Agents in medical decision support have been extensively researched, particularly in areas like evidence support, multimorbidity management, and patient-specific needs. However, current approaches lack a unified method to address real-time evidence updates, manage multiple diseases independently, and incorporate personalized patient needs. Existing Large Language Model (LLM) agents are limited by their reliance on static knowledge bases, hindering their ability to promptly update clinical guidelines and meet diverse patient requirements. Moreover, the interpretability of LLMs remains a significant concern, leading to skepticism in their medical application. To address these challenges, we developed MedGen, a multi-agent architecture that decomposes the clinical decision-making process into stages such as clinical goal setting, data collection, argumentation linking, and plan selection. This structured approach allows LLM agents to provide both reasoning evidence and a transparent reasoning process, enhancing the reliability and interpretability of outcomes. Finally, a case study of breast cancer and depression is combined to illustrate our architecture.

Independently developed clinical guidelines present a systematic challenge in managing patients with multi-morbidity in a consistent and integrated manner. Existing approaches mainly focus on combining multiple guidelines and lack approaches that combine with additional medical resources. The correlations and conflicts between treatment plans in the management of multi-morbidity are well-documented in medical literature but are less explored in the Clinical Decision Support line of research. In this paper, we propose a literature-based guideline interlinking method to address these challenges through the integration of clinical guidelines and the harmonization of conflicting recommendations, thereby providing a more holistic and efficient way to manage patients with multi-morbidity conditions. This method employs an ontology model and knowledge graph technology to represent and analyze the complexity and interrelations of diseases, with the aim of transcending the limitations of traditional single disease guidelines and providing a holistic and integrated framework for multi-morbidity management. The objective is to construct a multi-morbidity knowledge graph by correlating medical literature with clinical guidelines and to provide optimal decision support for patients with multi-morbidity complications in a clinical decision support system (CDSS).

In today’s complex healthcare environment, shared decision making is increasingly being emphasized as an ideal model for healthcare decision making. The interaction of a set of arguments can represent the beliefs of agents, and the process of reaching consensus in a multi-agent system can be facilitated by argument merging. However, existing argumentation models are insufficient to support the application of shared decision making in healthcare. To fill this gap, this paper, we construct a cognitive representation model of agents supporting the computational task of argumentation and propose a novel interactive argumentation framework merging method based on it as a solution for shared decision support in healthcare environment. We used the Lightweight Social Calculus (LSC) to describe and standardize our interaction model. The usability of the method is demonstrated by definitional proofs and case study.

This paper describes the process of embedding ChatGPT-4 (Large Scale Natural Language Modeling) in a CDSS (Clinical Decision Support System) to generate multiple types of decision recommendations. First, given enquiry data, the CDSS generates a specific type of decision recommendations and gives the questions and answers to ChatGPT-4 to generate the related type of decision recommendations. To categorize the same type of decision recommendations in CDSS and ChatGPT-4 together, Word2Vec model was used to learn the semantic relationships of words in medical texts and evaluate the model. Then, the similarity between the decision recommendations generated by CDSS and ChatGPT-4 is determined by calculating the cosine similarity, and a suitable threshold is set to decide whether to fuse these decision recommendations or not. Through this method, the decision recommendations generated by CDSS and ChatGPT-4 can be effectively fused to provide more comprehensive and precise clinical decision support.

Clinical decision support systems (CDSS) are knowledge-driven tools that have the potential to improve diagnosis and decision-making for clinicians. However, no clinical decision model is infallible. Therefore, CDSS should enable clinicians to validate each decision recommendation and reject any incorrect ones while selecting the right ones. Although previous studies have aimed to explain the arguments behind each decision candidate, these arguments are often drawn from a knowledge base that is modeled from clinical guidelines and may not always be up-to-date with the latest medical research.To address this issue, we propose a different approach that provides accurate and relevant scientific evidence from the biomedical literature. Our proposed knowledge extension engine uses the BioBERT tool to efficiently identify clinical trial reports based on a range of clinical questions. This feature enables the system to identify clinical trials that are relevant to diagnostic or treatment hypotheses. Furthermore, the knowledge extension engine can extract essential information from clinical trial summaries, such as patient populations, interventions, and outcomes. This capability allows clinicians to quickly identify matches between clinical questions and clinical trials, and understand key elements of clinical trials without extensive reading.Additionally, we have designed a knowledge modeling approach that facilitates the rapid updating of the knowledge base, enabling domain experts to quickly update the knowledge base based on the latest literature provided by the system. By following this approach, we aim to provide clinicians with up-to-date and accurate information to improve the effectiveness of CDSS in clinical decision-making.

The most studied form of group intelligence in nature is the bee colony, where the process of finding a new nest site is similar to the democratic discussion in human group decision making. Some researchers are currently simplifying complex problems by modeling the behavior of bee swarms and interacting with them by building agents that behave like swarms. In this paper, we define a swarm decision language containing behavioral language, agents, protocols, decision rules and constraints based on the process of swarm addressing, on top of the Lightweight Social Calculus specification (LSC). The aim is that the group cooperation pattern is described by a protocol and the behavioral characteristics of individuals are restricted by rules. Complex group decisions are described by a dynamic combination of these two levels. We use this method to provide a group decision solution for cases that require site selection in human life, reducing a large amount of manpower and increasing the speed and reliability of decision making compared to traditional manual site search methods.

Conversational robots have been widely used in the medical field to monitor, diagnose, and manage patient diseases. The construction of question and answer systems relies on domain knowledge, and according to research, there are currently no conversational bots available to answer postoperative thyroid eye disease questions. To alleviate healthcare resources, we have developed a multi-agent dialogue system for answering post-operative thyroid eye disease questions. The system consists of three agents: an intent understanding agent, a knowledge processing agent, and an intelligent interaction support agent, each responsible for intent recognition, knowledge base building, and conversation generation. The system is scalable and reusable with a division of labour between the agents working together to complete the dialogue tasks. In the experimental part, we used the dialogue system to simulate a post-operative thyroid eye surgery dialogue scenario, and the system was able to accurately recognise intent and answer the user’s questions well.

It is important to incorporate patient values in healthcare decision-making systems so that the system gives patient-centered decision solutions in the decision-making process. The extant studies show less integration of both. Patient values can be summarized in several medical community review data. The existing knowledge of values is distilled and then combined with existing knowledge of clinical guidelines and rehabilitation. The patient's personalized information is obtained during the consultation process and matched with the general medical data, resulting in personalized recommendations that incorporate the patient's values. An ontology model is constructed based on five domains: population features, medical treatments, personal values, side effects and rehabilitation. Clinical guideline data, rehabilitation data and patient review data are mapped and linked through the ontology to obtain a general medical knowledge graph. The patient's voice, emotion and gesture information during the consultation process are analyzed and recognized as weighted entities, which are passed into the general medical knowledge graph to form a personalized multi-modal knowledge graph of the patient. A prototype system in a man with breast cancer was designed and implemented to demonstrate the method's feasibility.

Aspect-Based Sentiment Analysis (ABSA) aims to determine the sentiment polarity of certain aspect words in a sentence.Recently, it is a popular approach to fuse the sentences' syntactic information via the dependency tree into the graph neural network.However, how to efficiently utilize the obtained syntactic information is still a challenging problem of this kind of approach.Therefore, this paper proposes a novel Aspect-level Sentiment Analysis model based on Convolutional network with Dependency Tree, named ASAC-DT in short.First, the attention mechanism is utilized to obtain the attention score of the sentence and the aspect word respectively, to improve the connection of the words related to the aspect word in the sentence.Afterwards, by relying on the syntactic information obtained from the dependency tree, the connections of words that are not related to the aspect words are reduced.Finally, the feature information most relevant to the aspect words in the proposed model is extracted through the graph convolutional neural network and the interactive network.Through extensive experimental baselines the proposed ASAC-DT model shows effectiveness in aspect-level sentiment classification and outperforms baselines in accuracy.

Clinical decision-making is closely related with the activity of argumentation among alternative options. In recent years, theories and languages have been developed for argumentation and evidence-based decision support. However, a systematic study of argument representation using evidence in the medicine domain is missing. In this paper, an Evidence-based Argumentation Graph is proposed. A Clinical Argumentation scheme and a Patient Preference Argumentation scheme guide their construction. Arguments can be represented using clinical and patient preference evidence and semantically integrated in the graph. Clinical decision support is delivered to clinicians and patients together. The method is demonstrated using a case study of decision support for patients suspected with breast cancer.

The treatment of patients with multimorbidity has always been a matter of importance. Due to the complexity of patients' conditions, physicians need to consider not only the cumbersome consultation process and complex care plans., but also potential clinical decision conflicts between different diseases. Currently, most clinical guidelines focus on a single medical condition, and the emergent and random nature of illness in patients with multiple conditions makes it difficult to take good account of the potential conflicts between various clinical decisions. Current clinical decision models on the treatment of complications are limited to specific types of complications and usually detect conflicts in a declarative method, which is difficult to cover various types of clinical decision conflicts and is not scalable. We model the treatment process of patients with multimorbidity as a goal forest and propose a goal-driven clinical support model for group decision making. This model is applicable to distributed settings and can integrate multiple clinical guidelines to concurrently treat patients with multimorbidity. A clinical decision conflict ontology is constructed that defines various decision conflict types for clinical decision conflict detection, and providing solutions for conflict resolution.

Traditional medical guidance becomes increasingly unsatisfactory, as the care of patients should be centered around not just clinical symptoms but also their values and preferences. A method is proposed, in this paper, to fuse clinical knowledge and patient preferences into an integrated knowledge graph. Objective data was extracted from semi-structured online medical service interfaces, and subjective emotional data from patient review pages. A prototype system was designed and implemented to demonstrate the feasibility of the method. The system can recommend a ranked list of doctors with the best matched clinical background as well as patient preferences. An evaluation was conducted via carrying out a survey of user groups upon the medical guidance options of a human nurse, the “We Doctor” system, and our prototype system.

Recent years have seen a surge in many deep learning research approaches to predict links in structured network data, however, these approaches seem to falter in its applicability. This paper seeks to propose a new model, named HLPGAM (Heuristics Link Prediction Graph Attention Mechanism), which combines probabilistic heuristics and attention mechanism to learn a more suitable way of predicting links in a given structured-network without relying on sophisticated feature engineering based on the statistical properties of a given node. The paper first aligns graphs and performs graph2Vec conversion using graph convolutions operation then it overcomes entity classification and link prediction limitation via an attention mechanism, i.e. to replace the normalization with data-dependent attention weights. For the entity classification problem, the experimental results have demonstrated that the HLP-GAM model can act as a competitive, end-to-end trainable graph-based encoder. For link prediction, the HLP-GAM model outperformed direct optimization of the factorization model and achieved competitive results on standard link prediction benchmarks. Our model achieves much better performance than other algorithms when he experimented both based on AIFB and AM dataset.