Abstract
Automated medical dialogue systems have the potential to streamline patient pre-screening and clinical data collection. However, existing systems often struggle with inefficient dialogue policies due to a lack of structural understanding of medical knowledge. This paper presents a novel Graph-DQN architecture that leverages a weighted heterogeneous graph to model symptom-disease relationships. By integrating Graph Convolutional Networks (GCN) for feature extraction and Deep Q-Networks (DQN) for policy learning, the system achieves superior diagnostic accuracy with significantly reduced dialogue length.
Introduction
The development of automated dialogue systems for medical diagnosis is a critical step toward accessible healthcare. A primary challenge lies in the Dialogue Manager's ability to ask informative questions. Traditional approaches often rely on flat statistical features or unweighted graphs, which fail to capture the varying clinical importance of different symptoms. Consequently, these agents frequently ask redundant or irrelevant questions, leading to prolonged conversations and decreased user satisfaction. This research addresses these limitations by introducing a graph-based reasoning framework that prioritizes discriminative symptoms.
Methodology: Graph-DQN Architecture
The proposed system architecture consists of three main components: Natural Language Understanding (NLU), Dialogue Management (DM), and Natural Language Generation (NLG). The NLU module extracts user intents, symptoms, and demographic information (e.g., age group), which are passed to the DM.
The core innovation is the weighted heterogeneous graph used within the DM. Unlike binary graphs, this model employs specific weighting schemes to capture clinical relevance:
- Symptom-Symptom Edges: Weighted using Point-wise Mutual Information (PMI) to capture the statistical co-occurrence of symptoms in clinical dialogues.
- Symptom-Disease Edges: Weighted using a novel sf-idf (symptom frequency-inverse disease frequency) metric. Similar to tf-idf in text processing, sf-idf highlights symptoms that are uniquely characteristic of specific diseases, allowing the agent to prioritize highly discriminative queries.
The Dialogue Manager utilizes a GCN to process this graph, generating rich embeddings for symptoms and diseases. These embeddings are then fed into a DQN, which determines the optimal next action—either requesting more information about a symptom or providing a final diagnosis—based on the current dialogue state.
Experimental Evaluation
The model was evaluated on two real-world disease diagnosis dialogue datasets:
- MZ Dataset: Collected from Baidu Muzhi Doctor, containing 710 dialogue goals across 4 disease types and 66 symptoms.
- DX Dataset: Collected from dxy.com, containing 527 dialogues across 5 disease types and 41 symptoms.
Results and Discussion
The Graph-DQN demonstrated significant improvements in both accuracy and efficiency. On the MZ dataset, the model achieved 69.7% accuracy in an average of 2.24 turns, nearly halving the 5.11 turns required by a standard DQN baseline (65.0% accuracy). Similar trends were observed on the DX dataset, where the model reached 74.0% accuracy in 2.21 turns.
Qualitative analysis illustrates how the Graph-DQN avoids the pitfalls of baseline models. By leveraging the weighted graph, the system identifies key symptoms—such as decreased urine output to distinguish between infantile dyspepsia and diarrhea—much earlier in the conversation than traditional RL agents.
Visualization of the learned embeddings confirmed that the Graph-DQN successfully clusters diseases with their relevant symptoms in the latent space. This structural awareness is key to the system's ability to differentiate between diseases with overlapping clinical presentations, a property missing in standard DQN models where embeddings appear scattered.
Conclusion
This work demonstrates that incorporating structured medical knowledge via weighted heterogeneous graphs significantly enhances the performance of diagnostic dialogue systems. By prioritizing discriminative symptoms, the Graph-DQN provides a more efficient and accurate user experience. Future research will focus on scaling the model to larger datasets and deeper integration of patient-specific background data, such as detailed demographics and hereditary factors, to further personalize the diagnostic process.
Summary updated