Risk assessment for canine periodontal disease using a hybrid causal Bayesian network.

Periodontal disease is among the most common diagnoses in canine primary care yet remains significantly underdiagnosed. The disconnect between prevalence and detection represents a critical gap in veterinary preventive medicine. Disease risk depends on non-modifiable factors (breed, age, morphology) and modifiable factors (dental hygiene, professional care), yet evidence-based interventions remain underutilized. We developed and validated a hybrid Bayesian network for canine periodontal disease risk assessment that integrates multiple factors to quantify disease probability. A first directed acyclic graph (DAG) for periodontal disease was constructed to define and map causal relationships between risk factors. This was followed by the construction of a Bayesian network that integrated data from 9.5 million electronic health records, 2,600 owner questionnaires, previous studies and expert elicitation. The final network comprised 19 nodes with 101 states and over 33,200 conditional probabilities. The model successfully differentiated high-risk from low-risk breeds and captured associations with age, size, head shape and dental hygiene practices. Key clinical indicators showed strong predictive value: a prior periodontitis probability was 12.4%, which increased to 17.6% with biofilm presence, 24.0% with poor dental conformation and 47.0% with gingivitis. The network demonstrated robust performance across four independent validation datasets, with ROC AUC values ranging from 0.583 to 0.962, sensitivity from 0.639 to 0.913 and specificity from 0.300 to 0.906. This hybrid Bayesian network integrated diverse data sources whilst accounting for complex interactions between morphological, clinical and preventive factors. The model's bidirectional inference enables risk calculation using any combination of the 19 nodes and can operate as both a probabilistic inference tool (capturing observed associations) and causal inference tool (predicting intervention outcomes). This approach provides a framework to support clinical decision-making and demonstrates the utility of hybrid Bayesian networks for complex veterinary conditions where traditional epidemiological approaches face limitations.