Latent-space clustering and feature interpretation in GraphCROC-Encoded periodontal disease, focusing on anterior and posterior bone loss prediction.

BACKGROUND

Periodontitis is a widespread inflammatory disease marked by progressive loss of the periodontal ligament and alveolar bone, often leading to tooth loss. Its complex etiology involves dental plaque, the immune response, behavior, and systemic factors such as diabetes. While traditional diagnostic tools such as probing depth, clinical attachment loss, and radiographs are useful, they often miss the heterogeneity of disease progression. Conventional assessments evaluate clinical and radiographic data separately, missing key patterns and interactions. Recent advances in artificial intelligence, especially graph neural networks (GNNs), model the oral cavity as a network of teeth as nodes and relationships as edges. This approach provides a more comprehensive and biologically relevant understanding of periodontal disease.

METHODS

We developed GraphCROC, a dual-stream model that separately encodes 11 node-level clinical features and five edge-level bone loss metrics, integrating them via a cross-correlation module. The model was trained on data from 90 patients using a clinically weighted reconstruction loss and evaluated against VAE, GCN, and DEC baselines.

RESULTS

GraphCROC outperformed all baselines in reconstruction (MAE = 0.402, R2 = 0.607) and predicted CAL with the lowest RMSE (0.369). Unsupervised clustering revealed three biologically meaningful disease patterns. Dimensionality reduction revealed clear gradients that represent inflammation and structural loss.

CONCLUSION

GraphCROC successfully reconstructs clinical data, identifies disease subtypes, and supports personalized diagnosis by connecting individual risk factors to spatial bone loss. This graph-based model demonstrates potential for AI-driven periodontal risk assessment and treatment planning.