From classification to clinical understanding

Key points

• A list, a hierarchy, a controlled vocabulary, and an ontology are four different tools, not four names for the same thing. Carried through one synthetic example, each makes something visible that the previous one could not, and each still hides something.
• A hierarchy always has an organizing criterion, whether or not anyone states it. Two correct hierarchies of the same items, built on different criteria, disagree about what belongs together, and neither can show a relationship that cuts across its branches.
• The history of classification is a history of such choices: Wilkins fused the name with the classification, Linnaeus separated them, the International Classification of Diseases built bins for counting, and the SNOMED lineage built multi-axial descriptions and then formal definitions.
• A classification sorts each thing into one place for a purpose. An ontology states what things are and how they relate, so that many purposes can be served from one account. Using one for the work of the other is a common and expensive mistake.
• This publication is built as a reference network: essays carry an argument, monographs supply depth, and lexicon entries give a compact definition and relationships.

Six things on a shelf

Everything in this section is a synthetic example. The agents, targets, and procedures are invented for teaching and do not refer to any marketed product, real molecular target, or terminology code.

Imagine a small nuclear medicine department that uses exactly three radiopharmaceuticals and performs exactly three procedures with them.

• Agent A: a diagnostic PET agent labeled with fluorine-18 that binds a cell-surface protein we will call Target P.
• Agent B: a therapeutic agent labeled with lutetium-177 that binds the same Target P.
• Agent C: a diagnostic agent labeled with technetium-99m that localizes to bone through a mechanism unrelated to Target P.
• Procedure 1: PET/CT imaging after administration of Agent A.
• Procedure 2: therapeutic administration of Agent B.
• Procedure 3: SPECT/CT imaging after Procedure 2, using the gamma emissions of Agent B to see where it went.

Six items. The question of this essay is how to arrange them so that a physician, a radiochemist, a physicist, a scheduler, and a database can all understand them, and what each arrangement costs.

A list

The simplest arrangement is the one above: a list. A list is more useful than it looks. It fixes membership (these six, and no others), it gives each item a name, and it can be numbered so that the names can be referred to. Every department has such lists, usually in a spreadsheet, and they do real work.

What a list cannot do is say anything about how its members are related. The reader may notice that Agents A and B bind the same target, or that Procedure 3 depends on Procedure 2, but the list does not notice. If "Agent A" is also called by its chemical name in the radiopharmacy and by a nickname in the clinic, the list has no way to record that these are one thing. A set of names is where every terminology begins and where none should end.

A hierarchy

The next step is to group the items. Here the central question of this essay appears for the first time: group them by what?

One natural answer, for a radiopharmacy, is by radionuclide: fluorine-18 agents (A), lutetium-177 agents (B), technetium-99m agents (C), with procedures under the agent they use. This hierarchy is correct and useful for ordering, storage, shielding, and waste. It makes visible that Agent B and Procedure 3 share a nuclide. It hides completely that Agents A and B share a target.

Another natural answer, for a clinician, is by intended use: diagnostic (A, C, Procedures 1 and 3) and therapeutic (B, Procedure 2). This is also correct and useful, for scheduling, consent, and billing. It makes visible that Procedure 3 is an imaging procedure. It hides that Procedure 3 is imaging the therapeutic agent, and it separates Agent B from Agent A even though the two are, for the purpose of theranostics, a pair.

A third answer, for a biologist, is by target: Target P agents (A, B) and bone-seeking agents (C). This one finally shows the pair. It hides the nuclide and the use.

None of these is wrong. Each is a Taxonomy: a classification organized by a stated criterion, usually as a hierarchy. The criterion is not optional. A hierarchy that does not state its criterion is still built on one, and a reader who assumes a different criterion will misread it. The monograph on Taxonomy develops this point and shows why a taxonomic parent is not a causal explanation: putting Agent B under "lutetium-177 agents" says what it is labeled with, not why it works.

The deeper limitation is structural. A tree can express only one way of belonging. Anything that cuts across the branches, such as "binds the same target as," cannot be drawn in the tree at all. That is what a hierarchy conceals: not facts it got wrong, but relationships it has no place for.

A controlled vocabulary

Suppose the department's three disciplines each keep their own hierarchy and go on calling Agent A by three different names. The next tool addresses the names rather than the grouping.

A Controlled vocabulary is a managed set of terms used consistently for a defined purpose. Managing a vocabulary means, at minimum, four things. Each concept gets a preferred label, so that there is one name to use in the record. Alternative names are recorded as synonyms pointing at the same concept, so that a search for the radiopharmacy's chemical name finds the clinic's nickname. Each concept gets a scope note saying what it does and does not cover: "Agent B refers to the formulated product as administered; the unlabeled ligand and the bare radionuclide are separate concepts." And each concept gets an Identifier that does not change when the label does. Cimino's desiderata (1998) are the classic statement of why each of these matters, and why the concept and its names must be separate records.

A Thesaurus is a controlled vocabulary with relationships among its terms added: broader term, narrower term, and a looser "related term." The international standard for retrieval thesauri (ISO 25964-1, 2011) defines those three relations and nothing more. In our example, a thesaurus could record that Agent B has broader term "lutetium-177 agent" and related term "Agent A." That is better than three disconnected hierarchies. But "related term" is deliberately vague: it says the two are associated for retrieval, not how. A reader who needs to know whether Agent A can be used to select patients for Agent B gets no answer from "related." The monograph on Thesaurus explains why an associative link must not be read as equivalence.

An ontology

The last tool keeps everything the vocabulary provided and adds one thing: relations that are typed and directed.

Instead of "Agent A is related to Agent B," an ontology records "Agent A binds Target P" and "Agent B binds Target P," and separately records what kind of thing Target P is. Instead of placing Procedure 3 under "imaging," it records "Procedure 3 images the distribution of Agent B" and "Procedure 3 follows Procedure 2." Instead of forcing Agent B into one branch, it records three facts that three different hierarchies had each shown and the others hidden: labeled with lutetium-177, intended for therapy, binds Target P. From those facts any of the three hierarchies can be generated on demand, by choosing a criterion. The reverse is not possible: a hierarchy does not remember why it was built.

This is what the lexicon entry on Relationship means by recording type and direction rather than an unlabeled connection, and what the monograph on Ontology treats as the difference between a vocabulary and an ontology. It is also where context enters. Agent B is a therapeutic agent in Procedure 2 and an imaging agent in Procedure 3. A hierarchy must choose; an ontology can say that the agent's role depends on the procedure. The lexicon entry on Context gives the short form: the conditions needed to interpret a statement are part of the statement.

Figure 1. The same six items under four arrangements (text table, synthetic example). Organizing criterion for the hierarchy column: radionuclide. Agents, targets, and procedures are invented; the identifiers are local to this page.

Item	List	Hierarchy (by radionuclide)	Controlled vocabulary	Ontology (typed relations)
Agent A	named, position 1	under fluorine-18 agents	preferred label "Agent A"; synonyms recorded; local ID X-001	labeled with fluorine-18; binds Target P; intended for diagnosis
Agent B	named, position 2	under lutetium-177 agents	preferred label "Agent B"; scope note excludes bare ligand and nuclide; local ID X-002	labeled with lutetium-177; binds Target P; intended for therapy; imaged by Procedure 3
Agent C	named, position 3	under technetium-99m agents	preferred label "Agent C"; local ID X-003	labeled with technetium-99m; localizes to bone; intended for diagnosis
Procedure 1	named, position 4	under Agent A	preferred label; local ID X-004	uses Agent A; is a PET/CT imaging procedure
Procedure 2	named, position 5	under Agent B	preferred label; local ID X-005	administers Agent B; is a therapeutic procedure
Procedure 3	named, position 6	under Agent B	preferred label; local ID X-006	images Agent B; follows Procedure 2; is a SPECT/CT imaging procedure
What it shows	membership and names	shared nuclide	one name per concept, stable ID, scope	shared target; role by context; dependency between procedures
What it hides	everything else	shared target; role of Agent B in Procedure 3	how related terms are related	nothing in principle, but every relation must be asserted and maintained

Alt text: a table with one row per synthetic item and one column per arrangement, ending with two rows that summarize what each arrangement shows and hides.

The last cell is the honest one. An ontology hides nothing in principle, and that is exactly its cost. Every relation in it was put there by someone, and will be wrong the day the department changes a protocol unless someone changes the relation too. A list is cheap and dumb. An ontology is expensive and explicit. The history of classification is largely the history of people discovering which they could afford.

How people have classified

The same choices recur across three centuries, and the history clarifies them because the stakes were visible.

In 1668 John Wilkins, a founder of the Royal Society, published An Essay towards a Real Character and a Philosophical Language. He proposed to divide all of reality into forty top-level genera, each subdivided into differences and then species, and to assign each node a written symbol and a spoken syllable, so that a word in his language would be its own classification (Wilkins, 1668; Eco, 1995, traces the project and its predecessors). The project could not work. The world refused to stay in forty boxes, and a change anywhere in the classification would change the words themselves. Wilkins had fused the identifier with the hierarchy, and the fusion made both brittle.

Carl Linnaeus learned the opposite lesson. In the Systema Naturae, first published in 1735, he arranged minerals, plants, and animals into nested classes, orders, genera, and species. His scheme for plants was frankly artificial, grouping them by the number and arrangement of their reproductive parts, and he knew it did not track natural affinity; he chose it because anyone with a flower and a hand lens could apply it (Koerner, 1999). The organizing criterion was explicit, and chosen for use rather than for truth. The lasting innovation came with Species Plantarum in 1753 and the tenth edition of Systema Naturae in 1758: a consistent two-word name for every species, separated from the description that had previously served as the name. A stable short identifier, bound to a position in a hierarchy, with a separate description that may grow, is the architecture of every serious terminology since. Roget's Thesaurus of 1852, a classification of ideas rather than a dictionary, is the ancestor of the retrieval thesaurus described above.

Medicine's classifications grew from a humbler need: to count the dead. In the nineteenth century William Farr, at the General Register Office in England, worked for decades to impose a usable nosology on death returns, sorting diseases by distinctions a registrar could make. Jacques Bertillon's Classification of Causes of Death was adopted by the International Statistical Institute in 1893 and revised roughly every ten years; with the sixth revision in 1948 the World Health Organization took responsibility, and the list became the International Classification of Diseases (Moriyama, Loy, and Robb-Smith, 2011; WHO, 2004). The tenth revision was endorsed in 1990; the eleventh was adopted by the World Health Assembly in 2019 and came into effect in 2022.

The move from nosology to coding is the important one. A nosology is a physician's ordering of diseases by their nature. A coding scheme is an administrator's ordering of cases into mutually exclusive bins so that totals can be computed. The ICD inherited the shape of a nosology but the purpose of a tally sheet, and the tension shows: categories such as "other" and "unspecified" exist because every case must land somewhere, not because nature contains such kinds. Bowker and Star (1999) remain the clearest account of how these invisible choices shape what a health system can see. In the terms of this essay, the ICD is a hierarchy with a stated criterion (one bin per case, for counting) that is excellent at its purpose and cannot, by design, express a relationship between two bins.

The lineage behind SNOMED CT started from the other end. In 1965 the College of American Pathologists published the Systematized Nomenclature of Pathology, developed under Roger Cote, which described a specimen along several independent axes such as topography and morphology rather than in a single tree. It became the Systematized Nomenclature of Medicine, SNOMED, in the 1970s and kept its multi-axial design (Cote and Robboy, 1980). In the late 1990s a logic-based successor, SNOMED RT, replaced fixed axes with formal definitions that software could check (Spackman, Campbell, and Cote, 1997). In 2002 it was merged with the United Kingdom's Clinical Terms Version 3, the descendant of the Read Codes, to create SNOMED CT, and in 2007 ownership passed to the body now called SNOMED International. The result is, in the terms of this essay, a controlled vocabulary with typed relations: each concept has an identifier, descriptions, and relationships to other concepts, including at least one "is a" parent and defining attributes such as a finding site (SNOMED International, logical model). The lexicon entry on SNOMED CT gives the compact description, and the monograph on SNOMED CT and the NucLex niche asks what a nuclear medicine terminology should reuse from it.

The Unified Medical Language System, begun at the National Library of Medicine in 1986, took a third road: it mapped the existing vocabularies rather than replacing them, giving each distinct meaning a concept identifier and linking every source term that expressed it (Lindberg, Humphreys, and McCray, 1993). It is the clearest institutional admission that the plurality of classifications is permanent.

Alan Rector's question from 1999, why clinical terminology is so hard, stands over all of this. Clinical meaning is compositional, context-dependent (a family history of a disease is not the disease), and endlessly productive of new distinctions, so any fixed list is out of date on publication. That is why maintenance is not an afterthought in a terminology but its central activity, and why the "what it hides" row of Figure 1 ends with the word "maintained."

Relations and context

Two ideas carry the weight of the shift from classification to understanding, and both deserve to be stated plainly.

The first is that a relationship has a type and a direction. "Agent A binds Target P" is not the same statement as "Target P binds Agent A," because the first is about what the agent does and the second about what the target admits. "Procedure 3 follows Procedure 2" cannot be reversed at all. A hierarchy has exactly one relation, "is a kind of," pointing upward, and it works because that relation is so constrained. The moment a second relation is needed, the tree is no longer the right picture. An ontology is, in large part, the discipline of naming the relations one needs and refusing to leave any of them as an unlabeled line.

The second is that a statement's meaning depends on conditions that have to travel with it. Agent B is "a therapeutic agent" in Procedure 2 and "the imaged substance" in Procedure 3. "Uptake" means one thing at one hour after administration and another at seven days. A value measured in a phantom is not a value measured in a patient. Classification tends to strip context away, because a bin cannot hold it. Understanding depends on keeping it. In nuclear medicine, where the same molecule changes role with the nuclide attached to it, the time since injection, and the question being asked, this is not a refinement. It is most of the work.

The separation drawn here, between what a hierarchy makes visible and what it conceals, is the writer's interpretation. Whether any particular software system can represent a cross-cutting relation or a context-dependent role is a matter of that system's documentation and tests, not of the argument here.

The reference network

A publication about shared language should practice what it describes, and this one is built accordingly.

NucLex has three kinds of page. An essay, such as this one, carries a narrative and an argument and is meant to be read from beginning to end. A monograph supplies the depth needed at one term or question, in a fixed structure that moves from definition and distinctions through history and theory to biomedical and nuclear medicine relevance, disagreements, and references. A lexicon entry, or concept page, gives a compact definition, a scope note, justified synonyms, and related entries, so that a reader who meets an unfamiliar word can orient in a minute and return to the essay.

Links are written as double brackets around a slug. A plain link such as Taxonomy resolves to the lexicon entry first, then to a monograph, then to an essay, so the default destination is the short definition. A prefixed link such as Taxonomy forces the longer treatment. The essays link the first meaningful use of a concept, preferring the monograph where the reader needs an argument or history and the concept page where a definition will do. Each lexicon entry carries a local editorial identifier of the form NX-C followed by a number. Those identifiers are deliberately local: they do not claim membership in any released terminology and are not mappings to any external code system. The reference network is an editorial structure, implemented in this release as pages and links. It is not a terminology service, and nothing on the site should be read as a live endpoint.

In the terms of this essay, the site is a small controlled vocabulary with typed relations between pages: a Concept for each term, one preferred label, a stable identifier, a scope note, and links whose type is explicit. It is not yet an ontology of nuclear medicine. It is the place where one is being argued for, one definition at a time.

Limitations

This is an AI-assisted draft that has not been source-checked or reviewed by a domain expert or an editor.

The six-item example is wholly synthetic. Agents A, B, and C, Target P, and the three procedures do not correspond to any real product, target, or coded procedure, and the local identifiers X-001 through X-006 are teaching labels.

The historical section compresses a large literature and depends on secondary accounts for the ICD and SNOMED lineages. Dates for institutional events (the 1893 Bertillon adoption, the 1948 WHO assumption of responsibility, the 2002 SNOMED CT merger, the 2007 transfer to SNOMED International, the 2019 and 2022 ICD-11 milestones) are given as commonly reported and must be verified against primary records by the source checker. The characterization of SNOMED CT's structure follows the public logical model and does not describe its authoring rules, editions, or release process.

The claim that an ontology "hides nothing in principle" is a statement about representational capacity, not about any implemented system. This essay separates philosophical interpretation from demonstrated software behavior and asserts the latter nowhere.

The reference network section describes this publication's editorial design as implemented in its first release. It does not describe a terminology server, an API, or a validated mapping to any external system; none exists.

Sources and further reading

• Wilkins J. 1668. An Essay towards a Real Character and a Philosophical Language. London.
• Linnaeus C. 1735. Systema Naturae. 1st ed. Leiden. 10th ed., Stockholm, 1758.
• Linnaeus C. 1753. Species Plantarum. Stockholm.
• Koerner L. 1999. Linnaeus: Nature and Nation. Harvard University Press.
• Roget PM. 1852. Thesaurus of English Words and Phrases. Longman.
• Eco U. 1995. The Search for the Perfect Language. Translated by Fentress J. Blackwell.
• Moriyama IM, Loy RM, Robb-Smith AHT. 2011. History of the Statistical Classification of Diseases and Causes of Death. National Center for Health Statistics.
• World Health Organization. 2004. History of the development of the ICD. In: International Statistical Classification of Diseases and Related Health Problems, 10th revision, Volume 2: Instruction Manual, 2nd ed. WHO.
• Bowker GC, Star SL. 1999. Sorting Things Out: Classification and Its Consequences. MIT Press.
• Lindberg DAB, Humphreys BL, McCray AT. 1993. The Unified Medical Language System. Methods of Information in Medicine 32(4):281-291.
• Cote RA, Robboy S. 1980. Progress in medical information management: Systematized Nomenclature of Medicine (SNOMED). JAMA 243(8):756-762.
• Spackman KA, Campbell KE, Cote RA. 1997. SNOMED RT: a reference terminology for health care. Proceedings of the AMIA Annual Fall Symposium: 640-644.
• Rector AL. 1999. Clinical terminology: why is it so hard? Methods of Information in Medicine 38(4-5):239-252.
• Cimino JJ. 1998. Desiderata for controlled medical vocabularies in the twenty-first century. Methods of Information in Medicine 37(4-5):394-403.
• SNOMED International. SNOMED CT logical model. SNOMED CT Starter Guide. https://docs.snomed.org/snomed-ct-practical-guides/snomed-ct-starter-guide/5-snomed-ct-logical-model (accessed 10 October 2026).
• ISO 25964-1:2011. Information and documentation. Thesauri and interoperability with other vocabularies. Part 1: Thesauri for information retrieval. International Organization for Standardization.