Mitigating bias in multilabel medical text classification: a cooperative training framework with dynamic debiasing.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Deep learning models are crucial for medical text classification in clinical decision support and research.
• Dataset biases, including label and keyword bias, compromise the reliability and generalization of these models.
• Existing debiasing techniques struggle with overcorrection, lack interpretability, and are less effective for multi-label classification.

Purpose of the Study:

• To introduce a novel cooperative training framework, Cooperative Debiasing Network (CoDeNet), to mitigate dataset-induced biases in medical text classification.
• To enhance the robustness and interpretability of deep learning models in clinical applications.
• To address limitations of current debiasing methods in multi-label medical text classification.

Main Methods:

• CoDeNet employs a cooperative training framework with a primary classifier and a debias estimator.
• Dynamic sample reweighting and an elastic scaling mechanism regulate the optimization process based on bias quantification.
• Interpretable counterfactual inference and post-processing isolate label-level and keyword-level biases.

Main Results:

• CoDeNet demonstrated consistent performance improvements over Transformer-based baselines (BERT, MentalBERT) on DepressionEMO and BDI-Sen datasets.
• The framework achieved significant gains, including up to +6.57% macro-F1 on BDI-Sen and +2.16% macro-F1 on DepressionEMO.
• CoDeNet showed particularly strong improvements for low-frequency clinical labels, indicating effective bias reduction.

Conclusions:

• CoDeNet effectively reduces dataset-induced bias in medical text classification.
• The proposed framework enhances model robustness and interpretability.
• CoDeNet offers a promising solution for reliable and interpretable medical AI applications.