Diabetes is a global health challenge, requiring early detection to mitigate severe complications. This study explores a deep learning (DL) approach for predicting diabetes stages, incorporating social determinants of health (SDOH) and medical indicators. Using a multiclass dataset, the study addresses class imbalance through SMOTE-Tomek resampling. It adapts non-structural classifiers, Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) for structured data analysis. The study compares these models against a Feedforward Neural Network (FNN) baseline. Results indicate that while accuracy declined post-resampling, minority class predictions improved, enhancing model fairness. Notably, LSTM demonstrated the highest post-resampling accuracy (77.27%). This study advances DL applications in healthcare by integrating SDOH and reconfiguring CNN and LSTM for structured data, expanding their utility beyond traditional domains. These findings contribute to unbiased, clinically relevant diabetes prediction, aligning with Sustainable Development Goal 3 to promote well-being for all.