Medical Specialty Classification Based on Semiadversarial Data Augmentation

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic health record (EHR) adoption necessitates automated medical specialty classification for efficient data retrieval and clinical decision support.
• Current challenges include imbalanced and insufficient medical note data, multicategory classification complexity, and difficulties in de-identifying sensitive information.

Purpose of the Study:

• To propose a novel data augmentation method using adversarial attacks to address data limitations in medical specialty classification.
• To design a classification framework that incorporates probabilistic noun information and confidence recalculation for improved accuracy.

Main Methods:

• A data augmentation method based on adversarial attacks, generating semi-adversarial examples to expand training data coverage.
• A classification framework integrating probabilistic noun information with confidence recalculation post-softmax layer.
• Validation on an 18-class, highly imbalanced dataset.

Main Results:

• The proposed method significantly improved accuracy and F1 score compared to four benchmark methods.
• An average improvement of 14.9% in accuracy and F1 score was observed.
• The technique effectively expanded the training data's decision space coverage.

Conclusions:

• The adversarial attack-based data augmentation and probabilistic noun-focused classification framework offer a robust solution for imbalanced medical specialty classification.
• This approach enhances the efficiency and reliability of EHR data analysis for clinical applications.