A Bayesian Modelling Approach with Balancing Informative Prior for Analysing Imbalanced Data.

Area of Science:

• Statistics
• Computational Biology
• Pharmacometrics

Background:

• Imbalanced datasets pose challenges in predictive modeling, potentially biasing results towards majority subpopulations.
• Accurate prediction for underrepresented groups is crucial in various scientific applications, including clinical dose optimization.

Purpose of the Study:

• To develop a Bayesian modeling approach incorporating a balancing informative prior to mitigate the impact of data imbalance.
• To enhance the prediction accuracy and robustness of models dealing with datasets where certain subpopulations are scarcely sampled.

Main Methods:

• Development of a Bayesian modeling framework utilizing a balancing informative prior to re-weight data in favor of smaller subsets.
• Assessment of the method's performance using simulated datasets, evaluating bias and precision of model parameter estimates.
• Application of the approach to a real-world example of predicting optimal tobramycin dosage across different age groups.

Main Results:

• The balancing informative prior approach demonstrated reduced bias compared to conventional methods that do not account for data imbalance.
• Superior precision estimates were observed with the proposed Bayesian method.
• Predictions for optimal tobramycin dosage aligned well with existing literature, validating the approach's practical utility.

Conclusions:

• The proposed Bayesian balancing informative prior approach effectively addresses data imbalance by appropriately weighting smaller data subsets.
• This method offers a robust strategy for generating reliable prediction models, particularly in scenarios with significant subpopulation disparities.
• The approach shows significant potential for improving predictive accuracy in fields such as pharmacometrics and personalized medicine.