A MULTI-AGENT GENERATIVE AI FRAMEWORK FOR HEALTHCARE DECISION SUPPORT

Abstract

With the growing number of electronic health records (EHRs) and medical knowledge databases, there has been an increase in the adoption of Artificial Intelligence (AI) for intelligent healthcare decision support systems. However, current healthcare AI systems face limitations related to lack of explainability, lack of integration of medical knowledge, and lack of collaborative reasoning capabilities. In order to overcome these limitations, this research paper aims to design and develop a Multi-Agent Generative AI Framework for Healthcare Decision Support involving the use of MIMIC-IV clinical dataset, SMART-RAG, Clinical BERT embeddings, FAISS vector database, GPT-4 based LLMs, XAI and Human-AI Decision Support. The SMART-RAG module fetches information from medical guidelines, protocols of treatment, and scientific articles that help in enriching patient specific information with scientific evidence based on the literature available in the field. The framework uses Multi-agent architecture including Diagnosis Agent, Risk Assessment Agent, Treatment Recommendation Agent, and Explainability Agent. These agents collaborate together for the analysis of patient information and give clinical advice that is clear and transparent to understand. Experiments prove its promising performance by obtaining 0.93, 0.91, and 0.95 precision, recall, and faithfulness scores for the SMART-RAG retrieval process. The Diagnosis Agent, Risk Assessment Agent, Treatment Agent, and Explainability Agent obtained 94.5%, 92.7%, 93.6%, and 95.4% accuracies, respectively. Moreover, the GPT-4 decision engine managed to score 95.1% for Accuracy, 94.6% for Precision, 94.2% for Recall, and 94.4% for F1-Score, whereas the physician validation gave rise to a recommendation acceptance rate of 92.4%. The Explainability Agent improves the transparency of the process through interpretive reasoning behind the clinical recommendations, thus making physicians more confident about the decisions. Knowledge-driven reasoning, agent intelligence, explainable AI, and human-in-the-loop can be viewed as a robust solution for intelligent health care decision support.