PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Shubham Vatsal, Harsh Dubey, Aditi Singh

arXiv:2602.04813v1cs.AIcs.CY

TL;DR

Healthcare research on LLM-based agents lacks a common framework for comparing their multifaceted capabilities and designs. This paper reviews 49 studies with a seven-dimensional, 29-sub-dimension taxonomy and explicit implementation labels, revealing strong information-centric and multi-agent patterns alongside major gaps in adaptation, activation, safety, and action-oriented tasks.

  • Problem

    Existing literature lacks a structured framework for evaluating the multifaceted capabilities and system designs of LLM-based agents in healthcare and medicine.

  • Method

    The survey maps 49 healthcare and medicine studies to 7 dimensions and 29 sub-dimensions using explicit inclusion criteria and Fully Implemented, Partially Implemented, and Not Implemented labels.

  • Results

    The analysis finds strong external knowledge integration and multi-agent design, while event-triggered activation, drift detection, safety maturity, and treatment-oriented tasks remain substantially underimplemented.

  • Takeaways & Limitations

    The taxonomy provides a structured basis for identifying capability clusters and prioritizing more robust, reliable, and ethical healthcare agents.

  • Takeaways & Limitations

    The evidence and analysis are limited to peer-reviewed or archival research papers rather than production systems, leaving scalability as an open challenge.

Abstract

from arXiv · show

Large Language Model (LLM)-based agents that plan, use tools and act has begun to shape healthcare and medicine. Reported studies demonstrate competence on various tasks ranging from EHR analysis and differential diagnosis to treatment planning and research workflows. Yet the literature largely consists of overviews which are either broad surveys or narrow dives into a single capability (e.g., memory, planning, reasoning), leaving healthcare work without a common frame. We address this by reviewing 49 studies using a seven-dimensional taxonomy: Cognitive Capabilities, Knowledge Management, Interaction Patterns, Adaptation & Learning, Safety & Ethics, Framework Typology and Core Tasks & Subtasks with 29 operational sub-dimensions. Using explicit inclusion and exclusion criteria and a labeling rubric (Fully Implemented, Partially Implemented, Not Implemented), we map each study to the taxonomy and report quantitative summaries of capability prevalence and co-occurrence patterns. Our empirical analysis surfaces clear asymmetries. For instance, the External Knowledge Integration sub-dimension under Knowledge Management is commonly realized (~76% Fully Implemented) whereas Event-Triggered Activation sub-dimenison under Interaction Patterns is largely absent (~92% Not Implemented) and Drift Detection & Mitigation sub-dimension under Adaptation & Learning is rare (~98% Not Implemented). Architecturally, Multi-Agent Design sub-dimension under Framework Typology is the dominant pattern (~82% Fully Implemented) while orchestration layers remain mostly partial. Across Core Tasks & Subtasks, information centric capabilities lead e.g., Medical Question Answering & Decision Support and Benchmarking & Simulation, while action and discovery oriented areas such as Treatment Planning & Prescription still show substantial gaps (~59% Not Implemented).

I. INTRODUCTION

LLM-based agents are increasingly applied to healthcare and medicine, but the literature lacks a shared framework for evaluating their capabilities, designs, and scope. This survey addresses that gap by organizing 49 studies within a seven-dimensional taxonomy.

  • Background: LLM scaling and prompt engineering have advanced reasoning, planning, and context-sensitive interactions across domains.These developments have supported the transition from predictive modeling toward more interactive agent behavior.
  • Healthcare Applications: Healthcare applications include clinical documentation, medical question answering, differential diagnosis, patient triage, laboratory interpretation, and EHR management.Reported uses span workflow automation, decision support, and clinical research assistance.
  • Agentic AI: LLM-based agents combine reasoning, planning, memory management, and interactive execution to support context-aware problem solving.Related approaches include modular reasoning architectures and feedback-based behavioral refinement.
  • Survey Scope: The survey evaluates 49 healthcare and medicine studies across 7 dimensions and 29 sub-dimensions.The taxonomy covers cognitive capabilities, knowledge management, interaction patterns, adaptation, safety, framework typology, and core tasks.
  • Scope Boundary: The review focuses on agents in which an LLM drives planning and decision-making while using tools, maintaining memory, or taking actions.Purely symbolic, rule-based, or non-LLM-controlled architectures are excluded; hybrid and multimodal systems qualify when the LLM coordinates behavior.

B. LITERATURE SEARCH PROCESS

The survey uses a staged literature-search and selection process to produce a curated corpus of 49 studies. Its taxonomy and scoring rubric emphasize reproducible cross-study comparison while acknowledging boundary ambiguity.

  • Search Process: The review begins with a structured three-step filtering process for curating the research corpus.The process covers search, eligibility screening, and publication-date restriction.
  • Search Process: The initial search across Google Scholar, PubMed, DBLP, Scopus, and ArXiv yielded 137 articles after preliminary manual filtering.Queries targeted LLM-based agents across healthcare, medicine, EHR management, clinical documentation, and related tasks.
  • Eligibility Criteria: Studies were retained when an LLM was central to the agent and the work explicitly addressed healthcare or medicine.Interdisciplinary biomedical studies were assessed case by case according to their predominant contribution.
  • Corpus Construction: Publication dates were restricted to October 2023 through June 2025, producing a curated set of 49 research papers.A PRISMA-style diagram summarizes identification, screening, eligibility assessment, and final inclusion.
  • Evaluation Framework: The taxonomy combines open coding, structured synthesis, and cross-referencing with prior frameworks to define 7 dimensions and 29 sub-dimensions.The resulting process is intended to keep the taxonomy grounded in the literature and empirically reproducible.
  • Scoring Rubric: The Fully Implemented, Partially Implemented, and Not Implemented labels are decision aids rather than strict mutually exclusive thresholds.Conservative assignment uses verifiable descriptions, with partial labels for implicit or simulation-only evidence and full labels for explicit end-to-end demonstrations.

A. COGNITIVE CAPABILITIES

Healthcare agents must transform heterogeneous clinical inputs into representations, reason over them, and convert decisions into executable actions. The survey operationalizes these capabilities through planning, perception, tool use, and reflective control.

  • 1) Planning: Planning decomposes long-horizon clinical goals into milestones, decision points, and executable action structures.Structured lookahead supports care-pathway drafting, triage routing, prior-authorization checks, and treatment planning.
  • Perception: Perception converts notes, laboratory data, images, device streams, and guidelines into clinically meaningful representations for downstream reasoning.Multimodal alignment and domain-tuned language models help ground plans in patient-specific context.
  • Action and Control: Agents interleave reasoning with tool calls and use self-monitoring to critique intermediate products and support escalation when evidence is weak.Reflective loops and diversity-aware decoding can stabilize conclusions and surface deferral conditions.
  • 1) Planning: Planning is Fully Implemented when agents dynamically construct multi-step plans, compare alternatives or refine plans, and select tools appropriately.Task breakdown or dynamic role assignment without substantive strategy generation is scored Partial, while fixed workflows are Not Implemented.

2) Perception (Input Processing)

Reliable healthcare agents depend on robust input processing, controlled execution, knowledge integration, memory, and contradiction handling. The evaluation framework distinguishes explicit operational capabilities from simple text generation or unsupported claims.

  • Perception: Perception transforms heterogeneous clinical inputs through normalization, denoising, temporal alignment, and cross-modal fusion into model-usable representations.The resulting representations are designed for downstream reasoning and action, with provenance sensitivity where supported.
  • Action: Action is Fully Implemented only when systems demonstrate tool use and external execution beyond text generation.Basic API calls or database lookups are Partial, while text-only outputs are Not Implemented.
  • Meta-Capabilities: Consistency and conflict resolution reconcile contradictions from model memory, external retrieval, and multi-agent outputs.Layered strategies include contradiction detection and ensemble agreement protocols to keep recommendations coherent and evidence-based.
  • Knowledge Management: Knowledge management routes between parametric knowledge, retrieval systems, and non-parametric memory when internal recall is insufficient.External resources can provide current guidelines and trial evidence, while short- and long-term memory support continuity across encounters.
  • Knowledge Management: Self-reflective retrieval and targeted parametric edits support knowledge updating, citation reliability, and reduction of stale information.These mechanisms are presented as important for aligning agent knowledge with changing clinical evidence.

1) External Knowledge Integration

External Knowledge Integration augments an agent’s parametric knowledge with retrieved non-parametric information, while Dynamic Updates & Forgetting keeps stored knowledge current by adding new evidence and removing stale facts.

  • External Knowledge Integration: External Knowledge Integration augments model-weight knowledge with non-parametric information retrieved from domain-specific repositories at inference time.Repositories may include clinical databases, medical knowledge graphs, regulatory guidelines, and vector-indexed corpora.
  • External Knowledge Integration: Fully realized implementations combine structured retrieval with arbitration between internal recall and external evidence.This arbitration supports outputs that are temporally current and contextually precise.
  • Memory Module: Memory modules support continuity and longitudinal reasoning by storing and retrieving information beyond the transient context window.Long-term stores may capture episodic consultation narratives and semantic abstractions such as knowledge graphs.
  • Dynamic Updates & Forgetting: Dynamic Updates & Forgetting requires continuous indexing, temporal re-weighting, and explicit removal of stale or retracted facts.The rubric treats both active addition and explicit forgetting as necessary for full implementation.

1) Conversational Mode

Conversational Mode emphasizes persistent, resumable dialogue rather than single-shot replies, while broader interaction patterns include event-driven activation and human control during execution.

  • Conversational Mode: Conversational Mode requires turn-based interaction that preserves context across turns and supports resumption after interruptions.Single-query processing without dialogue state, session continuity, or conversational context is not considered conversational mode.
  • Interaction Patterns: Event-Triggered Activation enables autonomous workflows initiated by external signals rather than explicit user prompts.Examples include webhooks, message buses, scheduled polling, and monitoring of clinical streams.
  • Interaction Patterns: Human-in-the-Loop inserts confirmation, approval, redirection, uncertainty pauses, or rollback controls during execution.Feedback is stored audibly to refine memory or downstream policies.
  • Error Recovery: Error Recovery detects, diagnoses, and corrects failures across parsing, planning, tool use, and system integration.Recovery may use disambiguation prompts, alternative tools, parameter repair, retries, and fallback procedures.
  • Adaptation & Learning: Adaptation and learning keep deployed agents calibrated as data distributions, task definitions, or usage contexts change.Drift monitoring and targeted mitigation are presented as safeguards against performance degradation.

1) Drift Detection & Mitigation

Drift Detection & Mitigation concerns maintaining agent reliability as inputs and tasks evolve, using monitoring to identify shifts and interventions to limit resulting degradation.

  • Drift Detection & Mitigation: Drift Detection & Mitigation monitors distributional and behavioral changes before clinical performance erodes.Detection can use embedding distributions, activation probes, or black-box statistical tests.
  • Drift Detection & Mitigation: Drift manifests when incoming inputs diverge from training data or task requirements shift subtly enough to degrade performance.Healthcare shifts may arise from changing populations or evolving clinical guidelines.
  • Adaptation & Learning: Reward-based adaptation translates evaluative feedback into structured signals that shape future decisions.Full implementations use explicit reinforcement learning such as RLHF, whereas partial implementations use simpler success signals or heuristic optimization.
  • Adaptation & Learning: Meta-Learning & Few-Shot adaptation targets rapid reconfiguration to new tasks or domains with minimal supervision.It combines meta-learning methods with in-context demonstrations at inference time.

1) Safety Guardrails & Adversarial Robustness

Safety and ethics require layered defenses against adversarial behavior, equitable performance across clinical populations, privacy controls, and evidence-based regulatory compliance.

  • Safety Guardrails & Adversarial Robustness: Safety Guardrails & Adversarial Robustness uses input hardening, resilient planning, output governance, automated detection, and adversarial stress testing.The framework treats mature safeguards as multi-stage controls with rapid remediation pathways.
  • Bias & Fairness: Bias & Fairness requires representative data, constrained disparate error, subgroup analysis, and implemented mitigation across clinically salient populations.Privacy measures alone do not satisfy the fairness criteria.
  • Privacy: Privacy controls restrict access to patient data through mechanisms such as scoped retrieval, deidentification, encryption, and differential privacy.The text also calls for threat modeling, leakage audits, and redaction verification.
  • Regulatory & Compliance Constraints: Regulatory & Compliance Constraints cover consent, lawful bases, role-scoped access, retention, cross-border transfers, and formal risk assessments.Compliance should be evidenced through testing and linked technical safeguards rather than merely asserted in policy.
  • Framework Typology: Multi-agent framework designs distribute cognitive labor across specialized roles that collaborate through discussion and consensus.Examples include multidisciplinary doctor agents with a supervising agent and centralized planning that delegates subtasks.

1) Multi-Agent Design

Multi-Agent Design evaluates how responsibilities are distributed among agents and how they coordinate, distinguishing genuine specialization from partial or absent coordination. The surrounding framework grounds these designs in concrete healthcare workflows and outputs.

  • Multi-Agent Design: Multi-Agent Design assesses responsibility distribution across one or multiple agents and their interactions during task completion.Specialized roles can include planners, retrievers, verifiers, imaging parsers, guideline validators, and patient-summary drafters.
  • Multi-Agent Design: Centralized Orchestration evaluates control over task, data, and decision flow, including sequencing, monitoring, and result integration.Healthcare orchestrators may reconcile evidence and escalate to human experts when uncertainty is high.
  • Multi-Agent Design: Core tasks define agents’ operational responsibilities and verifiable outputs across workflows from clinical data processing to recommendations and evaluation.This dimension connects agent design to concrete workflows and measurable endpoints.
  • Multi-Agent Design: Clinical record processing converts heterogeneous documentation and EHR artifacts into structured, clinician-usable outputs such as summaries, extracted entities, codes, predictions, and reports.The sub-dimension includes chart summarization, ICD mapping, longitudinal risk prediction, and draft report generation.
  • Multi-Agent Design: Medical Question Answering & Decision Support covers clinical question answering, retrieval-augmented recommendations, diagnostic questioning, and guideline-grounded support.The rubric distinguishes broad validated capabilities from narrow, simulation-only, or absent decision support.

3) Triage & Differential Diagnosis

Triage and differential diagnosis are evaluated through acuity decisions, defensible diagnostic alternatives, and multi-step evidence-based inference. Related treatment and discovery dimensions extend from patient-specific recommendations to translational and trial workflows.

  • Triage & Differential Diagnosis: Triage and Differential Diagnosis requires an acuity decision alongside a defensible list of diagnostic alternatives from first-contact information.Systems may elicit history, synthesize vitals, ask targeted follow-ups, align with ESI or KTAS, and recommend a receiving service.
  • Triage & Differential Diagnosis: Diagnostic Reasoning evaluates multi-step inference linking targeted evidence gathering, test selection, result interpretation, hypothesis revision, and guideline alignment.It can also include specialist-agent input and knowledge-graph traversal.
  • Treatment Planning & Prescription: Treatment Planning & Prescription translates diagnostic context into individualized plans and executable orders, including medication selection, dose personalization, contraindication checks, and counseling.Robust systems account for allergies, renal function, comorbidities, formularies, provenance, and pharmacist or prescriber verification.
  • Drug Discovery & Clinical Trial Design: Drug Discovery & Clinical Trial Design spans target interaction prediction, toxicity screening, hypothesis generation, lead optimization, protocol authoring, matching, and outcome prediction.The rubric expects task-specific benchmarks, tool-integrated workflows, and multi-tier validation for full implementation.

7) Patient Interaction & Monitoring

Patient Interaction & Monitoring covers communication, engagement, scheduling, symptom follow-up, adherence, and physiologic monitoring, while Benchmarking & Simulation evaluates agents under reproducible clinical conditions. Across the broader findings, knowledge integration is strong but maintenance and adaptation remain weak.

  • Patient Interaction & Monitoring: Patient Interaction & Monitoring includes two-way messaging, education, symptom follow-ups, adherence checks, reminders, scheduling, device integration, and personalized conversation management.Implementation is earned when at least one interaction, engagement, or appointment-management capability is present.
  • Patient Interaction & Monitoring: Benchmarking & Simulation Environment evaluates agents with standardized protocols, metrics, simulated roles and artifacts, multi-turn workflows, baselines, ablations, error analyses, and expert review.The emphasis is on rigorous and reproducible conditions that reflect clinical work.
  • Cross-cutting findings: External Knowledge Integration is strong, while Dynamic Updates & Forgetting and Error Recovery are weak across the surveyed dimensions.The cross-cutting analysis also reports relatively prevalent Conversational Mode and overall low adoption of Adaptation & Learning.
  • Cognitive Capabilities: Planning is absent in ∼45% of surveyed papers, Meta-Capabilities in ∼53%, and Consistency & Conflict Resolution in ∼61%.Perception and Action show more fully implemented systems, at ∼46% and ∼41% respectively.
  • Knowledge Management: External Knowledge Integration is fully implemented in 37/49 studies (∼76%), whereas Dynamic Updates & Forgetting is fully implemented in only 1/49 (∼2%) and not implemented in 33/49 (∼67%).The findings distinguish widespread retrieval from the much rarer continuous addition and explicit forgetting of knowledge.

C. DO CURRENT AGENTS BALANCE CONVERSATION,

Healthcare agents show uneven interaction and assurance maturity: conversational exchange is relatively common, while event-triggered activation, human oversight, recovery, adaptation, and safety mechanisms remain limited.

  • Interaction Patterns: Conversational Mode is fully implemented in ∼43% of studies and partially implemented in ∼6%, indicating relatively prevalent multi-turn exchange and state carryover.
  • Interaction Patterns: Only 3/49 (∼6%) studies fully implement Event-Triggered Activation, while 45/49 (∼92%) do not implement it.The surveyed systems rely mainly on manual prompts rather than external signals such as device telemetry or EHR events.
  • Interaction Patterns: Human-in-the-Loop is fully implemented in ∼12% of studies and not implemented in ∼86%, reflecting scarce mid-execution checkpoints and escalation paths.
  • Adaptation & Learning: Drift Detection & Mitigation is fully implemented in 0% of studies, partially implemented in ∼2%, and not implemented in ∼98%.Reinforcement-Based Adaptation and Meta-Learning & Few-Shot are also mostly not implemented, at ∼84% and 78%, respectively.
  • Safety & Ethics: Safety Guardrails & Adversarial Robustness are fully implemented in only ∼6% of studies, while ∼65% do not implement them.Regulatory & Compliance Constraints are even less developed, with ∼86% not implemented.

F. FRAMEWORK TYPOLOGY TRENDS ACROSS LLM-BASED HEALTHCARE AGENTS

Framework choices favor modular multi-agent architectures, while centralized orchestration is often only partial; task maturity is highest for information-centric work and lower for action, discovery, and clinical adoption.

  • Framework Typology: Multi-Agent Design is fully implemented in ∼82% of studies, with only ∼16% not implemented.Role-based compositions such as planners, retrievers, verifiers, and explainers are the dominant architectural pattern.
  • Framework Typology: Centralized Orchestration is partially implemented in ∼57% of studies, indicating that coordination layers often stop short of fully realized controllers.The cited analysis identifies gaps in global state, policy enforcement, and auditable sequencing.
  • Core Tasks & Subtasks: Medical Question Answering & Decision Support is fully implemented in 28/49 studies (∼57%), while Clinical Documentation & EHR Analysis is fully implemented in 22/49 (∼45%).
  • Core Tasks & Subtasks: Benchmarking & Simulation Environment is fully implemented in 39/49 studies (∼80%), with only 1/49 (∼2%) not implemented.
  • Core Tasks & Subtasks: Treatment Planning & Prescription is fully implemented in only 10/49 studies (∼20%), while Drug Discovery & Clinical Trial Design is fully implemented in 9/49 (∼18%) and not implemented in 39/49 (∼80%).The conclusion states that wider clinical adoption remains limited until adaptation, safety, and compliance become more robust.
Loading 2602.04813v1…