RECONSTRUCTING PERIODONTAL-CLINICAL GRAPHS: AN AUTOENCODER APPROACH TO UNVEIL ANAEMIA RISK

Background: Periodontal disease is increasingly recognized as a systemic condition with broader implications
beyond the mouth. An underexplored link is its association with anemia, where chronic inflammation and
bleeding may affect blood health. Traditional methods often miss the complex, non-linear interactions between
periodontal parameters and systemic biomarkers like haemoglobin. This study introduces a graph autoencoder
(GAE) framework to model periodontal data alongside hematological markers, aiming to better detect anemia
with improved interpretability and accuracy.
Materials and Methods: Clinical records from 350 patients, including periodontal measures (probing pocket
depth, clinical attachment loss, bleeding on probing, plaque index) and systemic features (haemoglobin levels,
systemic diseases, smoking status), were pre-processed using one-hot encoding for categorical variables and
standardisation for continuous values. A 10-nearest-neighbour graph was constructed to represent patient
similarity, and a single-layer GAE with a latent dimension of 16 was trained for 100 epochs using binary crossentropy loss. Logistic regression on latent embeddings predicted anaemia status. Performance was evaluated
through cross-validation accuracy and reconstruction metrics.
Results:The model achieved stable classification accuracy (~81.4%) across epochs. Reconstruction metrics
demonstrated robust graph representation learning, with an AUC of 0.898, average precision of 0.241, and
cross-entropy loss of 0.693. Additional analyses revealed a mean squared error of 0.250, high recall (0.996),
but low precision (0.074), reflecting the imbalanced distribution of anaemic versus non-anaemic cases and
precision–recall curves further illustrated the trade-offs in predictive performance.
Conclusion: Graph autoencoders offer a powerful way to learn latent structures in periodontal–systemic data.
They show how graph-based deep learning can detect conditions like anaemia early from dental records,
advancing integrated health strategies linking oral and systemic health.