HeartAgent: An Autonomous Agent System for Explainable Differential Diagnosis in Cardiology

HeartAgent is an autonomous, cardiology-specific agent system that leverages specialized sub-agents and curated data to deliver explainable, high-accuracy differential diagnoses, significantly outperforming existing AI methods and enhancing clinical decision-making when assisting human experts.

The Gist

Imagine you are a detective trying to solve a very tricky medical mystery. A patient walks into the hospital with chest pain and shortness of breath. Is it a heart attack? A blood clot in the lung? Or maybe just severe anxiety? In cardiology, these different possibilities are called differential diagnoses. Getting this right is life-or-death, but it's incredibly hard because symptoms often overlap.

For a long time, computers tried to help doctors solve these mysteries, but they were like overconfident interns: they could guess, but they often made up facts, couldn't explain why they guessed, and didn't know enough about the specific rules of heart medicine.

Enter HeartAgent. Think of HeartAgent not as a single robot doctor, but as a highly organized, super-smart medical task force working together to solve the case.

The HeartAgent "Task Force"

Instead of one AI trying to do everything, HeartAgent breaks the job down into a team of specialized agents, each with a specific role:

1. The Specialist Predictor (The Lead Detective): This agent looks at the patient's data (heart rhythms, blood tests, notes) and makes the first list of suspects. "I think it might be Condition A or B."
2. The Generalist Examiner (The Skeptic): This agent is the "devil's advocate." It asks, "Wait, could this actually be a problem with the lungs or the stomach instead?" It makes sure the team doesn't get tunnel vision and miss non-heart causes.
3. The Specialist Reviewer (The Senior Consultant): This agent double-checks the work. It looks at the initial list and says, "Actually, based on these specific symptoms, we should also consider Condition C." It refines the list and adds the "why."
4. The Reference Verifier (The Librarian): This is the most crucial part for trust. Before the team presents its final answer, this agent runs to the library (a massive database of medical textbooks and guidelines) to find proof. It doesn't just say "It's a heart attack"; it says, "It's a heart attack, and here is the specific medical textbook page that proves it."

How It Works in Real Life

Imagine the HeartAgent system is a kitchen team preparing a complex meal (the diagnosis).

• The Predictor chops the vegetables (identifies symptoms).
• The Examiner tastes the soup and says, "Is it too salty? Maybe we need to check the spices (other diseases)."
• The Reviewer checks the recipe book to ensure the cooking method is correct.
• The Librarian pulls out the original cookbook to show the head chef (the human doctor) exactly which page says this recipe is safe and correct.

If one part of the team makes a mistake, the others catch it. If the team can't find a recipe in the book for a specific ingredient, they admit they don't have proof, rather than making something up.

Why This Is a Big Deal

The researchers tested HeartAgent on thousands of real patient records and found some amazing results:

• It's Smarter: HeartAgent was 36% more accurate than the best existing AI methods at guessing the right diagnosis. It's like upgrading from a standard GPS to a satellite-guided navigation system that knows every backroad.
• It's Honest: Unlike other AIs that might "hallucinate" (make things up), HeartAgent provides evidence. It shows its work, just like a math student showing their steps.
• It Makes Doctors Better: When human cardiologists used HeartAgent as a partner, their own accuracy jumped by 27%. It's like giving a master chef a sous-chef who has read every cookbook in the world and never forgets a detail. The human brings the experience and intuition; the AI brings the massive memory and strict adherence to rules.

The Bottom Line

HeartAgent is a new kind of medical assistant that doesn't just give an answer; it builds a case. It acts like a team of expert detectives who check each other's work and bring the evidence to the table. By combining the speed of computers with the reliability of verified medical books, it helps doctors make faster, safer, and more trustworthy decisions for patients with heart trouble.

In short: It's not about replacing the doctor; it's about giving the doctor a super-powered, fact-checking sidekick to ensure no stone is left unturned.

Technical Summary

Here is a detailed technical summary of the paper "HeartAgent: An Autonomous Agent System for Explainable Differential Diagnosis in Cardiology."

1. Problem Statement

Cardiovascular diseases are a leading cause of global mortality and morbidity, requiring precise differential diagnosis (identifying and prioritizing multiple plausible etiologies). Current AI-based diagnostic tools face three critical limitations in this domain:

• Insufficient Domain Knowledge: Generalist models often lack deep, specialized cardiology knowledge and adherence to clinical guidelines.
• Inadequate Reasoning: Existing models struggle with the complex, multi-step reasoning required to distinguish between conditions with overlapping symptoms (e.g., differentiating aortic dissection from myocardial infarction).
• Lack of Interpretability: Many models operate as "black boxes," providing predictions without transparent reasoning or verifiable evidence, which hinders clinician trust and adoption.

2. Methodology: The HeartAgent Framework

HeartAgent is an autonomous, multi-agent system designed specifically for cardiology. It orchestrates specialized sub-agents to perform structured reasoning, leveraging Large Language Models (LLMs) as a backbone while integrating external tools and curated knowledge.

Core Architecture & Workflow

The system processes multimodal clinical data (notes, ECGs, echocardiograms, labs) through a four-stage pipeline:

1. Data Preprocessing & Tool Integration:
   • Modality-Specific Tools: Vision-language models analyze images (ECG, Echo, CT); NeuroKit2 processes ECG waveforms; rule-based calculators compute risk scores.
   • Knowledge Retrieval: Agents dynamically access a Curated Knowledge Base (guidelines, textbooks), a Case Repository (similar past cases), and Web Searchers (PubMed/Wikipedia) to bridge knowledge gaps.
2. Collaborative Agent Reasoning:
   • Specialist Predictor Agent: Generates initial diagnostic hypotheses and rationales based on processed data.
   • Generalist Examiner Agent: Explores non-cardiology differential diagnoses to prevent missed diagnoses (e.g., pulmonary embolism).
   • Specialist Reviewer Agent: Critiques and refines the initial predictions, adding plausible cardiac conditions and verifying logic.
3. Self-Verification:
   • The Predictor Agent performs iterative self-assessment, integrating feedback from the Examiner and Reviewer to filter out unlikely diagnoses and refine rationales.
4. Reference Verification:
   • A dedicated Reference Agent retrieves supporting literature for the final explanations. It uses a hybrid retrieval strategy (BM25 for lexical matching + MedCPT for semantic re-ranking) to cite specific sections from medical textbooks and guidelines, ensuring factual accuracy.

Key Technical Components

• Base Models: Evaluated using open-source LLMs (Llama-3.3-70B, Qwen-2.5-32B, MedGemma-27B) and commercial models (GPT-5, Claude, DeepSeek-R1).
• Knowledge Resources: Includes 109 medical guidelines, 10 textbooks, and a repository of 4,000 curated clinical notes.
• Verification Mechanism: A two-stage retrieval system ensures that every diagnostic claim is backed by authoritative evidence, with a manual verification step to confirm the retrieved text supports the claim.

3. Key Contributions

1. Specialized Agentic Architecture: The first system to orchestrate multiple specialized agents (Predictor, Examiner, Reviewer) specifically for cardiology differential diagnosis, enabling complex, collaborative reasoning.
2. Explainability & Traceability: Unlike standard "black box" models, HeartAgent generates transparent reasoning trajectories and provides verifiable references to medical literature for every claim, significantly reducing hallucinations.
3. Human-AI Collaboration: Demonstrated that the system acts as a powerful decision-support tool, significantly enhancing the performance of human experts when used in tandem.

4. Experimental Results

The system was evaluated on three datasets: MIMIC-IV (ICU data), UMN (private hospital data), and NEJM (public case challenges).

• Diagnostic Accuracy:
  • MIMIC Dataset: HeartAgent improved top-3 diagnostic accuracy by >36% over baseline methods (Chain-of-Thought) across all tested LLMs.
  • UMN Dataset: Achieved a >20% improvement over baselines.
  • NEJM Dataset: Outperformed standalone commercial LLMs (e.g., GPT-5) by 29% on average. Notably, HeartAgent built on smaller open-source models (e.g., MedGemma-27B) matched or exceeded the performance of massive commercial models.
• Explanation Quality:
  • Outperformed baselines by 37–46% in explanation scores on the MIMIC dataset.
  • Factuality: 86% of HeartAgent's explanations scored >4/5 on a factual accuracy scale (vs. 56% for baselines), indicating a drastic reduction in medical hallucinations.
• Reference Verification:
  • Achieved 92% precision and 80% recall in retrieving correct supporting references for diagnostic claims.
• Human-AI Collaboration:
  • Clinicians assisted by HeartAgent showed a 26.9% increase in diagnostic accuracy and a 22.7% increase in explanation quality compared to unaided experts.
  • The system helped clinicians identify the most probable diagnosis (Top-1) by providing clear, evidence-backed rationales.

5. Significance and Impact

• Clinical Trust: By providing transparent reasoning and verifiable citations, HeartAgent addresses the "black box" problem, making AI outputs actionable and trustworthy for clinicians.
• Bridging the Knowledge Gap: The system effectively compensates for the lack of specialized cardiology knowledge in general LLMs by integrating curated guidelines and dynamic web search.
• Scalability: The framework is model-agnostic, performing robustly across various LLM sizes, suggesting that architectural design (agentic workflow) is more critical than sheer model parameter count for specialized medical tasks.
• Future of Care: HeartAgent represents a shift toward evidence-based AI, where diagnostic support is not just a prediction but a reasoned argument backed by medical literature, potentially reducing misdiagnosis rates and improving patient outcomes in cardiology.