A1283
Title: Integrating Bayesian generative AI with weighted logistic regression for imbalanced biomedical data classification
Authors: Charlotte Wang - National Taiwan University (Taiwan) [presenting]
Yu-Ting Huang - National Taiwan University (Taiwan)
Abstract: Analyzing imbalanced data is a critical challenge in biomedical and public health research, where rare events are often of primary interest. Traditional data-level approaches frequently fail to capture underlying distributions in complex scenarios, leading to inconsistent classification performance. A hybrid framework is proposed that integrates a Bayesian generative AI approach with subject-weighted logistic regression. A BNNVAE model is introduced that incorporates Bayesian Neural Networks into a Variational Autoencoder to generate representative samples. These synthetic samples augment the dataset to achieve class balance. Subject-weighted logistic regression is then employed, assigning weights based on sample representativeness and proximity to decision boundaries. Simulation and real-world applications demonstrate that the BNNVAE framework significantly outperforms traditional methods such as SMOTE. It achieves superior stability and performance across F1-score, G-mean, and balanced accuracy metrics. By leveraging Bayesian neural networks, the model provides more effective estimates of data distributions, reducing reliance on large training data, a common limitation in practical applications. The proposed framework can generate more representative synthetic data and, through sample reweighting, construct interpretable classification models, thereby enhancing its practical value and applicability in biomedical and public health research.
