An Agentic AI system for disease diagnosis with explanations

Abstract

With the increasing complexities in the medical domain, the demand for autonomous, adaptable, scalable, and personalized Artificial Intelligence (AI)-based disease diagnostic systems is growing. However, disease diagnosis is an intricate task that involves complex subproblems such as processing multimodal clinical data, generating predictions using AI models, and diagnosis reasoning. The order in which subproblems are solved affects the output of disease diagnosis. In existing AI-based diagnosis systems, this order is managed by physicians, increasing their workload and the possibility of errors due to a lack of relevant expertise in computers. Our work addresses this challenge by proposing a framework for disease diagnosis with explanations that automates the process of disease diagnosis using Agentic AI. The proposed framework consists of autonomous multiple agents that can learn proactively, make decisions, and adapt to the environment without human intervention, compared to the reactive responses of traditional agents restricted to the humanly pre-coded rules. The proposed Agentic AI framework is backed by the Large Language Models (LLMs) that minimise human intervention and proactively take appropriate decisions through self-reflection. The framework uses different customised modules as tools to perform complex tasks. A Natural Language Processing-based framework, N2K mapper, and a Deep Learning model, CheXzero, are used as a tool for semantical processing of multimodal data from Electronic Health Records to a Knowledge Graph. Similarly, QLoRA is used to fine-tune an LLM for disease prediction, and Retrieval Augmented Generation is used as a tool for explanation generation based on standard medical guidelines. Different AI agents and tools are invoked autonomously to complete the given task. A web application has been developed based on the proposed Agentic AI framework, which has been tested on MIMIC-Eye, a real-world multimodal dataset. The application predicts two conditions: Congestive Heart Failure (CHF) and Urinary Tract Infection (UTI), with the relevant explanations for the predicted disease. Additionally, five LLMs, DeepSeek, LLaMA-3, Mistral, Qwen, and LLaMA-2, were fine-tuned for disease predictions, from which LLaMA-3 shows the best results with 0.88 AUROC and 0.90 AUPRC values in CHF and UTI predictions, respectively. In this work, we propose an Agentic AI framework for disease diagnosis and develop a web application based on the framework. It aims to provide a comprehensive disease diagnosis system that assists medical physicians and increases confidence in clinical decisions. • Proposes Agentic AI framework for disease diagnosis with explanations. • It uses multimodal clinical data from EHR to predict plausible diagnoses. • The agents use tools backed by technologies, LLM, Knowledge graphs, QLoRA, and RAG. • Developed framework-based application to diagnose CHF and UTI on MIMIC-Eye dataset.

Ähnliche Arbeiten