Agentic AI in Healthcare & Medicine: A Seven-Dimensional Taxonomy for Empirical Evaluation of LLM-based Agents

Technical Summary: Agentic AI in Healthcare & Medicine: A Seven-Dimensional Taxonomy for Empirical Evaluation of LLM-based Agents

Problem Statement
The rapid emergence of Large Language Model (LLM)-based agents in healthcare and medicine has demonstrated competence in tasks ranging from Electronic Health Record (EHR) analysis and differential diagnosis to treatment planning. However, the existing literature lacks a unified framework for evaluating these systems. Current studies are fragmented, consisting either of broad surveys that lack depth or narrow dives into specific capabilities (e.g., memory or reasoning) without a common operational frame. This absence of a structured taxonomy hinders the ability to compare methodologies, assess capability prevalence, and identify gaps between early-stage prototypes and dependable clinical systems.

Methodology
To address this gap, the authors conducted a systematic review of 49 research papers published between October 2023 and June 2025. The selection process followed a three-step filtering protocol using resources such as Google Scholar, PubMed, DBLP, Scopus, and ArXiv, adhering to PRISMA-style guidelines. Inclusion criteria required that the LLM be the core controller of the agent's planning and decision-making and that the study explicitly address healthcare or medicine tasks.

The authors developed a novel seven-dimensional taxonomy comprising 29 operational sub-dimensions to evaluate the selected studies. The dimensions are:

1. Cognitive Capabilities (Planning, Perception, Action, Meta-Capabilities, Consistency & Conflict Resolution)
2. Knowledge Management (External Knowledge Integration, Memory Module, Dynamic Updates & Forgetting)
3. Interaction Patterns (Conversational Mode, Event-Triggered Activation, Human-in-the-Loop, Error Recovery)
4. Adaptation & Learning (Drift Detection & Mitigation, Reinforcement-Based Adaptation, Meta-Learning & Few-Shot)
5. Safety & Ethics (Safety Guardrails & Adversarial Robustness, Bias & Fairness, Privacy-Preserving Mechanism, Regulatory & Compliance Constraints)
6. Framework Typology (Multi-Agent Design, Centralized Orchestration)
7. Core Tasks & Subtasks (Clinical Documentation, Medical Question Answering, Triage, Diagnostic Reasoning, Treatment Planning, Drug Discovery, Patient Interaction, Benchmarking)

Each study was mapped to this taxonomy using a labeling rubric: Fully Implemented (✓), Partially Implemented (∆), or Not Implemented (✗). The authors employed a conservative assignment strategy, defaulting to "Partially Implemented" for implicit claims and "Not Implemented" for assertions lacking concrete procedure or evaluation.

Key Contributions

• Structured Taxonomy: The primary contribution is the definition of a 7-dimensional, 29-sub-dimension taxonomy specifically designed for the empirical evaluation of LLM-based agents in healthcare.
• Empirical Baseline: The paper provides the first large-scale quantitative analysis of capability prevalence and co-occurrence patterns across 49 studies, moving beyond qualitative overviews.
• Visualization and Mapping: The authors provide detailed tables and diagrams mapping specific papers to the taxonomy, serving as a navigational index for the field.
• Identification of Asymmetries: The analysis reveals clear disparities between well-developed capabilities (e.g., information retrieval) and critical gaps in safety, adaptation, and action-oriented tasks.

Key Results and Findings
The empirical analysis surfaces significant asymmetries in the current state of Agentic AI in healthcare:

• Cognitive Capabilities: There is a marked lack of robust mechanisms for long-horizon task decomposition and self-monitoring. Planning is absent in ~45% of studies, Meta-Capabilities in ~53%, and Consistency & Conflict Resolution in ~61%. While Perception and Action show higher implementation rates, they often lack end-to-end reliability.
• Knowledge Management: External Knowledge Integration is widely realized (76% fully implemented), often via Retrieval-Augmented Generation (RAG). However, Dynamic Updates & Forgetting is largely absent (98% not implemented), indicating agents struggle to maintain knowledge hygiene or adapt to evolving guidelines.
• Interaction Patterns: While Conversational Mode is relatively common (43% fully implemented), Event-Triggered Activation is rare (92% not implemented), suggesting a heavy reliance on manual prompting rather than autonomous monitoring of external signals. Human-in-the-Loop and Error Recovery mechanisms are also sparse.
• Adaptation & Learning: This dimension shows the lowest adoption. Drift Detection & Mitigation is essentially absent (0% fully implemented, ~98% not implemented). Reinforcement-Based Adaptation and Meta-Learning are similarly underdeveloped, with most systems relying on static prompts.
• Safety & Ethics: Implementation maturity is limited. Safety Guardrails and Bias & Fairness are fully implemented in only 6% and ~10% of studies, respectively. Regulatory & Compliance Constraints are the most underbuilt (86% not implemented), with few studies demonstrating verifiable runtime compliance.
• Framework Typology: Multi-Agent Design is the dominant pattern (82% fully implemented), reflecting a shift toward role-based specialization. However, Centralized Orchestration remains mostly partial (57%), indicating that while agents are modular, the control layers managing them are not yet fully robust.
• Core Tasks: Information-centric tasks like Medical Question Answering (57% fully implemented) and Benchmarking (80% fully implemented) lead in maturity. In contrast, action-oriented and discovery tasks lag significantly; Treatment Planning & Prescription is fully implemented in only ~20% of studies, and Drug Discovery in ~18%.

Significance and Claims
The paper claims that its findings provide an empirical baseline indicating that current LLM-based agents in healthcare excel at retrieval-grounded advising and information synthesis. However, the field requires stronger adaptation, compliance platforms, and safety mechanisms to transition from early-stage research prototypes to dependable clinical systems.

The authors emphasize that while the field has made significant strides in modular design and information retrieval, critical gaps in drift detection, dynamic knowledge updates, error recovery, and regulatory alignment persist. They argue that without addressing these specific asymmetries, the deployment of these agents in high-stakes clinical environments remains risky. The paper positions itself as a foundational resource to guide future research toward robust, safe, and ethically responsible agent development, explicitly noting that scalability and production-scale engineering challenges lie beyond its scope.