Abstract:Automatic diagnosis, as an important research direction in artificial intelligence engineering, has advanced significantly in recent years. In real diagnostic scenarios, after the patient informs the doctor about their most obvious symptoms, the doctor is often required to guide the patient to uncover potential symptom information. The doctor then makes the final diagnosis after obtaining sufficient evidence. Previous work has modeled this as a sequential decision-making process and trained an optimal policy using RL-based methods, achieving a high diagnosis correctness rate. However, the RL-based method only finds symptom sequences that improve reward through random trials, failing to directly capture the inter-relationships between symptom sequences in the inquiry process. Also, the doctors' diagnostic experience, especially symptoms that often co-occur and are located in close proximity to each other in the doctors' inquiry symptom sequence, has not been well explored in previous implementations. To address this, we propose a model that further enhances the learning of doctors' diagnostic experience based on a two-part transformer model that completes symptom inquiry and disease diagnosis separately. We introduce a Doctor Diagnosis Experience Reinforcement Module (DDEM) to the symptom sequence inquiry part, reducing the model's sensitivity to the order of symptom sequences within the same symptom cluster. Then, we further enhance the learning of this symptom distribution by corresponding reinforcement learning reward rules. Also, we propose three new transition rules when transitioning the symptom inquiry model to the disease diagnosis model to enable the symptom inquiry model and the disease diagnosis model to collaborate in a manner that closely aligns with the diagnostic logic used by doctors. Experiments on three public real-world medical dialogue datasets demonstrate that the proposed model improves diagnostic accuracy by 2%, 1.4%, and 0.7% and shows a clear advantage in symptom recall rate, highlighting its effectiveness.

A sequential decision-theoretic model for medical diagnostic system

A Patient-Similarity-based Model for Diagnostic Prediction

A Novel Sequential Three-Way Decision Model for Medical Diagnosis

An Interactive Task Analysis Framework and Interactive System Research for Computer Aided Diagnosis

A Knowledge Representation Method Based on Two-Layer Modeling for Constructing Medical Knowledge Base

Test sequencing and diagnosis in electronic system with decision table

Knowledge-Based Decision Model Construction for Hierarchical Diagnosis: A Preliminary Report

Representation and Acquisition of Expert Knowledge for Medical Diagnosis System

A Comparison of Decision Analysis and Expert Rules for Sequential Diagnosis

A multiperiod hybrid decision support model for medical diagnosis and treatment based on similarities and three‐way decision theory

Decision Curve Analysis: a Technical Note

Decision-Making System for the Diagnosis of Syndrome Based on Traditional Chinese Medicine Knowledge Graph

A Medical Decision Making Study based on Markov Chain Model

A novel intelligent model for visualized inference of medical diagnosis: A case of TCM

A doctor's diagnosis experience enhanced transformer model for automatic diagnosis

Group Decision-Making Method with Incomplete Intuitionistic Fuzzy Soft Information for Medical Diagnosis Model

A bipolar intuitionistic fuzzy decision-making model for selection of effective diagnosis method of tuberculosis

A Novel Interval Three-Way Concept Lattice Model with Its Application in Medical Diagnosis

Medical Dialogue Generation via Intuitive-then-Analytical Differential Diagnosis

A Knowledge-Based Modeling System for Time-Critical Dynamic Decision-Making

Sequential Diagnosis by Abstraction