Hierarchical reinforcement learning for automatic disease diagnosis.

Area of Science:

• Artificial Intelligence
• Computational Biology
• Medical Informatics

Background:

• Current disease diagnosis dialog systems often use flat policy structures, limiting efficiency with large numbers of symptoms and diseases.
• These systems model consultations as Markov decision processes, typically solved with reinforcement learning.

Purpose of the Study:

• To propose a novel hierarchical policy structure for disease diagnosis dialog systems.
• To enhance the accuracy and efficiency of AI-driven medical consultations.

Main Methods:

• Developed a two-level hierarchical policy: a master model triggering specialized symptom checkers and a disease classifier.
• Applied reinforcement learning to train the hierarchical policy structure.

Main Results:

• The hierarchical framework demonstrated superior accuracy and symptom recall compared to existing systems on real-world and synthetic datasets.
• Experimental results validated the effectiveness of the proposed approach in handling complex diagnostic problems.

Conclusions:

• The hierarchical policy structure offers a more effective approach for disease diagnosis dialog systems, especially in complex scenarios.
• This research provides a benchmark for future studies in AI-powered medical diagnosis.