Multivariate and Online Transfer Learning With Uncertainty Quantification

Area of Science:

• Biostatistics
• Dental Research
• Machine Learning

Background:

• Periodontitis, a common dental condition, can lead to tooth loss if untreated.
• Accurate modeling of periodontal outcomes is challenging due to measurement difficulties.
• Existing models may fail or pose risks when applied to underrepresented demographic groups.

Purpose of the Study:

• To extend the RECaST Bayesian transfer learning framework for improved periodontal outcome modeling.
• To address disparities in representation within demographic groups for predictive modeling.
• To develop a method that enhances model performance for underrepresented populations without data sharing.

Main Methods:

• Proposed an extension to the RECaST Bayesian transfer learning framework.
• Developed a joint multivariate outcome modeling approach.
• Introduced an online method for sequential datasets and mitigated negative transfer.

Main Results:

• The proposed method significantly improved upon the previous univariate RECaST approach.
• Demonstrated effective predictive performance and uncertainty quantification on simulated and real dental data.
• Successfully mitigated negative transfer, protecting underrepresented groups from detrimental model application.

Conclusions:

• The novel Bayesian transfer learning framework enhances periodontal outcome prediction accuracy and reliability.
• The method is particularly valuable for applications in healthcare where demographic representation is critical.
• The approach offers robust uncertainty quantification and ensures data privacy by not sharing data between domains.