PSO-CNN, CHOA-CNN, and conventional machine learning methods, including SVM, MLP, and XGBoost. The results demonstrated that ICHOA-CNN outperformed all compared methods, achieving an accuracy of 0.994, a precision of 0.981, a recall of 1.000, and an F1-score of 0.991. Moreover, an AUC of 0.9896 and the results of paired statistical tests with p<0.001confirmed the stability and statistical superiority of the proposed model. This framework can be considered an effective approach for developing intelligent screening and clinical decision-support systems for diabetes diagnosis.