Gestational diabetes is a severe condition in which pregnant women are affected and which causes difficulties to the mother and the child if it is not diagnosed. Identification of the threat of evolving diabetes early in pregnancy is important to trigger early intervention and increase maternal healthcare needs. The paper presents a deep learning (DL) model based on Deep Belief Networks (DBN) for binary classification of the risk of diabetes in pregnant females using the PIMA Indian Diabetes Dataset. Three preprocessing methods used to ameliorate the results in terms of prediction performance were, namely the Interquartile Range (IQR)-based outlier removal, the Robust Scaler (RS), and the Min-Max normalization. All the methods were applied in preprocessing the data before training the DBN model. Four major metrics performance evaluation was applied including accuracy, precision, recall, and log loss. All combinations were tested but the Min-Max+DBN combination yielded the closest and consistent results thereby indicating the significance of proper preprocessing in DL pipelines. Although the performance of DBN together with IQR and RS was also decent, it was not as effective as Min-Max normalization. The results affirm that medical prediction problems require preprocessing to be very crucial. The study identifies the possibility of DL models particularly DBNs, in the development of reliable and efficient clinical decision support systems to help in the assessment of diabetes risk.