Diabetes and chronic disease prediction has a seminal responsibility in ensuring the provision of early treatment for diabetes patients in order to enhance health stages. It’s the first work which presents GAE-MedBERT—a complex framework derived from the MedBERT design that incorporates methods for processing sequential patient records. Integral to this framework is the integration of the GNN for appropriate modeling of disease dependency based on a weighted comorbidity graph. These are intended to improve data quality and relevance via one-hot encoding of dichotomous variables such as sex and symptoms, and dynamic imputation for numerical variables like age and tidal volume, as well as a comorbidity graph to represent the connectedness of diseases. For contextual embeddings, MedBERT is only fine-tuned on structured clinical data so that the model encapsulates operative medical features. At the same time, the GNN nodes create embeddings that encompass disease associations, which allow the model to identify relationships in the data. To learn vital features within medical timelines, a multi-head self-attention mechanism is considered at this point and is utilized to improve predictions of the model. The empirical evidence indicates that the proposed model, GAE-MedBERT, provides higher gains over both MedBERT and Clinical BERT in terms of several evaluation measures, including accuracy, precision, recall, and F1 score, primarily for elderly chronic patients. This showcases the possible enhancement of chronic disease risk prediction and contributes to directions in managing and providing treatment methodologies to enhance patient prognosis through GAE-MedBERT.