DxFormer: a decoupled automatic diagnostic system based on decoder-encoder transformer with dense symptom...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Linguistics

Background:

• Symptom-based diagnostic systems aim to predict diseases through patient interaction.
• Current methods often focus on diagnosis, neglecting optimal symptom inquiry, limiting accuracy.
• Reinforcement learning (RL) has been explored but struggles with joint symptom-disease optimization.

Purpose of the Study:

• To propose DxFormer, a novel framework for automatic disease diagnosis that decouples symptom inquiry and disease prediction.
• To enhance diagnostic accuracy by optimizing symptom recall and interaction turns.
• To leverage Transformer architecture for improved representation learning in medical dialogues.

Main Methods:

• DxFormer treats symptoms as tokens, formalizing inquiry as language generation and diagnosis as sequence classification.
• A decoder-encoder Transformer structure is employed to learn symptom representations.
• Joint optimization using reinforcement learning reward and cross-entropy loss is utilized.

Main Results:

• Experiments on three real-world medical dialogue datasets demonstrate improved diagnostic accuracy.
• The proposed method significantly enhances symptom recall compared to previous approaches.
• DxFormer achieves state-of-the-art diagnostic accuracy by effectively decoupling the two core tasks.

Conclusions:

• Decoupling symptom inquiry and disease diagnosis is a viable strategy for improving automatic diagnostic systems.
• DxFormer offers a significant advancement in medical diagnosis accuracy and efficiency.
• The framework's modular design allows for independent optimization of its components.