The Same Scissors Have 20 Different Names

Three Hospitals, Three Names

Consider these three entries from three different hospital sterile processing departments:

Hospital	Instrument Name in System
Hospital A	Metzenbaum curved 18cm
Hospital B	Metz cvd 7in
Hospital C	Scissors, Metz, curved, TC

Same instrument. Three hospitals. Three completely different names. Now multiply this across 5,000 instruments per hospital and 200 surgical trays, and you begin to understand the scale of a problem that quietly costs healthcare systems millions in validation errors, surgical delays, and compliance risk.

This isn't sloppy data entry. It's a systems problem: one that emerges naturally from decades of decentralised healthcare operations, manufacturer variations, regional naming conventions, and the practical reality that sterile processing technicians need shorthand that works for their workflow, not a standardised taxonomy designed in a committee room.

The Hidden Cost of Naming Chaos

The surgical instrument tracking system market is projected to reach $751 million by 2030, growing at 15.2% annually.¹ But this investment in tracking technology often fails to deliver its promised value because of a fundamental upstream problem: if you can't reliably identify what an instrument is, you can't reliably track where it goes.

The numbers: nearly 87% of all surgical tray errors result from failures in instrument visualisation: inspection, identification, and function verification.² The annual cost of surgical instrument-associated delays at just one academic healthcare campus was estimated between $6.7 million and $9.4 million in lost chargeable operating room minutes.

In sterile processing departments, approximately 1 in 10 instrument trays arrives at the operating room with missing instruments, and another 1 in 20 contains broken instruments. Operating room delays average 17 minutes per incident, with OR time costing $100 or more per minute. Case cancellations can cost $1,500 or more per hour of planned surgery.

The Scale of the Problem

87% of tray errors stem from identification failures
$6.7M–$9.4M annual cost at a single hospital campus
1 in 10 trays arrive with missing instruments
17 minutes average delay per incident
$100+/minute OR time cost

Why Standardisation Has Failed

Healthcare interoperability has been a regulatory priority for decades. The 21st Century Cures Act mandated data sharing standards. FHIR adoption has grown from 66% to 71% of countries using it for at least a few use cases by 2025. Yet only 4 in 10 health information exchanges routinely send data aligned with USCDI standards, and only about one-third routinely receive such data.

The problem isn't lack of standards; it's that standards describe ideal states, not messy realities. Surveyed physicians most often name interoperability as their primary obstacle in getting value from health information technology. The difference between being technically 'FHIR-compliant' and truly interoperable remains substantial.

Surgical instrument naming exemplifies this gap. The FDA's Unique Device Identification (UDI) system provides a framework for device traceability, and reusable surgical instruments are explicitly covered: they must carry both human-readable and machine-readable UDI codes.³ But UDI tracks the device, not what hospitals call the device in their local systems.

Each hospital develops its own naming conventions over years, sometimes decades. Sterile processing technicians create abbreviations that work for rapid identification during tray assembly. Surgeons request custom trays with preference cards that use their own terminology. Manufacturers update product names across versions. The result is a Tower of Babel where the same physical instrument exists under dozens of different identities across the healthcare ecosystem.

What Doesn't Work

Rigid Keyword Matching

Traditional database matching breaks on the first abbreviation. 'Metzenbaum' becomes 'Metz' becomes 'Metzb' becomes a zero-match lookup. Keyword systems require exact matches or carefully curated synonym tables: tables that must be maintained manually and inevitably fall behind actual usage patterns.

"Just Use AI"

The temptation to throw a large language model at the problem is understandable. LLMs excel at understanding semantic relationships and can handle abbreviations, synonyms, and contextual variation. But healthcare demands accountability that pure AI approaches cannot provide.

Retrieval augmentation helps but does not rescue them. Grounding an LLM in retrieved documents significantly reduces hallucination rates compared to the base model, yet retrieval noise, domain shift, and limited explainability persist. When an LLM confidently identifies an instrument incorrectly, the consequences can propagate through surgical workflows, compliance audits, and patient safety incidents.

Perhaps more critically, pure AI approaches fail audits. Healthcare regulators require explainable decision trails. When asked 'why did the system identify this instrument as a Metzenbaum scissors?', 'the neural network said so' is not an acceptable answer.

Manual Mapping

The fallback approach (having staff manually map every instrument name across systems) doesn't scale. With 10,000 to 100,000 instruments tracked per hospital and constant churn from new equipment, manufacturer updates, and evolving terminology, manual mapping burns out staff and creates maintenance backlogs that undermine system reliability.

What Actually Works: A Layered Architecture

The solution isn't to choose between keyword matching, AI, and manual curation. It's to build a layered system that combines the strengths of each approach while compensating for their individual weaknesses. This is the architecture I've developed and refined through production deployment.

Layer 1: Canonical Instrument Model

The foundation is representing what an instrument is, not what it's called. A canonical model captures the essential attributes: instrument class (scissors, forceps, retractor), subtype (Metzenbaum, Mayo, Kelly), configuration (curved, straight), size (measurements in both metric and imperial), material composition (stainless, tungsten carbide tips), and functional characteristics (serrated, smooth, ratcheting).

This canonical representation becomes the 'source of truth' against which all local naming variations are mapped. It's not about forcing hospitals to change their terminology; it's about creating a translation layer that preserves local workflows while enabling system interoperability.

Layer 2: Ruthless Normalisation

Before any AI touches the data, aggressive preprocessing normalises input text: abbreviations are expanded ('cvd' → 'curved', 'TC' → 'tungsten carbide', 'Metz' → 'Metzenbaum'), measurements are parsed and converted to canonical units, noise words are stripped, and manufacturer-specific prefixes are identified and tagged.

This normalisation layer dramatically improves downstream matching accuracy by reducing the variation that semantic models must handle. Tokenisation, unit parsing, and abbreviation expansion each yield measurable gains in retrieval precision: the work is bridging the informal-to-formal language gap before any model sees the text.

Layer 3: Semantic Retrieval with Domain-Specific Embeddings

After normalisation, semantic search handles the naming chaos that humans create. Modern embedding models can understand that 'Metz scissors curved' and 'Metzenbaum dissecting scissors, curved' refer to the same instrument class, even without explicit synonym tables.

Domain-specific embeddings make the difference here. Cross-lingual medical procedure mapping with LLM-augmented semantic embeddings has reached 90% accuracy at top-5 retrieval across English, German, French, and Italian clinical terminologies.⁴ The margin over general-purpose models looks small on paper (a few points of F1), but at the scale of a hospital's instrument catalogue it translates to hundreds fewer misclassifications in downstream workflows.

Layer 4: Hybrid Scoring

The critical innovation is combining lexical precision with semantic understanding through hybrid scoring. BM25-based keyword matching excels at exact term matches, essential for catching specific model numbers, catalogue codes, and precise terminology. Semantic embeddings capture conceptual relationships that keyword matching misses.

Neither approach alone is sufficient: published evaluations of hybrid search architectures report gains of more than 20 percentage points over single-method baselines.⁵ The key is score normalisation: converting the different scales of BM25 scores and embedding distances into comparable confidence measures that can be combined through weighted fusion.

This hybrid approach produces stable, explainable confidence scores: 'Matched with 94% confidence based on 85% semantic similarity and 98% lexical match on normalised instrument name.' Auditors can trace exactly how the system reached its conclusion.

Layer 5: Constrained LLMs for Edge Cases

Large language models enter the pipeline only for adjudicating edge cases where hybrid scoring produces ambiguous results. The LLM is never the final authority; it provides reasoning that a human can verify, with outputs constrained to structured decisions rather than free-form text generation.

This constrained approach addresses the hallucination concerns that make pure LLM solutions unsuitable for healthcare. Constrained properly, retrieval-augmented models have outperformed human evaluators on structured medical assessment tasks while producing no hallucinated outputs.⁶ Unconstrained, the same models offer neither guarantee.

Layer 6: Human-in-the-Loop with Full Audit Trails

Healthcare demands accountability. Every mapping decision, whether automated or human-verified, is logged with complete provenance: input data, normalisation steps, matching scores, confidence levels, and approval chain. This audit trail satisfies regulatory requirements while enabling continuous system improvement through feedback loops.

Architecture Overview

Layer	Function	Key Benefit
Canonical Model	Define what instruments ARE	Single source of truth
Normalisation	Expand abbreviations, parse units	Reduces AI input variation
Semantic Search	Domain embeddings retrieval	Handles naming chaos
Hybrid Scoring	BM25 + embeddings fusion	Stable, explainable confidence
Constrained LLM	Edge case adjudication	Handles ambiguity safely
Human Review	Verification with audit trail	Regulatory compliance

Beyond Matching: Modelling Equivalence for Tray Validation

For tray validation, the goal isn't equality; it's equivalence. A surgical tray doesn't fail validation because it contains a different brand of the same instrument. It fails when clinically significant requirements aren't met.

This requires modelling several additional dimensions beyond simple name matching:

Substitution Rules: Which instruments can functionally replace others? A 7" Metzenbaum may substitute for a 7" Mayo scissors in some contexts but not others.

Critical vs. Non-Critical Items: Missing a backup retractor is different from missing the primary scalpel. Validation rules must weight items by surgical impact.

Surgeon and Site Preferences: Preference cards encode acceptable variation per surgeon. Dr. Smith always wants tungsten carbide tips; Dr. Jones accepts either. Site-specific policies add another layer.

Quantity Tolerance: Some trays require exact counts; others allow variance within ranges.

The hybrid scoring system extends naturally to handle these requirements. Equivalence scoring combines instrument identity matching with rule-based validation logic that encodes clinical knowledge, producing actionable results: 'Tray validated with 98% confidence. Note: Substitute Aesculap Metzenbaum (catalogue #BC456) used in place of Sklar Metzenbaum (catalogue #BC123). Functional equivalence confirmed.'

State of the Art: What Research Shows

The architecture I've described tracks the research frontier in healthcare AI. Four findings from the 2024–2025 literature underpin it:

Hybrid Retrieval Outperforms Single Methods: Across information retrieval evaluations, combining sparse (BM25) and dense (embedding) retrieval beats either method alone.

Domain-Specific Embeddings Matter: Embeddings trained on EHR data and medical literature capture the semantics of medical terms better than those trained on general corpora. General-purpose models underperform on exactly the vocabulary that matters here.

RAG Reduces But Doesn't Eliminate Hallucinations: Grounding generation in retrieved knowledge improves factual consistency, but retrieval noise, domain shift, and limited explainability persist. Retrieval is a mitigation, not a cure.

Preprocessing Yields Measurable Gains: Text normalisation, tokenisation, and abbreviation expansion provide significant accuracy improvements before any model is involved.

Regulatory Context and Compliance

The layered architecture addresses key regulatory requirements that pure AI solutions cannot satisfy:

FDA UDI Compliance: The canonical model layer can incorporate UDI data as authoritative identifiers, with the matching system bridging between UDI codes and local naming conventions.

Joint Commission Standards: The audit trail component satisfies requirements by documenting every identification decision with complete provenance.

AAMI ST79 Compliance: Instrument tracking systems must support compliance with sterile processing standards, which the hybrid architecture enables through reliable identification.

Key Takeaways for Healthcare Technology Leaders

Naming inconsistency is a systems problem, not a data quality problem. Solutions must work with messy reality, not assume it away.
Pure AI approaches fail healthcare's accountability requirements. Explainability and audit trails are non-negotiable.
Hybrid scoring combining lexical and semantic matching outperforms either alone. Research consistently validates this approach.
Preprocessing and normalisation provide significant gains. Clean the data before letting AI touch it.
For tray validation, model equivalence, not equality. Substitution rules, critical items, and surgeon preferences matter.

Where Healthcare AI Is Actually Heading

The surgical instrument naming problem is a microcosm of healthcare's broader AI challenge. The industry is filled with messy, inconsistent, locally-optimised data that resists standardisation. The solutions that work aren't the ones that promise magic; they're the ones that acknowledge complexity and engineer around it.

Less magic. More engineering. Deep respect for domain complexity.

The layered architecture I've described delivers a system that hospitals can trust, regulators can audit, and that actually scales. It represents the direction applied AI in healthcare is heading: hybrid approaches that combine the strengths of traditional software engineering with modern machine learning, always with human oversight and accountability.

Anyone who still thinks a single model solves everything has not met this problem.

Footnotes

Analyst forecast for the surgical instrument tracking market; market-size estimates vary by firm and will date. ↩
BMC Surgery, 2024. The delay-cost estimate that follows is from a study of a single US academic medical campus. ↩
Under 21 CFR Part 801.20, which mandates direct marking of reusable surgical instruments. ↩
A 2025 study of cross-lingual medical procedure mapping using LLM-augmented semantic embeddings, evaluated across English, German, French, and Italian terminologies. ↩
One 2025 evaluation of hybrid search architectures reported a 23.3 percentage point gain over its baseline; the exact margin varies with corpus and query mix. ↩
A 2025 study of retrieval-augmented medical fitness assessment: a properly constrained GPT-4 reached 96.4% accuracy with no hallucinated outputs, against 86.6% for human evaluators. ↩