Agentic Clinical Dialogue Guide

Overview

A curated collection of papers on AI agents for clinical dialogue — systems that conduct patient interviews, perform differential diagnosis, explain medical information, and support clinical decision-making through conversation. Covers medical QA benchmarks, patient simulation, clinical reasoning chains, and safety considerations unique to healthcare AI.

Research Landscape

Agentic Clinical Dialogue
├── Patient-Facing Agents
│   ├── Symptom checkers
│   ├── Triage systems
│   ├── Health information
│   └── Follow-up management
├── Clinician-Facing Agents
│   ├── Diagnostic support
│   ├── Treatment recommendation
│   ├── Clinical documentation
│   └── Literature integration
├── Clinical Reasoning
│   ├── Differential diagnosis
│   ├── History taking
│   ├── Physical exam interpretation
│   └── Test ordering
├── Patient Simulation
│   ├── Standardized patients (SP)
│   ├── Medical education
│   └── Agent evaluation
└── Safety & Ethics
    ├── Hallucination in medicine
    ├── Bias in clinical AI
    ├── Liability frameworks
    └── Informed consent

Key Systems

System	Focus	Approach
AMIE	Diagnostic dialogue	LLM with clinical reasoning
Med-PaLM	Medical QA	Finetuned on medical data
ChatDoctor	Patient consultation	LLaMA + medical knowledge
AgentClinic	Clinical evaluation	Simulated clinical encounters
ClinicalAgent	Decision support	Multi-step clinical reasoning

Benchmarks

benchmarks = {
    "MedQA (USMLE)": {
        "task": "US Medical Licensing Exam questions",
        "size": "11,450 questions",
        "metric": "Accuracy",
    },
    "PubMedQA": {
        "task": "Biomedical yes/no/maybe QA",
        "size": "1,000 expert-labeled",
        "metric": "Accuracy",
    },
    "AgentClinic": {
        "task": "Simulated clinical encounters",
        "size": "Various patient scenarios",
        "metric": "Diagnostic accuracy + safety",
    },
    "MedMCQA": {
        "task": "Indian medical entrance MCQs",
        "size": "194k questions",
        "metric": "Accuracy",
    },
    "HealthSearchQA": {
        "task": "Consumer health search questions",
        "size": "3,375 questions",
        "metric": "Expert evaluation",
    },
}

for name, info in benchmarks.items():
    print(f"\n{name}:")
    print(f"  Task: {info['task']}")
    print(f"  Size: {info['size']}")