Enhancing medical text classification with GAN-based data augmentation and multi-task learning in BERT.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Electronic medical records offer opportunities for intelligent medical text classification.
• Traditional models face challenges like class imbalance, semantic heterogeneity, and data sparsity.
• These limitations hinder the effectiveness of current classification approaches.

Purpose of the Study:

• To propose an enhanced medical text classification framework.
• To address limitations of traditional models in handling imbalanced and sparse medical data.
• To improve the accuracy and robustness of clinical decision-support systems.

Main Methods:

• Integration of a self-attentive adversarial augmentation network (SAAN) for data augmentation.
• SAAN utilizes adversarial self-attention to generate high-quality minority class samples and reduce noise.
• Employment of a disease-aware multi-task BERT (DMT-BERT) strategy for enhanced feature extraction.
• DMT-BERT learns medical text representations and disease co-occurrence relationships simultaneously.

Main Results:

• The proposed framework significantly outperforms baseline models on private clinical and public CCKS 2017 datasets.
• Achieved the highest F1-score and ROC-AUC values in experiments.
• Demonstrated improved handling of class imbalance and data sparsity in medical text classification.

Conclusions:

• The developed framework effectively addresses key limitations in medical text classification.
• The integration of SAAN and DMT-BERT enhances the performance of intelligent medical text analysis.
• Contributes to the development of more robust and reliable clinical decision-support systems.