Patient Agent in Digital Healthcare

• Patient Agent is a computational entity that mimics human patient behavior through LLMs and structured workflows.
• It is applied in clinical benchmarking, medical education, self-triage, and scheduling to enhance digital healthcare interactions.
• Emerging methods emphasize temperament-driven dialogue, robust evaluation metrics, and privacy-centric protocols for realistic simulations.

A patient agent is a computational entity—frequently realized as a LLM or structured multi-agent workflow—that simulates the perspective, knowledge, and behaviors of a human patient within digital healthcare systems. Across clinical benchmarking, medical education, and agentic telemedicine, the patient agent operationalizes the patient's role in tasks ranging from symptom reporting and dialogue participation to autonomous reasoning for scheduling and resource allocation. Recent advances span temperament-driven conversational simulation, multi-turn interaction fidelity, privacy-centric self-triage, and dynamic response verification. The following sections analyze prominent paradigms and implementations of patient agents, their operational architectures, behavioral policies, evaluation strategies, and the significance for emergent multi-agent medical systems.

1. Architectural Paradigms and Core Definitions

Patient agents are instantiated in a variety of forms, reflecting divergent goals such as conversational realism, information control, or operational optimization. The predominant paradigm leverages LLM-backed simulators where the agent embodies (a) a fixed memory (case profile, symptoms, or EHR-derived features) and (b) a deterministic or stochastic dialogue policy, often modulated by explicit behavioral attributes or clinical parameters.

In 3MDBench, the Patient Agent is a text-only LLM (Llama-3-8B) receiving a structured case context—basic complaint, atomic additional complaints, and a temperament profile—governing its turn-taking and symptom disclosure (Sviridov et al., 26 Mar 2025). Other frameworks, such as MAQuE, programmatically encode additional layers of behavioral complexity: the agent is initialized with a clinical vignette decomposed into atomic information units (AIUs), then parameterized at runtime with linguistic, cognitive, and emotional variation (Gong et al., 29 Sep 2025). In scheduling/optimization scenarios, patient agents are lightweight objects tracking arrival time, task list, and priority, enabling distributed, partially observed resource allocation (Mageshwari et al., 2012).

Multi-agent designs frequently partition the patient-facing logic across special-purpose agents (symptom checker, medication, appointment) orchestrated by a protocol engine (e.g., Model Context Protocol in Agentic-AI Healthcare (Shehab, 25 Sep 2025)). In clinical simulation environments, the patient is an LLM process consuming both static (demographics, symptoms) and dynamic (dialogue history) context buffers, producing responses sampled from prompt-conditioned LLMs (Almansoori et al., 28 Mar 2025, Rashidian et al., 4 Jun 2025).

2. Behavioral Policies and Variability Mechanisms

Patient agent behaviors are derived from prompt engineering, conditional logic, and stochastic sampling over underlying LLMs, or by explicit finite automata in structured decision support. In 3MDBench, four temperament prompts govern dialogue verbosity, question-asking tendency, communication style, treatment acceptance, and emotional involvement—each encoded in the initial system prompt and enforced turn-by-turn through deterministic rules and sampling at temperature $\tau=0.6$ (Sviridov et al., 26 Mar 2025). The dialogue loop is bounded: after a maximum number of utterances or a conclusive diagnosis, a termination token is returned.

MAQuE introduces variational axes:

• Linguistic style: stylized paraphrasing from formal to colloquial or dialectal,
• Cognitive status: induction of slips, hesitations, misunderstanding,
• Emotional state: dynamic modulation in response to inquiry tone or dialogue events. Per-turn, the agent performs AIU selection (disclosure control), noise injection (linguistic/cognitive/emotional), and rendering to output, producing high-fidelity diversity and realism (Gong et al., 29 Sep 2025).

In triage and scheduling, behavioral policy is replaced by procedural logic dictated by case state—e.g., in DOPSG, each Patient Agent passively advances through its task list, surrendering migration decisions to resource agents (Mageshwari et al., 2012). In privacy-first orchestrators, policies are modular: symptom and medication agents validate and escalate user reports, with each message adhering to strict schema and compliance wrappers (Shehab, 25 Sep 2025).

3. Dialogue Management, Information Control, and Turn-Taking

Patient agents implement historically-informed information ordering and release, simulating real patient-doctor interactional dynamics. In multi-turn dialogue environments such as DoctorAgent-RL and MedAgentSim, the agent maintains a hidden profile and sequentially reveals symptom attributes only in response to targeted clinician queries; out-of-scope or repetitive requests are flagged, while refusal behaviors are explicitly modeled (Feng et al., 26 May 2025, Almansoori et al., 28 Mar 2025).

In AIPatient, six agents coordinate to retrieve graph-based facts from EHR-derived knowledge graphs, generating responses mapped to personality profiles for realism (Yu et al., 2024). TriageMD’s chat agent operationalizes clinical questions in sync with structured flowchart logic, managing response restatement, uncertainty clarification, and final action recommendation generation (Liu et al., 16 Nov 2025).

Protocols are frequently defined in JSON-based message structures passing between system (“system”/“user”/“assistant” role tags), allowing seamless interface with other clinical agents or downstream assessment modules (Sviridov et al., 26 Mar 2025, Shehab, 25 Sep 2025). Termination is strictly managed, either through explicit intent detection, action endpoints, or token limits.

4. Evaluation Criteria and Calibration

Patient agent calibration is typically a prompt-driven, zero-shot/few-shot design, with model selection and hyperparameter tuning conducted through performance audits on held-out dialogue sets. In 3MDBench, LLM candidates were ranked on instruction following, relevance, and a factuality metric defined as the proportion of patient utterances with cosine similarity $>$0.8 to true symptoms (Sviridov et al., 26 Mar 2025). MAQuE documents an ablation regime where the addition of disclosure gating, linguistic variation, and noise injection to the patient agent significantly alters downstream doctor agent performance, confirming the patient model’s behavioral impact (Gong et al., 29 Sep 2025).

Clinical consistency and relevance are benchmarked via human expert review and standardized rubrics. In EHR-mapped simulations (Rashidian et al., 4 Jun 2025), patient agent consistency with vignette is evaluated as the percent of conversations matching key vignette facts ($97.7\%$), with case summary relevancy reaching $99.2\%$. PatientHistory agent outputs are scored through claim-level model-based entailment, measuring recall and succinctness against reference summaries (Codella et al., 8 Sep 2025).

5. Transparency, Safety, Privacy, and Compliance

Multiplexed patient-agent systems emphasize explainability, traceability, and regulatory compliance. Agentic-AI Healthcare applies role-based access control, AES-GCM field-level encryption (mathematically: $C = \mathrm{AES\_Encrypt\_CTR\_Mode}(K, IV, P)$ and $$T = \mathrm{GHASH}(H, AAD, C) \oplus \mathrm{AES\_Encrypt}(K, J_0)$$), and a hash-chained audit log, with policy-permitted actions enforced at each transaction (Shehab, 25 Sep 2025). TriageMD encodes every decision node and transition in auditable graph structures, allowing clinicians to reconstruct and verify the diagnostic logic (Liu et al., 16 Nov 2025).

Dialogue explainability is supported by explicit reasoning traces—symptom-checking agents return not only results but the chain of decision rules that led to each structured output (enumerated in JSON fields such as "trace") (Shehab, 25 Sep 2025). Privacy-first triage and summary-generation workflows are further facilitated by local, data-free evaluation infrastructures (e.g. TBFact) that avoid transmission of clinical data (Codella et al., 8 Sep 2025).

6. Applications, Limitations, and Future Directions

Patient agents underpin a broad range of applications:

• Benchmarking and Evaluation: As realistic conversational partners or simulation environments for LVLM or RL-based doctor agents, patient agents stress-test information-seeking, empathy, and diagnostic capability across linguistic and behavioral variances (Sviridov et al., 26 Mar 2025, Gong et al., 29 Sep 2025, Almansoori et al., 28 Mar 2025).
• Medical Education: Simulated patients constructed from EHR-derived or synthetic (Patient-Zero) knowledge bases support reproducible, diverse, and privacy-preserving training at scale (Yu et al., 2024, Lai et al., 14 Sep 2025).
• Self-Triage and Scheduling: Structured flowchart-driven or modular dialog agents enable robust, auditable, multilingual, and privacy-compliant triage and coordination (Liu et al., 16 Nov 2025, Shehab, 25 Sep 2025, Mageshwari et al., 2012).
• Specialized Assessment: Conversational agents for ADRD early detection illustrate domain-specific design, balancing systematic coverage with patient comfort and response latency (Breithaupt et al., 14 Sep 2025).

Limitations include the risk of LLM factual inaccuracy or hallucination, restricted scope imposed by dataset or knowledge base constraints (e.g., EHR specificity, coverage gaps), and the tendency for prompt-driven agents to omit longitudinal, follow-up, or multi-party context unless explicitly engineered (Rashidian et al., 4 Jun 2025, Yu et al., 2024, Gong et al., 29 Sep 2025). Many frameworks are evaluated retrospectively or in simulation; real-world clinical integration, ongoing safety validation, and regulatory approval remain areas for extension.

7. Significance and Outlook

Patient agents have become foundational in the architecture of multi-agent clinical AI, enabling controlled, reproducible, and systematically variable evaluation of diagnostic and interactional capabilities. Their sophistication—ranging from temperament modulation and atomic information control to full-resolution clinical triage and compliance integration—supports not only empirical progress in LVLM performance but also the design of scalable, transparent, and privacy-preserving digital health systems. Pioneer works have established reproducible evaluation metrics, behavioral ablation tools, privacy-centric orchestration, and knowledge-grounded dialogue, providing a reference foundation for subsequent generation and benchmarking of medical AI (Sviridov et al., 26 Mar 2025, Shehab, 25 Sep 2025, Gong et al., 29 Sep 2025, Yu et al., 2024, Rashidian et al., 4 Jun 2025, Liu et al., 16 Nov 2025).