Master Health Checkup Agent (MHCA)

• Master Health Checkup Agent (MHCA) is a modular, multi-agent system designed for longitudinal, preventive, and personalized health evaluations with strict privacy and multilingual interfaces.
• It integrates clinical workflow automation, advanced machine learning for risk stratification, and human-in-the-loop oversight to optimize clinical processes.
• The system employs dynamic task orchestration, robust encryption, and RBAC to ensure compliance and efficient inter-agent communication across healthcare platforms.

The Master Health Checkup Agent (MHCA) is a modular, multi-agent system designed for longitudinal, preventive, and personalized health evaluations, orchestrated by intelligent agents operating under strict privacy, compliance, and multilingual accessibility requirements. MHCA integrates clinical workflow automation, advanced machine learning for risk stratification, and human-in-the-loop mechanisms, purposely engineered to satisfy regulatory requirements and interface seamlessly with both consumers and clinicians across healthcare ecosystems.

1. System Architecture and Orchestration

The MHCA architecture is typically structured in three principal logical layers:

1.1 User Interface Layer:

A multilingual web front-end (React) enables both patients and clinicians to interact with the system via text or voice modalities. FastAPI middleware supports authentication, input sanitization, language detection (e.g., fastText), and communicates with the orchestration server (Shehab, 25 Sep 2025).

1.2 Agent Orchestration Layer:

Centralized orchestration is achieved via a Master Orchestration Agent, which maintains a global conversation context ($C^{master}$), delegates clinical subtasks—such as symptom analysis, medication guidance, and appointment scheduling—to dedicated sub-agents, and enforces workflow logic. Sub-agents operate with strictly locked-down, JSON-based schema prompts to ensure repeatability and control.

Communication across agents implements the Model Context Protocol (MCP), a formal message-based scheme with rigorously defined message tuples: $$m = (\text{from}, \text{to}, \text{tid}, \text{type}, \text{payload})$$ where “from” and “to” span all agents, UI, and persistent stores; “type” includes {REGISTER, INVOKE, UPDATE, RESPONSE, COMPLETE}; and the payload is strictly JSON, context-specific (Shehab, 25 Sep 2025).

1.3 Privacy & Compliance Layer:

Comprehensive, field-level AES-GCM encryption is applied to all sensitive health information in the backing database (e.g., MongoDB) (Shehab, 25 Sep 2025). Role-Based Access Control (RBAC) with least-privilege policies, tamper-evident audit logging (hash-chained logs), and consent management provide compliance with major health data protection laws (HIPAA, PIPEDA, PHIPA).

2. Agent Workflow and Protocols

2.1 Task Decomposition and Dynamic Scheduling:

MHCA decomposes the health checkup process into granular subtasks spanning triage, structured history, vitals integration, lab and imaging ingestion, guideline-based recommendations, and follow-up (Yu et al., 3 Nov 2025). These subtasks are monitored and scheduled dynamically:

• The Monitor continuously evaluates subtask completeness with a score $S(t_{ij}; \text{Context}^k) \in [0,1]$, where thresholding dictates pending actions.
• The Controller selects and schedules the next subtask using an explicit utility function:

$$U(t_{ij}; \text{Context}^k) = \alpha \cdot \text{Priority}(t_{ij}) + \beta \cdot [1 - S(t_{ij}; \text{Context}^k)] - \gamma \cdot C_t(t_{ij})$$

where $C_t$ is a time-budget cost (Yu et al., 3 Nov 2025).

2.2 Model Context Protocol (MCP) Pseudocode:

Formalized master agent loop:

on INVOKE(tid, payload): C[tid] := init_context(payload) next_task := "symptom_check" dispatch_to_agent(tid, next_task) on UPDATE(tid, from_agent, result): C[tid] := merge(C[tid], result) if from_agent == "SymptomChecker": if C[tid].symptoms.severity != "high": dispatch_to_agent(tid, "MedicationAgent") else: escalate_to_clinician(tid) elif from_agent == "MedicationAgent": dispatch_to_agent(tid, "AppointmentAgent") elif from_agent == "AppointmentAgent": send_COMPLETE_to_UI(tid, C[tid])

(Shehab, 25 Sep 2025)

2.3 Context Management and Message Schema:

All agent states and outputs are coordinated through a global context object containing structured history, vitals, labs, and system logs. Inter-agent messages encapsulate the sender, receiver, action type, timestamp, and serialized payload, ensuring atomicity and auditability (Yu et al., 3 Nov 2025).

3. Privacy, Security, and Compliance

3.1 Role-Based Access Control (RBAC):

RBAC is defined as a policy function $\varphi$, mapping roles $R$ and operations $O$ to binary decisions, with user roles established by mapping $\mu: U \rightarrow 2^R$ (Shehab, 25 Sep 2025). Every access or action request must satisfy:

$$\exists r \in \mu(u) \text{ such that } \varphi(r, o) = 1$$

3.2 Encryption and Auditability:

All personally identifiable health information (PHI) fields are encrypted as: $C_f = \text{AES\_GCM\_Enc}_K(\text{IV}, P_f, A_f)$ with decryption requiring correct nonce, ciphertext, associated data, and authentication tag (Shehab, 25 Sep 2025).

Audit logs are hash-chained:

$$H_0 = H(\text{seed}), \quad H_i = \text{SHA256}(H_{i-1} \| \text{serialize}(e_i))$$

where a break in chain integrity signals tampering. Every MCP message that processes PHI is logged; unauthorized actions are blocked and logged for post hoc review (Shehab, 25 Sep 2025).

3.3 Compliance with Standards:

The platform aligns with HIPAA, PIPEDA, and PHIPA: at-rest and in-transit encryption, granular access controls enforcing minimum-necessary use, explicit consent recording, and transparent, tamper-evident auditability (Shehab, 25 Sep 2025).

4. Multilingual and Multimodal Interactions

4.1 Language Support:

Language detection is performed at UI ingestion (fastText). Context objects are tagged per session (e.g., $$C[\text{tid}].\text{lang} \in \{\text{en, fr, ar}\}$$), and agents leverage language-specific prompt templates while preserving invariant JSON schemas—there is no inline translation of protected data (Shehab, 25 Sep 2025). Clinicians can switch UI languages without reprocessing PHI; data fields remain encrypted.

4.2 Multimodal Data:

MHCA extends input processing to structured (labs, biosignals), unstructured (text, voice), and image data, with each modality preprocessed (e.g., ASR for voice, CNN feature extractors for images) and unified in a “percept vector” that feeds agent models or LLM pipelines (Abbasian et al., 2023).

5. Clinical Logic, Workflow, and Evaluation

5.1 Use-Case: Health Checkup Session Flow:

A typical orchestrated workflow involves sequential agent invocations for symptom checking, low-risk stratification, medication suggestions, appointment scheduling, and context-sensitive escalation. The privacy layer is invoked at each step: authentication, RBAC enforcement, encryption upon arrival/storage of PHI, and audit logging. Final recommendations are encrypted, stored, and displayed in the user’s preferred language (Shehab, 25 Sep 2025).

5.2 Comprehensive Task Taxonomy:

MHCA extends the agent-driven scheduling to a multilayered taxonomy of health tasks, including review of systems, family and lifestyle history, vitals, laboratory and imaging ingestion, and generation of personalized recommendations. Each domain is divided into atomic subtasks, scored for completion and clinical priority, and scheduled to optimize information gain and time budget (Yu et al., 3 Nov 2025).

5.3 Evaluation Metrics:

Empirical benchmarks on MHCA variants include triage accuracy (87.0% primary; 80.5% secondary), subtask completion rates (98.2% for agent-driven scheduling vs. 93.1% sequential), and expert-rated clinical quality scores (e.g., Chief Complaint, HPI, PH all $>4.4$ on a $5$-point scale) (Yu et al., 3 Nov 2025). Latency, throughput, energy efficiency, and user satisfaction are assessed in extension studies (Gawade et al., 7 Mar 2025).

5.4 Comparison Table: Sub-Agent Effectiveness

Agent	Key Task	Specialized MHCA	Single LLM	Parallel Agents
DS	Plan Generation	75.6%	53.7%	N/A
DS	Code Pass Rate	79.0%	58.4%	N/A
DE	MCQ Accuracy	83.6%	81.8%	N/A
DE	DDx Top-5	75.6%	71.8%	N/A
HC	Active Listening (win rate)	71.0%	29.0%	N/A
PHA	End-User Preference	48.7% (#1)	26.5/24.8%	–
PHA	Expert Preference	80.0% (#1)	4.4/15.2%	–

(Heydari et al., 27 Aug 2025)

6. Personalization, Explainability, and Prevention

6.1 Disease Risk Stratification & HDPD

MHCA incorporates individualized, machine learning-based risk prediction using Health–Disease Phase Diagrams (HDPD). For each patient vector $x^0$, XGBoost models $f: \mathbb{R}^p \to [0,1]$ generate probabilities for future NCD onset. Intervention boundaries are visualized using the projected-multivariate ICE (p-mICE) approach, enabling per-patient 2D “phase diagrams” that recommend minimal biomarker adjustments to shift the health state from “onset” to “non-onset” (Nakamura et al., 2022):

$$\Omega_{onset} = \{(u, v)\ |\ f(x(u,v)) \ge \tau\},\quad B_{ij} = \{(u,v)\ |\ f(x(u,v)) = \tau\}$$

This enables actionable, individualized prevention targets—for example, to lower predicted diabetes risk, an intervention may aim to decrease HbA1c and fasting glucose to subject-specific cutpoints.

6.2 Explainability and Human-in-the-Loop:

Chain-of-thought prompting, causal tracing, and guideline-backed agent reasoning deliver interpretable outputs suitable for clinician oversight (Almansoori et al., 28 Mar 2025). All recommendations, orders, and escalations are auditable and, optionally, subject to human approval before execution.

7. Deployment Models, Extensibility, and Limitations

MHCA can be instantiated on-device (e.g., via Qwen-7B models quantized to 2GB with LoRA adapters for planner/caller roles) or in secure clinical-cloud environments with full model-agnostic LLM and compliance support (Gawade et al., 7 Mar 2025, Yu et al., 3 Nov 2025). The architecture is extensible—task, measurement, and guideline domains can be added modularly. Smart orchestration, edge optimization, and RBAC enforcement remain invariant under extension.

MHCA is presented as a research prototype and not a certified medical device (Shehab, 25 Sep 2025). Clinical deployment for diagnostic or therapeutic purposes requires additional validation, oversight, and regulatory certification.

• Agentic-AI Healthcare: Multilingual, Privacy-First Framework with MCP Agents (Shehab, 25 Sep 2025)
• Self-Evolving Multi-Agent Simulations for Realistic Clinical Interactions (Almansoori et al., 28 Mar 2025)
• Multi Agent based Medical Assistant for Edge Devices (Gawade et al., 7 Mar 2025)
• From Passive to Proactive: A Multi-Agent System with Dynamic Task Orchestration for Intelligent Medical Pre-Consultation (Yu et al., 3 Nov 2025)
• Individual health-disease phase diagrams for disease prevention based on machine learning (Nakamura et al., 2022)
• Conversational Health Agents: A Personalized LLM-Powered Agent Framework (Abbasian et al., 2023)
• The Anatomy of a Personal Health Agent (Heydari et al., 27 Aug 2025)