Mitigating Bias in Multi-Label Medical Text Classification: A Cooperative Training Framework with Dynamic Debiasing.

Abstract

Medical text classification plays a critical role in clinical decision support, automated diagnosis, and biomedical research. However, deep learning models are highly susceptible to dataset-induced biases, such as label bias and keyword bias, which can lead to unreliable predictions and poor generalization in real-world clinical applications. Existing debiasing methods often either overcorrect informative samples or lack interpretability during inference, limiting their effectiveness in multi-label medical text classification tasks. We propose Cooperative Debiasing Network (CoDeNet), a cooperative training framework that mitigates dataset bias through dynamic sample reweighting and interpretable counterfactual inference. The framework consists of a primary classifier and a debias estimator, where the debias estimator quantifies sample-level bias and dynamically regulates the optimization process through an elastic scaling mechanism. In addition, a counterfactual post-processing strategy explicitly isolates label-level and keyword-level biases to improve interpretability. Experiments conducted on the DepressionEMO and BDI-Sen datasets demonstrate that CoDeNet consistently improves classification performance over strong Transformer-based baselines, including BERT and MentalBERT. In particular, CoDeNet achieves improvements of up to +6.57% macro-F1 on BDI-Sen and +2.16% macro-F1 on DepressionEMO, with especially strong gains on low-frequency clinical labels. The results indicate that CoDeNet effectively reduces dataset-induced bias while preserving model robustness and interpretability. The source code and implementation details of CoDeNet will be publicly available on GitHub: https://github.com/66ccff39C5BB/CoDeNet.