Precision nuclear oncology

Definition and scope

Precision nuclear oncology is the practice of choosing, delivering, and evaluating radiopharmaceutical diagnosis and therapy for an individual patient on the basis of that patient's own molecular and clinical characteristics, rather than on the basis of a diagnosis label alone. The word "precision" is borrowed from the broader precision medicine program, which Collins and Varmus (2015) described as treatment and prevention that take individual variability in genes, environment, and lifestyle into account, and which Hood and Friend (2011) had earlier framed as predictive, personalized, preventive, and participatory. The word "nuclear" narrows the tools to radionuclide-based imaging and therapy, and in particular to Theranostics, in which a diagnostic radiopharmaceutical and a therapeutic one share a molecular target so that imaging can show whether the target is present before therapy is attempted (Herrmann and colleagues 2020 describe the field and its open problems).

This monograph is about the knowledge problem inside that practice, not about any particular drug or disease. The NucLex discussion put it plainly: a good decision for one patient may require distilling thousands of records from many domains and stitching them through time (discussion record, section 16). The monograph maps the kinds of information involved, separates observations from interpretations of them, explains why time must be handled with more care than a date column provides, and states what role a shared ontology could play. It makes no claim that NucLex, or any system described here, provides validated decision support. Nothing here is a treatment recommendation, and nothing here asserts that any product is indicated for any condition. The patient in the worked example is invented.

Key distinctions

Observation and interpretation. An observation is something measured or seen: a serum concentration, a standardized uptake value in a region of an image, a pathologist's count of positively stained cells. An interpretation is a judgment made from observations under a rule or by a person: "progressive disease," "PSMA-positive," "high expression." The FHIR Observation resource (HL7 R5) carries both a value and an interpretation element, and the separation is deliberate; the A patient story across time and modality essay develops it narratively. The distinction matters because interpretations depend on criteria that change, and because two readers can share every observation and disagree about the interpretation.

Clinical and molecular information. Clinical information describes the patient as a person and a body: symptoms, examination findings, performance status, prior treatments, organ function. Molecular information describes the disease at the level of molecules and cells: histology, immunohistochemistry, sequencing results, target expression on imaging. The two are not separate worlds; a molecular finding is obtained from a clinical specimen on a clinical date and is read in a clinical context. But they come from different sources, in different formats, with different vocabularies, and the join between them is where the knowledge problem lives.

Observation time, report time, treatment time. An observation has a time at which it was true of the patient (the blood was drawn, the scan was acquired). It has a different time at which it entered the record (the result was released, the report was signed). A treatment has a time at which it was given. These three are routinely collapsed into "the date," and the collapse produces errors that the longitudinal section below makes concrete.

Biomarker and target. A Biomarker is, in the definition of the FDA-NIH BEST resource (2016), a defined characteristic measured as an indicator of normal biological processes, pathogenic processes, or responses to an exposure or intervention. A molecular target is the molecule a drug is designed to bind. In theranostics the two often coincide, since imaging of the target serves as a biomarker for therapy selection, but they are distinct ideas with distinct evidence requirements, and the Molecular targeting monograph treats the target side.

Historical development

The idea that a radiopharmaceutical could be chosen for a patient on the basis of a demonstrated molecular property is older than the word theranostics. Radioiodine for thyroid disease, in clinical use since the 1940s (a date the source check should confirm against a documented history of nuclear medicine), depends on the thyroid's uptake of iodine, and the uptake could be measured before treatment. What changed in the twenty-first century was the number and specificity of targets, the arrival of positron emission tomography (PET) tracers that image a target quantitatively, and the development of therapeutic radiopharmaceuticals that bind the same targets, so that the diagnostic study could serve as a selection test for the therapy (Weber and colleagues 2020 survey this trajectory).

Precision medicine as a program took its name and prominence from the 2015 United States initiative (Collins and Varmus 2015), though treatment selection by molecular marker was already established in oncology through targeted small molecules and antibodies. The informatics side developed more quietly. Hripcsak and Albers (2013) described the problem of deriving reliable phenotypes from records produced for care, not research. Weiskopf and Weng (2013) catalogued the dimensions along which such records fail: completeness, correctness, concordance, plausibility, and currency. The Observational Health Data Sciences and Informatics (OHDSI) community built a common data model so that records from many institutions could be queried the same way (Hripcsak and colleagues 2015). None of these efforts was specific to nuclear medicine; all bear on the question this monograph asks.

Response criteria have their own history. RECIST 1.1 (Eisenhauer and colleagues 2009) defines response by anatomic measurement; PERCIST (Wahl and colleagues 2009) proposed criteria based on PET uptake. Both are interpretations built on observations, and a patient can be classified differently under the two, which is one reason the distinction is not academic.

Philosophical or technical account

A map of the information types

The discussion record (section 16) lists the kinds of records a decision may need to draw on. Arranged by what they are, rather than where they are stored, they form the following map. Every row is a type, not an instance, and the examples are illustrative.

Information type	Typical form	Observation or interpretation	Characteristic time problem
Clinical notes	Free text, structured problem lists	Both, interleaved	Note date often differs from the events described
Procedure notes	Free text with structured elements	Both	Procedure time versus signing time
Laboratory results	Numeric with units and reference ranges	Observation; flags are interpretation	Collection time versus result time
Histopathology	Narrative report with structured synoptic elements	Observation (what was seen) and interpretation (diagnosis, grade)	Specimen date versus report date, sometimes weeks apart
Genomic results	Variant lists with classifications	Observation (variant present) and interpretation (significance)	Classification can change after the report
Imaging reports	Narrative with measurements	Both; measurements are observations, impressions are interpretations	Acquisition versus report versus addendum
Images	Pixel data with acquisition metadata	Observation	Acquisition time is reliable; the rest is derived
Medications	Orders, administrations, dispensing	Observation (what was given)	Order time versus administration time
Interactions	Derived from medication lists and knowledge sources	Interpretation	Depends on the knowledge source version

Synthetic framework: a map of information types relevant to an individualized question in nuclear oncology, with the observation and interpretation status and the characteristic time problem for each.

The table makes one point that is easy to miss in practice. Almost every row contains both observations and interpretations, and the interpretation is usually the more prominent part of the document while the observation is the more durable part of the knowledge. A pathology report's diagnosis line is read; its description of what was stained and how strongly is where the reusable information is.

Longitudinal alignment

Suppose each record carries a single date. Then the patient's history is a sorted list, and a question such as "what was the state of the disease when therapy began" is answered by looking at the records just before the therapy date. This is the model most charts implicitly use, and it fails in at least three ways.

First, the date on a document is often the date of reporting, not of observation. A pathology report dated after a therapy started may describe a biopsy taken before it. A genomic report dated months after the specimen may be the first time anyone knew the result, and the clinician who chose the therapy did not have it. Whether a piece of information was available at decision time is a different question from whether it was true at decision time, and only a record that keeps observation time and report time apart can answer both.

Second, interpretations change without the observation changing. A variant classified as of uncertain significance may be reclassified; a response assessment may be revised on re-read; a reference range may be updated. A record that stores only the latest interpretation has overwritten the state of knowledge that existed when a decision was made.

Third, treatments have durations and sequences, not points. A radioligand therapy given in cycles weeks apart, overlapping with a hormonal therapy started earlier and a supportive medication started later, cannot be represented by three dates. The question "what was the patient exposed to at the time of this scan" needs intervals and their overlap, and the question "what changed between two scans" needs every interval that started or stopped in between.

The Longitudinal record concept page states the minimal requirement: each entry carries the time it was true of the patient, the time it entered the record, its source, and its version, and interpretations are linked to the observations they rest on. That is a requirement on representation, not on any particular software.

Missingness and conflict

Weiskopf and Weng (2013) distinguish completeness from correctness and both from concordance. In an individual patient's record all three fail routinely. A laboratory value is missing because it was not ordered, because it was ordered elsewhere and never imported, or because it was imported into a field no one queries. Two reports give different values for the same measurement because they used different methods, different regions, or different criteria. A note says one thing and the structured problem list says another.

An honest representation does not resolve these; it records them. A missing value is recorded as missing with a reason when the reason is known. Two conflicting observations are both kept, each with its source, and the conflict is surfaced rather than averaged. This is the point at which Provenance stops being an informatics nicety and becomes a clinical necessity: a decision made on a value needs to be traceable to the value's origin, so that when the value turns out to be wrong, the decisions that depended on it can be found.

Where an ontology would sit

Nothing above requires an ontology. A careful database schema with the right time and source columns and a rule against overwriting would meet most of it. What an ontology adds is shared meaning across sources: that the "PSMA" in the pathology report, in the imaging report, and in the therapy order refer to the same molecular target; that uptake values from two scanners are instances of one observable measured under different conditions; that an order, an administration, and an exposure are three different relations between a patient and a product.

The NucLex proposal, as the discussion record states it (sections 3, 4, and 16), is that such a layer would be built by reusing existing terminologies and adding only what nuclear oncology lacks, with every addition carrying its evidence. That is a proposal. The demonstration that would establish progress is a synthetic longitudinal record, like the one below, coded against named terminologies, in which a stated question can be answered by query with provenance attached. No such demonstration has been built.

Biomedical relevance

The pattern described here is general oncology, not nuclear medicine in particular. Any targeted therapy selected by a marker faces the same questions: was the marker measured on the relevant tissue at the relevant time, by a comparable method, and interpreted under the same criteria? The informatics literature on phenotyping (Hripcsak and Albers 2013) and data quality (Weiskopf and Weng 2013) was written for research reuse of records, but the individual patient question is the same question with a sample size of one and without a cohort's statistical protections.

One caution belongs here. A longitudinal record, however well represented, is observational. It can show what was done and what followed; it cannot, by itself, show what would have happened under a different choice. Hernán and Robins (2020) set out the demanding conditions under which observational data can support a causal claim. A system that reads a record and reports "patients like this responded to therapy X" has made a causal-sounding statement from data that cannot, without much more, support it. That is one reason the review gate on this monograph says what it says.

Nuclear medicine relevance

The following timeline is entirely synthetic. The patient does not exist; the values are chosen to illustrate representational problems and have no clinical meaning; nothing in it is a recommendation about any treatment or any drug.

A man in his late sixties is being considered for a radioligand therapy that targets a cell-surface protein. His record, viewed as a flat list of documents sorted by document date, reads cleanly. Viewed with observation time and report time separated, it reads differently.

Entry	Observation time	Report time	Source	Observation	Interpretation attached
Serum marker	Day 0	Day 0	Lab A	42 units (unit per local lab method)	Flagged "high" against Lab A range
Biopsy	Day 3	Day 19	Pathology	Staining for the target: moderate intensity in a stated percentage of tumor cells	"Target expressed"
Genomic panel	Day 3 (same specimen)	Day 41	External lab	Variant in a DNA repair gene	"Uncertain significance" (later revised)
Target PET	Day 12	Day 13	Imaging	Uptake values per lesion; two lesions below liver reference	"Target-positive disease; two discordant lesions"
Therapy cycle 1	Day 30	Day 30	Nuclear medicine	Administered activity, product lot	None
Serum marker	Day 28	Day 28	Lab B	51 units (different method)	Flagged "high" against Lab B range
Genomic revision	Day 3 (same specimen)	Day 95	External lab	Same variant	Reclassified as "likely pathogenic"
Target PET	Day 72	Day 74	Imaging	Uptake values per lesion; one new focus	"Mixed response" (criteria named in report)

Synthetic example: an invented longitudinal record for a hypothetical patient, with observation and report times kept apart. Values are placeholders and carry no clinical meaning.

Several things are visible only because the columns are separated. The genomic result was not available on Day 30 when cycle 1 was given; a reviewer looking at document dates would see a report dated Day 41 and might not notice that the therapy decision predated it. The same result was reinterpreted on Day 95 without any new observation; a record that stored only the current interpretation would show "likely pathogenic" as if it had always been known. The two serum marker values came from two laboratories with two methods; whether 42 and 51 are comparable is a question about the observable, not about the numbers, and the Context in which each was measured has to travel with the value. The imaging interpretation "target-positive disease" on Day 13 rests on per-lesion observations that include two lesions below the reference the reader used; the interpretation is defensible, but it is an interpretation, and a different reference or a different criterion could yield a different one (the joint procedure guideline by Fendler and colleagues 2023 describes how such reads are standardized, and standardization is exactly what makes the criterion a thing that can be named and versioned).

Now ask a precise question: "At the time cycle 1 was given, what evidence of target expression existed, and from what?" The answer the separated record gives is: a pathology observation from Day 3 reported Day 19, and an imaging observation from Day 12 reported Day 13, each with its own method and its own interpretation, and no genomic information. The answer a flat record gives is a list of documents, and the reader has to reconstruct the rest.

The ontology role is in the words "target expression," "observation," "interpretation," "method," and "criterion." If each is a concept with a definition, the question can be posed to the record rather than to a person reading it. If the record's entries are coded against those concepts with their provenance, the answer can be returned with its evidence. NucLex proposes that layer. It has not built it, and this example is a specification of what a demonstration would have to show, not a report of one.

Disagreements and limitations

Validated decision support is not claimed and must not be read in. Everything on this page concerns representation: what a record must hold so that a question can be answered honestly. None of it establishes that a system built on such a record would improve decisions, and a system that answered the synthetic question above correctly would still have demonstrated representation, not clinical benefit. Clinical benefit requires prospective evaluation that this project has not designed, let alone conducted.

The information map is a framework, not a finding. The table of information types is the writer's organization of the discussion record's list. Other organizations are possible, and a domain reviewer may reasonably move rows or add them. What should survive review is the claim that each type carries both observations and interpretations and has its own time problem.

Classification has consequences beyond accuracy. Bowker and Star (1999) showed that classification systems shape what is recorded, what is noticed, and what becomes invisible. A record that codes target expression as present or absent will lose the "moderate intensity in a stated percentage" that the pathologist saw, and a later question about degree of expression will be unanswerable. The choice of what to make a concept is an editorial act with clinical consequences.

The synthetic example is simpler than practice. Real records contain hundreds of entries, multiple institutions, scanned documents without structured content, and errors of every kind catalogued by Weiskopf and Weng. The example isolates the time and interpretation problems; it does not represent the scale.

No real data were used. The timeline, the values, and the patient are inventions. The writer has not consulted any clinical record in producing this page, and no reviewer should treat the example as derived from one.

Related entries

• Precision nuclear oncology, Longitudinal record, Biomarker, and Theranostics: the lexicon entries for the central terms.
• Context: what must travel with a measurement for it to be comparable.
• Provenance, Evidence, and Uncertainty: the concepts a decision-time question depends on.
• Provenance and evidence: tracing one claim through an evidence record, including corrections such as the genomic reclassification above.
• Molecular targeting: the target side of the biomarker and target distinction.
• Biodistribution and pharmacokinetics: where the administered agent goes, which is the next question after target expression.
• A patient story across time and modality: the narrative companion to this monograph.
• Semantic interoperability: why the "PSMA" in three documents may or may not mean one thing.

References

1. Collins FS, Varmus H. A new initiative on precision medicine. New England Journal of Medicine. 2015;372(9):793-795. Supports: the definition and national framing of precision medicine.
2. Hood L, Friend SH. Predictive, personalized, preventive, participatory (P4) cancer medicine. Nature Reviews Clinical Oncology. 2011;8(3):184-187. Supports: the earlier P4 framing.
3. Herrmann K, Schwaiger M, Lewis JS, et al. Radiotheranostics: a roadmap for future development. Lancet Oncology. 2020;21(3):e146-e156. Supports: the description of theranostics and its open problems.
4. Weber WA, Czernin J, Anderson CJ, et al. The future of nuclear medicine, molecular imaging, and theranostics. Journal of Nuclear Medicine. 2020;61(Suppl 2):263S-272S. Supports: the historical trajectory toward target-specific imaging and therapy.
5. FDA-NIH Biomarker Working Group. BEST (Biomarkers, EndpointS, and other Tools) Resource. 2016. https://www.ncbi.nlm.nih.gov/books/NBK326791/. Supports: the definition of biomarker.
6. Eisenhauer EA, Therasse P, Bogaerts J, et al. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). European Journal of Cancer. 2009;45(2):228-247. Supports: anatomic response criteria as a named interpretation rule.
7. Wahl RL, Jacene H, Kasamon Y, Lodge MA. From RECIST to PERCIST: evolving considerations for PET response criteria in solid tumors. Journal of Nuclear Medicine. 2009;50(Suppl 1):122S-150S. Supports: PET-based response criteria as a distinct interpretation rule.
8. Fendler WP, Eiber M, Beheshti M, et al. PSMA PET/CT: joint EANM procedure guideline/SNMMI procedure standard for prostate cancer imaging 2.0. European Journal of Nuclear Medicine and Molecular Imaging. 2023;50(5):1466-1486. Supports: the existence of standardized reading criteria for a target PET study. Cited for the existence of such standardization, not for any clinical claim.
9. Hripcsak G, Albers DJ. Next-generation phenotyping of electronic health records. Journal of the American Medical Informatics Association. 2013;20(1):117-121. Supports: the difficulty of deriving reliable phenotypes from records produced for care.
10. Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. Journal of the American Medical Informatics Association. 2013;20(1):144-151. Supports: the dimensions of record data quality (completeness, correctness, concordance, plausibility, currency).
11. Hripcsak G, Duke JD, Shah NH, et al. Observational Health Data Sciences and Informatics (OHDSI): opportunities for observational researchers. Studies in Health Technology and Informatics. 2015;216:574-578. Supports: the existence of a common data model for cross-institutional query.
12. HL7. FHIR R5. Resource Observation. https://www.hl7.org/fhir/R5/observation.html. Supports: the separation of value and interpretation, and of effective time and issued time, in a widely used exchange resource.
13. Hernán MA, Robins JM. Causal Inference: What If. Chapman and Hall/CRC; 2020. Supports: the conditions under which observational data can support causal claims.
14. Bowker GC, Star SL. Sorting Things Out: Classification and Its Consequences. MIT Press; 1999. Supports: the claim that classification choices shape what is recorded and noticed.