(A-179) Transfer Learning Strategies for Addressing Class Imbalance in Drug Adverse Reaction Prediction: Simulation and Empirical Analysis

Background: Class imbalance is a critical challenge in drug adverse reaction (ADR) prediction, especially for severe yet rare ADRs of high clinical significance. Traditional machine learning models often struggle to address this issue, while transfer learning provides a novel solution by utilizing information from related but relatively common ADRs within the same patient population to enhance rare ADR prediction.

Objectives: This study aims to explore transfer learning strategies for addressing the class imbalance problem in ADR prediction models through systematic data simulation and empirical case studies.

Methods: This study constructed multiple ADR prediction scenarios based on Weibull distribution, and employed Copula functions to model the correlation between rare and common ADRs within the same drug-users. Using deep neural networks as the base model, we implemented a pre-training and fine-tuning framework: the first few layers of the common ADR network were frozen and transferred to the rare ADR model, followed by fine-tuning with rare ADR data. The performance of the transfer learning model was compared against models developed using traditional class imbalance methods, including sampling and cost-sensitive learning, in terms of discrimination and calibration. Furthermore, we examined the impact of ADR rates (both target and source ADR) and the strength of ADRs associations on the model performance, thereby identifying optimal modeling strategies for different scenarios. Finally, the proposed strategies were empirically validated using data on acute liver injury among statin users from the regional medical database of Yinzhou District, Ningbo, China.

Results: Simulation results demonstrated that transfer learning significantly improved discrimination of rare ADR prediction models. Specifically, the lower the target ADR event rate and the stronger the association between the target and source ADRs, the greater the performance enhancement of the transfer learning model compared to the unprocessed model. In contrast to traditional methods, the transfer learning model had a minimal impact on calibration and did not significantly overestimate ADR incidence rates. Furthermore, when combined with traditional class imbalance mitigation methods, the performance of the transfer learning model was further enhanced. Empirical validation supported these findings, showing that the transfer learning approach improved the prediction of acute liver injury among statin users.

Conclusions: By using transfer learning, this study proposes a novel approach to addressing class imbalance in ADR prediction. Combining transfer learning with traditional methods offers a comprehensive and practical solution to enhance predictive performance for rare ADRs.