An Explainable and Personalized Cognitive Reasoning Model Based on Knowledge Graph: Toward Decision Making for General Practice

Background

General medicine, as an important part of community and family healthcare, covers different ages, genders, organ systems, and various diseases. Its core concept is human-centered, family-based, emphasizing long-term comprehensive health maintenance and promotion. However, existing evidence shows that the quality of Primary Health Care (PHC) in China still falls short of satisfying standards. There is significant room for improvement in the accuracy of clinical diagnosis and treatment. To address this issue, AI-based decision tools have gradually become a powerful aid for general practitioners in diagnosing diseases. However, current research mainly has two issues: first, a lack of sufficient scalability and interpretability; second, most existing models are complex to operate and difficult to apply in actual general medical environments.

Paper Source

This research paper, titled “An Explainable and Personalized Cognitive Reasoning Model Based on Knowledge Graph: Toward Decision Making for General Practice,” is authored by Qianghua Liu, Yu Tian, Tianshu Zhou, Kewei Lyu, Zhixiao Wang, Yixiao Zheng, Ying Liu, Jingjing Ren, and Jingsong Li. It was published in the IEEE Journal of Biomedical and Health Informatics, Vol. 28, No. 2, in February 2024.

Research Summary

This paper proposes an Explainable and Personalized Cognitive Reasoning Model based on a Knowledge Graph (CRKG), aiming to utilize patients’ Electronic Health Records (EHRs) and a Knowledge Graph to provide personalized diagnosis and decision support. Focusing particularly on abdominal diseases, an abdominal disease Knowledge Graph (AKG) was first semi-automatically constructed. Combining the dual process theory from cognitive science, CRKG uses Graph Neural Networks and attention mechanisms to outperform in the diagnosis of common diseases. Experimental evidence shows that this model is superior to existing baseline models in both precision and recall.

Research Process

Knowledge Graph Construction

1. Knowledge Sources: The knowledge sources for this study include Chinese clinical guidelines, medical books, and information obtained through the UpToDate website. Relevant knowledge is also automatically extracted from the SemMedDB database in PubMed and the Chinese version of ICD-10.

2. Knowledge Extraction Templates: An extraction template was designed to add structured knowledge into the Knowledge Graph in the form of triplets (head, relation, tail). Entity types include diseases, symptoms, signs, processes, measurements, medical history, and drugs, while relations include “coexist,” “is a type of,” “causes,” “affects,” “prone to,” and “examines,” among others.

Knowledge Graph Update Strategy

To endow the Knowledge Graph with semantic information, a Message-Passing Neural Network (MPNN)-based update strategy was designed. The specific steps are as follows: 1. Node Information Exchange: Messages are passed from node h through relation r to the target node t. 2. Message Aggregation: The hidden representation of the node is generated by aggregating the information from all edges. 3. Node State Update: Using residual connections to avoid information loss and introducing learnable parameters to optimize graph structure learning.

Personal Cognitive Graph Construction

Each patient consultation can be modeled as a Personal Cognitive Graph (PCG). The specific process is as follows: 1. Initial Cognition Construction: The initial clinical data of the patient serves as the foundation for the PCG. 2. Intuitive Attention Capture Module: Based on patient data, relevant information from the Knowledge Graph is retrieved and loaded into the PCG, and the Graph Attention Network (GAT) is applied for attention transfer. 3. Explicit Reasoning Module: Disease reasoning is performed through explicit and conscious decisions, further updating node states, and ultimately predicting the patient’s potential diseases.

Experimental Results

The paper compares the CRKG model with several existing recommendation system baseline models: 1. Model Performance: CRKG performs excellently in terms of precision and recall. For instance, CRKG achieves a precision@1 of 0.7873, recall@10 of 0.9020, and hits@10 of 0.9340, which are significantly higher than those of other baseline models. 2. Interpretability: The model can intuitively display the reasoning process for each patient consultation, enhancing clinical doctors’ understanding of the model’s judgments and thereby increasing the model’s interpretability.

Research Conclusions and Value

The CRKG model has significant scientific and practical value: 1. Scientific Contribution: The research proposes a knowledge graph update strategy and reasoning method combining Graph Neural Networks and attention mechanisms, providing a new path for the development of cognitive intelligence, making AI decision-making tools more akin to human cognitive processes. 2. Application Prospects: The model can effectively improve the disease diagnosis accuracy and efficiency of general practitioners and can also identify key patients early and provide early referral recommendations, offering technical support for improving primary healthcare in China.

Research Highlights

1. Innovative Method: CRKG combines dual process theory (intuitive and analytical cognition) and innovatively designs a semi-automated knowledge extraction and reasoning process.
2. Superior Performance: Experiments reveal that CRKG outperforms existing baseline models on multiple metrics, demonstrating its outstanding diagnostic performance.
3. Interpretability: By visualizing the reasoning process, it enhances doctors’ trust and understanding of the model, increasing its potential for clinical application.

Through this research, not only are intelligent tools for general medical practice enriched, but a solid theoretical and practical foundation is also provided for the further exploration and application of health informatics technologies.