Article Preview
TopIntroduction
Medical dialogue services assume a crucial role in the healthcare domain by enhancing the accessibility of medical services. The systems can provide patient consultation services, address health-related inquiries, and facilitate improved understanding among patients of medical conditions, treatment procedures, and preventive measures. Li, Ren, et al. (2024) have applied the Bidirectional Encoder Representation from Transformers (BERT)+Slot-Gated deep learning model for the identification of traditional Chinese medicine entities and question intentions presented by users in their questions and the acquisition of the identified entities and intentions on the foundation of a comprehensive knowledge graph (KG) constructed from the textbooks Prescriptions of Chinese Materia Medica and Chinese MateriaMedica. Huang, Zhang, et al. (2021) proposed an inference method based on weighted path ranking on the KG to score the related entities according to the key information and intention of a given question. In summary, medical dialogue services improve patient experience, assist decision-making, and increase medical efficiency.
Current medical research predominantly focuses on question-answering systems with relatively limited exploration of medical dialogue services that involve multi-turn interactions. Compared to question-answering systems, dialogue services are more intricate, as they must consider the entire conversation history throughout multi-turn interactions. Process knowledge, which represents the potential trajectory of each conversation, involves the intents or actions of participants within the dialogue. To enhance the accuracy of handling multi-turn dialogues, we integrated process knowledge into the medical dialogue service. This integration helped the service better understand and respond to complex dialogue scenarios, thereby improving user experience and increasing the service's practicality.
In the medical domain, the requirements for result interpretability, credibility, and traceability surpass those of other fields. However, the black-box nature of large language models (LLMs) often results in biased outcomes. Integrating medical KGs with various artificial intelligence technologies can play a more critical role. KGs have a notable advantage over LLMs regarding interpretability, credibility, and traceability. Wong, et al. (2024) proposed DeepDR-LLM, a vision-LLM integrated system for diabetes diagnosis and treatment, which provides personalized diabetes management recommendations and assists in the diagnosis of diabetic retinopathy for physicians. Cao, et al. (2023) introduced the pancreatic cancer detection with artificial intelligence (PANDA) model, which integrates computed tomography (CT) imaging for early screening of pancreatic cancer. Consequently, the fusion of KGs and LLMs not only increases the reasoning capabilities of LLMs but also enhances their outcomes' practicality and reliability.