YU-CHAO CHANG2026-10-162026-10-162026-01-01Pure ID: 15483709Pure UUID: 14f34bf3-1b06-48a4-befa-a975c61dabc510.1016/j.jds.2025.12.027https://www.scopus.com/pages/publications/105028121745https://scholars.csmu.edu.tw/handle/123456789/26555Background/purpose Periodontitis is a common chronic inflammatory disease linked to systemic conditions. We applied a graph convolutional network (GCN) to site-level periodontal data to predict all-cause and cardiovascular mortality from National Health and Nutrition Examination Survey (NHANES). Materials and methods Adults aged ≥30 years with full-mouth periodontal exams and linked mortality data through December 31, 2019 were included. Periodontal probing depth and clinical loss of attachment were measured. Each chart was converted into a graph with 168 nodes and anatomically defined edges. Graph-level embeddings were combined with age and sex to predict mortality. Model performance was evaluated in an independent test set using receiver operating characteristic - area under the curve (ROC AUC) and precision-recall - area under the curve (PR AUC). Results Among 9034 participants (1000 deaths), deceased individuals had significantly greater mean probing depth (1.71 ± 0.68 mm vs. 1.58 ± 0.66 mm) and loss of attachment (2.46 ± 1.34 mm vs. 1.86 ± 1.06 mm) than survivors (both P < 0.001). Adjusted mortality probabilities rose from 2.2 % to 5.6 % across probing depth quartiles and from 2.2 % to 6.3 % across attachment loss quartiles. The GCN achieved strong discrimination, with ROC AUC = 0.831 and 0.845 and PR AUC = 0.397 and 0.203 for all-cause and cardiovascular mortality, respectively. Conclusion A GCN applied to site-level periodontal data achieved strong discrimination in predicting mortality. This finding highlights the prognostic significance of periodontal health and demonstrates the potential of graph-based deep learning for modeling complex periodontal–systemic interactions.enjournal article