1. Framing & Positioning

1.1 Title & Abstract

Structured Symptom Retrieval Intelligence (SSIR): Deterministic, High-Confidence Medical Triage via Retrieval-First Architecture

Abstract
Structured Symptom Retrieval Intelligence (SSIR) is a deterministic retrieval framework for medical triage that replaces the generative unpredictability of large language models (LLMs) with pre-engineered, clinically validated protocols and similarity search over precomputed embeddings. This design ensures every decision is auditable, reproducible, and cost-predictable—critical requirements for regulated healthcare environments.

SSIR is the healthcare-specific implementation of the broader Structured Intelligence Retrieval (SIR) paradigm, a domain-agnostic framework for retrieval-first, generation-last reasoning in high-stakes sectors. At its core, SSIR embeds a curated library of clinician-reviewed triage protocols into a vector space, processes patient narratives into canonical symptom terms via an NLP preprocessing layer, and iteratively refines the patient context until a confidence threshold (

$\tau_{high}$

) is reached.

The architecture includes multi-path orchestration to handle divergent symptom clusters, a controlled single-shot LLM fallback for resolving stalled reasoning loops, and strict guardrails to ensure that no diagnosis is generated at runtime. Retrieval is strictly non-generative—LLMs are used only for constrained clarifier generation from a pre-approved vocabulary.

SSIR's theoretical foundation, formalized mathematically in this paper, demonstrates convergence properties and introduces the Stochastic Structuring Effect (SSE)—the emergent tendency of unstructured patient input to resolve into structured, canonical symptom sets through iterative interaction. The framework delivers superior retrieval accuracy compared to traditional Retrieval-Augmented Generation (RAG) pipelines while operating at a fraction of the cost and with zero risk of hallucinated medical advice.

---

1.2 Executive Summary

The medical triage technology stack is at an inflection point. Generative LLM-first approaches—while impressive in unconstrained domains—struggle to meet the reproducibility, safety, and cost thresholds demanded by regulated clinical practice.

The problem is threefold:

1. Safety & Compliance Risk

   • Regulatory regimes such as the FDA's Software as a Medical Device (SaMD) guidance and the EU's MDR require deterministic outputs and traceable reasoning chains.
   • Generative outputs, by their nature, are non-deterministic—even identical prompts can yield different, unverifiable conclusions.
   • In clinical triage, such unpredictability is unacceptable: misclassification can lead directly to harm.

2. Operational Cost & Scalability

   • Continuous invocation of large generative models incurs high per-session costs and latency, making large-scale deployment economically unsustainable.
   • Token throughput for multi-turn conversations adds hidden cost multipliers.

3. Accuracy & Multi-Symptom Complexity

   • Current RAG pipelines still depend on generative synthesis, allowing error creep into final outputs.
   • Multi-symptom or divergent complaint handling (e.g., "headache and foot pain") often causes reasoning contamination—symptoms bleed into unrelated decision paths.

SSIR solves these constraints by design.
It is a retrieval-first, generation-last system where the terminal output is always a clinician-validated protocol, never a model-invented one. The user's query is the only element embedded at runtime, compared against a fixed, embedded library of protocols. An iterative context refinement loop progressively improves retrieval confidence until it crosses the acceptance threshold, ensuring both determinism and convergence.

The framework introduces:

• Multi-Path Orchestration: Splits unrelated symptom clusters into separate reasoning tracks, preventing contamination.
• Controlled LLM Fallback: A single-turn, JSON-only clarifier/symptom suggestion step from a pre-approved vocabulary, triggered only when similarity search stalls.
• NLP Preprocessing Layer: Normalizes messy, colloquial patient inputs into canonical medical ontology terms before retrieval.
• Offline Protocol Generation: All triage logic is encoded, validated, and embedded before runtime—no real-time protocol invention.

Why this matters now:
Healthcare AI is moving rapidly toward regulation-grade AI pipelines. SSIR is engineered to be certifiable under MDR/FDA frameworks, economically viable for nationwide deployment, and adaptable across domains that share similar constraints (finance, legal, logistics). By separating reasoning (retrieval) from articulation (optional LLM), it achieves safety without sacrificing user experience.

---

1.3 Industry Context

1.3.1 The Limitations of Current State-of-the-Art (RAG & Pure LLM Triage)

Most medical AI systems today fall into one of two categories:

1. End-to-End Generative LLMs

   • Flexible, conversational, but prone to hallucination and drift.
   • Hard to certify due to non-determinism.
   • Expensive to run continuously at clinical scale.

2. Retrieval-Augmented Generation (RAG)

   • Adds document retrieval before LLM synthesis.
   • Improves grounding, but the final synthesis step can still introduce unverified reasoning.
   • Vulnerable to subtle prompt-level errors and inconsistent phrasing.

Both approaches share weaknesses:

• Non-Deterministic Outputs: A reproducibility audit will often fail.
• High Latency: Multiple generative passes per session.
• Weak Multi-Symptom Handling: No native separation of unrelated complaint paths.
• Compliance Risk: Regulatory bodies are wary of unverifiable model reasoning.

1.3.2 The Retrieval-First Alternative

Structured Intelligence Retrieval (SIR) flips the paradigm:

• No generative synthesis at the terminal step.
• Retrieval terminates in a prevalidated, domain-specific object.
• LLMs, if used, are strictly constrained and sandboxed.

SSIR, as SIR's medical instantiation, brings these advantages to triage:

• Determinism → Same input, same output, auditable every time.
• Cost Predictability → Runtime embedding + vector search is orders of magnitude cheaper than continuous generation.
• Safety → All outputs are pre-engineered and clinically validated.
• Traceability → Every decision path is reconstructable from context state and protocol embeddings.

In regulated healthcare and similarly constrained domains, retrieval-first is not just an optimization—it's a necessity. SSIR represents a production-grade embodiment of that principle.