Regulated Medical Triage: A Neuro-Symbolic Architecture

The "Planning Rubicon" in Healthcare

A top-tier telehealth provider attempted to automate clinical intake using a standard generative AI "chatbot" (a ReAct loop). Because the LLM was trapped in a continuous loop of reasoning and tool-calling with no bounded action space, it failed catastrophically.

Unbounded Action Space

Tool calls were entirely LLM-decided. Without a deterministic graph, the agent could theoretically invoke any exposed API at any time, eliminating auditability.

Blast-Radius Failure

If the hospital's EHR API returned a timeout, the ReAct agent panicked, hallucinated fake patient history to fulfill its prompt, and attempted to route care based on fiction.

No Formal Verification

There was no physical barrier preventing the LLM from triggering a prescription API. It was only stopped by a "system prompt"—which is mathematically vulnerable to jailbreaks.

Clinical Domain Constraints

Real medical triage is not an abstract conversational flow. It requires mapping unstructured symptoms to standard clinical acuity frameworks (like ESI or CTAS).

The system must explicitly map patient presentations into rigid operational vectors:

Acuity Level: 1 (Resuscitation) to 5 (Non-urgent)
Risk Vector: Cardiac, Neurological, Trauma

Care Path: ER, Urgent Virtual, Async Self-Care
SLA Timer: e.g., 10 minutes to specialist

The Neuro-Symbolic Control Loop

Neural proposes. Symbolic disposes.

We separated probabilistic reasoning from deterministic execution. The LLM acts purely as a stochastic proposal engine; the surrounding infrastructure is strictly deterministic.

Ingress & Sanitization

Goal: Zero PHI exposure. Data enters a secure API gateway where a local edge-filtering service tokenizes Protected Health Information. The LLM never sees raw identity—only a clinical abstraction (e.g., Age Bucket: 40-50, Sex: M, Comorbidity: Hypertension).

The Constrained Brain (Planner LLM)

The LLM does NOT call tools. It does NOT write text to the patient. It ingests the clinical abstraction and outputs a strictly validated JSON execution plan. If the plan violates schema validation, maximum depth limits, or enum enforcement, it is deterministically rejected before execution begins.

Policy-as-Code Firewall (OPA)

Before any action is committed (scheduling, routing, EHR mutation), an Open Policy Agent (OPA) intercepts the request. Symbolic, hard-coded clinical rules (e.g., Chest pain + diaphoresis > 20 min = Automatic ER Escalation) override the neural network's routing priority.

Deterministic Execution Engine

The orchestration engine interprets the approved JSON plan step-by-step. It maintains durable state, writes to an immutable audit log, and tracks SLA timers. There is no recursion, no autonomous replanning, and no self-directed loops.

Failure Mode Engineering & Auditability

In healthcare, how an AI fails is more important than how it succeeds.

Deterministic Fallbacks

If EHR API fails: Mark status = EHR_UNAVAILABLE. Trigger fallback: Ask patient direct allergy questions; escalate to human nurse.
If LLM plan malformed: Reject. Log anomaly. Retry once. Escalate to human review.
If OPA blocks action: Log violation attempt. Alert compliance dashboard. Return safe fallback path.

Immutable Audit Trail

Every triage session generates a cryptographic footprint for SOC2/HIPAA compliance, stored in an append-only log:

Generated Plan Hash
Policy Decision Trace (OPA logic)
Timestamped State Transitions

Real Case Walkthrough

Scenario: 45-year-old male. Chest pain for 45 minutes. History: hypertension. On aspirin.

1_sanitized_clinical_abstraction.json

// Layer 1: PII stripped. Only structured clinical data reaches the LLM.
{
  "patient_id": "TOKEN_8832",
  "demographics": { "age_bucket": "40-50", "sex": "M" },
  "presentation": {
    "symptoms": ["chest_pain_acute"],
    "duration_mins": 45
  },
  "known_comorbidities": ["hypertension"],
  "current_medications": ["aspirin_81mg"]
}

2_execution_plan_proposal.json

// Layer 2: LLM proposes a structured plan. No actions are taken yet.
{
  "acuity_classification": {
    "esi_level": 2,
    "risk_vector": "cardiac"
  },
  "proposed_graph": [
    {
      "step": 1,
      "action": "ROUTE_DECISION",
      "parameters": { "care_path": "ER_IMMEDIATE", "sla_timer_mins": 10 }
    }
  ]
}

3_opa_firewall.rego

# Layer 4: Symbolic rules evaluate the neural proposal.
package triage.governance

default allow_routing = false

# RED FLAG DETECTED: Deterministic ER escalation
allow_routing {
    input.acuity_classification.esi_level <= 2
    input.proposed_graph[_].parameters.care_path == "ER_IMMEDIATE"
}

# FAILSafe: Block any non-ER routing for high-acuity cardiac vectors
deny_routing[msg] {
    input.acuity_classification.risk_vector == "cardiac"
    input.proposed_graph[_].parameters.care_path != "ER_IMMEDIATE"
    msg := "FATAL: Policy mandates immediate ER routing for cardiac vectors."
}

Regulated Medical Triage: A Neuro-Symbolic Approach