Spatiotemporal signal detection using continuous shrinkage priors.

Area of Science:

• Oral Health
• Biostatistics
• Data Science

Background:

• Periodontal disease (PD) is a chronic inflammatory condition affecting gums and tooth-supporting bone.
• Current PD assessment often uses average periodontal pocket depth (PPD), losing valuable site-specific, temporal information.
• Existing spatiotemporal statistical methods are often unscalable for large longitudinal dental databases.

Purpose of the Study:

• To introduce a scalable Bayesian spatiotemporal model for dynamic detection of diseased tooth sites in large oral health databases.
• To improve the precision of periodontal disease progression assessment.
• To address the limitations of whole-mouth average PPD in capturing disease activity.

Main Methods:

• Developed a Bayesian spatiotemporal model incorporating a spatial continuous sparsity-inducing shrinkage prior.
• Utilized a low-rank representation to capture nonstationary covariance structure of PPD outcomes.
• Implemented Markov chain Monte Carlo (MCMC) methods for scalability to large datasets.

Main Results:

• The proposed model demonstrated improved prediction performance on simulated and real-world electronic dental record data.
• Outperformed alternative models using standard Gaussian priors and conditionally autoregressive structures.
• Successfully identified problematic tooth sites dynamically within subjects.

Conclusions:

• The novel Bayesian spatiotemporal model offers a scalable and accurate approach for analyzing periodontal disease progression in large datasets.
• This method enhances the understanding of disease dynamics at the tooth-site level.
• Facilitates more precise identification and management of periodontal disease in clinical practice.