Globally, diabetes is one of the most dangerous diseases, which is a long-term illness produced through an unproductive utilization of insulin developed through the pancreas. The Internet of Things (IoT) sensor gathers the data from the diabetes patients, and machine leaning (ML) techniques are utilized for the identification of the data as well as prediction of this disease. The transformation of the large volumes of the sensitive medical data, dealing with IoT security problems, remains challenging. Hence, this research proposes the temporal attention layer-based long short-term memory (TAL-LSTM) for the diabetes prediction in IoT systems. Initially, The PIMA dataset is utilized for estimating the effectiveness of the TAL-LSTM approach. Then, the preprocessing is done by Synthetic Minority Over-Sampling Technique (SMOTE) to address class imbalance as well as min–max normalization technique. Then, the principal component analysis (PCA) is utilized for selecting the relevant features. The experimental results show that the propose TAL-LSTM attains the superior accuracy of 99.01% when compared to the state-of-the-art technique like convolutional neural network with LSMT (CNN-LSTM).