A patient story across time and modality

Key points

• A clinical record is not a list of facts about a patient. It is a list of acts of observation and interpretation, each performed by someone, at some time, with some method, and recorded at a different time. Strip those four things away and the "facts" stop being comparable.
• Observation time and report time differ for almost every kind of evidence, and the gap is not noise. A pathology addendum, a reclassified genomic variant, an outside discharge summary arriving three weeks late, and a medication list that no longer matches what the patient takes are all cases where the record's time and the world's time have come apart.
• Missingness is structured. Values are absent because a test was not ordered, because it was done elsewhere and never transmitted, because an image was acquired but never read, or because a patient did not come back. Each of these means something different, and a system that treats them alike will reason badly.
• An observation (a number, an image, a slide) and its interpretation (a category, a grade, a response call) are different kinds of things. Interpretations change when criteria, databases, and readers change, while the observation stays what it was. The record must keep both and keep them apart.
• The NucLex proposal in this area is a shared vocabulary for the kinds of things in such a record and the relationships among them, so that integration can preserve context. It is a proposal for knowledge representation. It is not a decision-support service, and this publication makes no treatment recommendation.

An invented patient

Everything in this essay is synthetic. The patient, his physicians, the institutions, the dates, the laboratory values, the agents, and the findings are invented for teaching. No real patient record was consulted or combined. The imaging and therapeutic agents are the local teaching labels used elsewhere in this publication (Agent A, Agent B, Target P) and refer to no product. No statement here is a recommendation about diagnosis or treatment, and the essay names no regimen as appropriate for anyone.

Call him Mr. Varga. He is sixty-eight, and over fourteen months he is seen by a primary care physician, a urologist, a pathologist, a molecular laboratory, two radiology departments, a nuclear medicine department, an outside emergency department, and a pharmacy system that has never met him. Each of them observes something, writes something down, and moves on. By the end, his record holds several hundred entries. The question this essay asks is not what should be done for Mr. Varga. It is what a person, or a program, would need to know about each entry in order to read the record as a single coherent story rather than a pile.

The first-section essays argued that a shared language needs the term and the concept kept apart (Why medicine needs a shared language). This essay makes a parallel argument about time and evidence: the record needs the observation and its interpretation kept apart, and it needs the time at which something happened kept apart from the time at which someone wrote it down. The lexicon entry on Longitudinal record gives the short form. The rest of this essay is the long one.

The timeline

Figure 1. A synthetic longitudinal timeline with separate event and report dates (text table, synthetic example). "Event date" is when the observation was made or the thing happened. "Report date" is when the corresponding record became available in the receiving system. Blank cells and the rows marked "gap" are deliberate. All content is invented.

Row	Event date	Report date	Source	What was observed	What was interpreted or asserted	Note on gap or conflict
1	2031-01-08	2031-01-10	Primary care note	Serum marker M1 elevated (lab drawn 01-08); patient reports pelvic discomfort "for a few weeks"	"Elevated M1; refer to urology"	Symptom onset is undated; note written two days after the visit
2	2031-01-22	2031-01-29	Histopathology report	Core biopsy of a pelvic lesion; tumor present; grade assigned	"Carcinoma, grade G2 (local scheme)"	Seven-day lag between specimen and report
3	2031-01-22	2031-02-05	Histopathology addendum	Immunohistochemistry on the same specimen: Target P expression "low to moderate, focal"	Addendum to row 2; "Target P expression present but focal"	Same specimen, second report date; the addendum revises the interpretation available on 01-29
4	2031-02-01	2031-02-26	Molecular laboratory	Tissue panel from the row 2 block; variant in gene G detected	"Variant of uncertain significance" (laboratory's classification, knowledge-base version 2030.4)	25-day lag; classification depends on a database version
5	2031-02-12	2031-02-12 (preliminary); 2031-02-14 (final)	Nuclear medicine, Agent A PET/CT	Intense uptake at the biopsied pelvic lesion and at two bone sites; measured uptake values recorded per lesion	Preliminary: "extensive Target P-avid disease." Final: same, with one bone site downgraded to "equivocal" after review with CT	Two report versions for one acquisition; the final contradicts the preliminary on one site
6	2031-02-12	(none)	Nuclear medicine, Agent A PET/CT images	The image data themselves (DICOM series)	No separate assertion; the images are the observation	Images and report are distinct objects with distinct identifiers
7	2031-02-20	2031-02-20	Tumor board note	Discussion of rows 2 to 5	Board records its conclusion (plan to discuss Agent B radioligand therapy with the patient); recorded as a discussion outcome, not an order	The note cites the preliminary PET read (row 5), not the final
8	2031-03-03	2031-03-03	Chemistry laboratory (in-house)	Hemoglobin 12.9 g/dL; creatinine 1.1 mg/dL	Reference range applied; flagged "normal"	Baseline for therapy
9	2031-03-05	2031-03-05 (hot lab log); 2031-03-05 (EHR)	Nuclear medicine, Agent B cycle 1	Hot lab log: activity measured in dose calibrator at 09:40, residual measured at 10:35, net administered activity recorded with both times. EHR: "administered activity" field shows the ordered activity	Two values for "administered activity" in two systems	Conflict: ordered versus measured net activity; the EHR value lacks a measurement time
10	2031-03-06	(none)	Nuclear medicine, post-therapy SPECT/CT	Images acquired the day after cycle 1	No report issued	Gap: observation without interpretation
11	2031-04-16	2031-04-16	Chemistry laboratory (in-house)	Hemoglobin 11.8 g/dL; creatinine 1.2 mg/dL	Flagged "normal"	Cycle 2 baseline
12	2031-04-18	2031-04-18	Nuclear medicine, Agent B cycle 2	As row 9	As row 9	Same conflict repeats
13	2031-05-28	(none)	Outside laboratory	Blood drawn before cycle 3 at a laboratory near the patient's home	Result never transmitted	Gap: the record shows no cycle 3 baseline; the test was done
14	2031-05-30	2031-05-30	Nuclear medicine, Agent B cycle 3	As row 9	As row 9; the pre-treatment note states "labs reviewed"	The note asserts a review of values the system does not contain
15	2031-06-20	2031-06-20	Pharmacy interaction check	Medication list includes oral Drug Q (started 2030-11 by primary care) and a new supportive medication	Interaction flag raised; severity "major" in knowledge base 1, "moderate" in knowledge base 2	Conflict between two drug knowledge bases on the same pair
16	2031-04 (approximate)	2031-06-20	Patient statement at visit	Patient reports that he stopped Drug Q "around April"	Medication list not updated until 06-20; still showed "active"	Event time is approximate and precedes the record time by two months; the row 15 flag concerned a drug he was no longer taking
17	2031-07-09	2031-07-11	Nuclear medicine, Agent A PET/CT (follow-up)	Decreased uptake at the pelvic lesion and one bone site; a new focus in a rib; measured values recorded	"Partial response by PET-based criteria; one new focus, indeterminate"	The same study would be "progression" by anatomic criteria that count any new lesion
18	2031-07-09	2031-07-11	Chemistry laboratory	Hemoglobin reported as 108 (no unit in the transmitted message)	Flagged "out of range, critical" by the receiving system, whose reference range is in g/dL	Conflict: the value is 10.8 g/dL, transmitted as 108 g/L without the unit; the flag is a unit artifact
19	2031-08-14	2031-09-04	Outside emergency department discharge summary (via health information exchange)	ED visit for rib pain; CT chest read by the outside radiologist	"New rib lesion, concerning for progression"	21-day lag; the in-house reader considers this the row 17 focus, measured on a different modality
20	2031-09-18	2031-09-18	Molecular laboratory, reissued report	No new tissue; the row 4 variant reclassified	"Likely pathogenic" (knowledge-base version 2031.2)	Same observation as row 4, new interpretation, seven months later
21	2031-10 to 2032-02	(none)	(none)	No encounters recorded		Gap: the record is silent; the patient may have been seen elsewhere, or not at all
22	2032-03-02	2032-03-02	Nuclear medicine note	Patient returns; reports he "felt fine and stopped coming"	"Lost to follow-up October to February; resumes surveillance"	The silence in row 21 is now explained, by the patient, retrospectively

Alt text: a twenty-two-row table following an invented patient from January 2031 to March 2032. Each row gives an event date, a report date, the source, the observation, the interpretation, and a note. Several rows have no report date (unreported images, an untransmitted laboratory result, a four-month silence). Several rows record the same event twice with different values or different dates, and several note that a later interpretation contradicts an earlier one.

Two clocks

Every row in Figure 1 has two dates, and in most rows they differ. This is not a quirk of a badly run hospital. It is a property of how evidence is produced.

The pathology report (row 2) carries a date seven days after the biopsy because tissue must be fixed, cut, stained, and read. The addendum (row 3) carries a date a week later still, because the immunohistochemistry was ordered after the first read. The genomic panel (row 4) took twenty-five days. The outside discharge summary (row 19) took three weeks to arrive through a health information exchange. The medication change (row 16) took two months to reach the record because nobody asked until a flag was raised. In each case there is a moment at which the thing was true of the patient, and a later moment at which the record began to say so.

Informatics standards recognize the distinction explicitly. The HL7 FHIR Observation resource carries separate elements for the clinically relevant time of the observation and for the time the result was made available, which the specification calls effective[x] and issued respectively (HL7, FHIR R5 Observation). The DICOM standard that governs medical images distinguishes the date and time of acquisition from the dates of the study and of the series and from the time any derived object was created (DICOM PS3.3). Hripcsak and Albers have argued that the time stamps in a record reflect the health care process as much as the patient's physiology, and that analysis which ignores the process will mistake its artifacts for biology (Hripcsak and Albers, 2013; Hripcsak et al., 2011). Agniel, Kohane, and Weber showed in a large retrospective study that the mere fact and timing of a laboratory test being ordered carried information about outcomes independent of the test's value (Agniel et al., 2018). A record is a trace of what clinicians did, when, and the "when" has two readings.

Why does it matter for Mr. Varga? Consider row 7. The tumor board met on 20 February and recorded its conclusion, citing the PET/CT read. The final read (row 5) had been available since 14 February and had downgraded one bone site to equivocal. The board's note cites the preliminary. Nothing in the note is false; the preliminary read existed and said what the note says it said. But a reader who sees only the board note and the final report, without the two report dates and the version history, will believe the board considered the final read and disagreed with it, when in fact the board never saw it. Whether that mattered clinically is not the point of this essay. The point is that the record cannot even pose the question unless both dates travel with the observation.

What is missing, and why

Figure 1 has four kinds of absence, and they are not the same.

Row 10 is an observation with no interpretation. The post-therapy images were acquired, exist as DICOM objects, and were never read. A system searching for "post-therapy imaging reports" finds nothing and may conclude that no imaging was done. A system searching for images finds them and may conclude that they were assessed. Both conclusions are wrong, and both are available from a record that stores the image and the report as the same kind of thing.

Row 13 is an observation that was made and never transmitted. The blood was drawn, the laboratory produced a result, and the result exists in the outside laboratory's system. In Mr. Varga's record it is as if the test had never been ordered. Row 14 then makes the absence worse: the pre-treatment note says "labs reviewed." A clinician may well have reviewed a printed result the patient brought in. The record does not know that, and a program that trusts the note will believe a value exists that it cannot find, while a program that trusts the table will believe the note is wrong.

Row 16 is a change in the world that reached the record late and approximately. The patient stopped a drug "around April." The record said "active" until June. Between April and June every interaction check (row 15 included) was run against a medication he was not taking. The event time is itself uncertain: "around April" is the best anyone will ever have.

Row 21 is a silence. For four months the record contains nothing. The possibilities are that the patient was well and chose not to come, that he was seen elsewhere and the records have not arrived, that he was ill and did not come, or that he was in the record under a different identifier. Row 22 resolves it, retrospectively and only on the patient's say-so.

The statistical literature on missing data distinguishes cases by whether the absence is independent of everything, related to observed values, or related to the missing value itself, because the right method of analysis differs for each (Little and Rubin, 2019). Weiskopf and Weng, reviewing how researchers assess electronic health record data quality, found that completeness was the most commonly assessed dimension and that definitions of it varied widely across studies (Weiskopf and Weng, 2013). The clinical record adds a dimension the statistics does not have: the reason for the absence is often recoverable, from a note, a transmission log, or the patient, and a representation that can record "absent because not ordered," "absent because not transmitted," "absent because not read," and "absent because the patient did not return" preserves information that "null" destroys.

When the evidence disagrees

The timeline also contains five conflicts, and they are of different kinds.

Rows 5 and 7 conflict in version: a preliminary and a final report of one acquisition disagree on one site, and a downstream document cites the superseded version.

Row 9 conflicts in method: two systems hold two values for "administered activity." The hot lab log records an activity measured in a dose calibrator before administration, a residual measured after, and the net, with times. The electronic record shows the activity that was ordered. In a radiopharmaceutical therapy the difference between ordered and net administered activity is a real quantity, and so is the decay between measurement and administration. Anyone later wanting to relate exposure to outcome, or to perform Dosimetry, needs the measured value with its time, not the ordered value without one. The two systems are not lying. They are recording different things under the same label.

Row 15 conflicts in source: two drug knowledge bases assign different severities to the same interaction. There is no fact of the matter in the record that resolves it; the severity is each base's interpretation. And row 16 then shows that the interaction concerned a drug the patient had stopped, so both severities were applied to a state of the world that no longer obtained.

Row 18 conflicts in units: a hemoglobin of 10.8 g/dL was transmitted as 108 (the same value in g/L) with no unit, read against a range in g/dL, and flagged critical. The value was right; the message lost its unit; the flag is an artifact. This is the kind of error that a unit-bearing representation prevents and a bare number invites.

Rows 17 and 19 conflict in criteria, and this is the conflict the first essay in this publication opened with. The in-house PET/CT follow-up calls the study a partial response by PET-based criteria, with a new indeterminate focus. The outside CT calls a rib lesion new and concerning. The in-house reader thinks the two are the same focus seen on two modalities, three weeks apart, one of them measured anatomically and one metabolically. Published response frameworks differ precisely on how a new lesion counts (Eisenhauer et al., 2009; Wahl et al., 2009). Neither reader is wrong under the reader's own criteria. The record holds two interpretations of overlapping observations, made under different frameworks, and a reader who does not know the frameworks will see a contradiction where there is a difference of criterion.

Row 20 is the quiet one. The same genomic observation, from the same block, is reclassified seven months later because the laboratory's knowledge base moved from one version to another. The ACMG and AMP guidelines for variant interpretation define classification tiers and the evidence required for each, and they anticipate that classifications will change as evidence accumulates (Richards et al., 2015). Nothing about Mr. Varga changed between February and September. What changed was the state of knowledge, and the record must be able to say so: same observation, new interpretation, new knowledge-base version, new date.

Observation and interpretation are different kinds of things

The pattern across every conflict is the same. Something was observed: a slide, a value, an image, a measured activity, a sequence. Then something was asserted about it: a grade, a flag, a response category, a variant class, a severity. The observation does not change. The assertion changes with the reader, the criteria, the knowledge base, and the date.

Rector, Nowlan, and Kay argued in 1991 that an electronic medical record should be a faithful record of what clinicians have observed, thought, and done, and that it must therefore be able to represent statements that are false, uncertain, superseded, or contradicted, because clinicians make such statements and the record is a record of them (Rector et al., 1991). The argument has aged well. A record that stores only the current best interpretation of each observation cannot answer what the tumor board knew on 20 February, cannot explain why an interaction check fired in June, cannot show that the variant was "uncertain" when treatment decisions were being made, and cannot distinguish a change in the patient from a change in the criteria.

The distinction has a cost. Keeping observations and interpretations apart means storing more, linking more, and resisting the convenience of a single "status" field. Bowker and Star observed that classification systems become invisible as they become infrastructure, and that the work of maintaining the distinctions they rely on is chronically under-resourced (Bowker and Star, 1999). Every "status" field is a classification that has become invisible. The timeline above is an attempt to make a few of them visible again.

What, then, must travel with an observation for it to remain usable? At least the following: what was observed, by what method, with what units; when the observation was made and when it was recorded; who or what made it; what specimen, image, or encounter it belongs to; which interpretations have been attached to it, by whom, under what criteria or knowledge-base version, and when; which of those interpretations supersede which; and, where something is absent, why. The lexicon entries on Provenance, Context, and Uncertainty each cover one piece. The W3C PROV ontology gives one formal vocabulary for the "who, when, derived from what" part (W3C, PROV-O, 2013), and the FHIR Provenance resource gives another, scoped to health data (HL7, FHIR R5 Provenance). Neither is a complete answer, and neither is the subject of this essay. They are evidence that the problem is recognized and that partial vocabularies exist.

Why a nuclear oncology record is the hard case

The kinds of evidence in Figure 1 are ordinary. What makes the nuclear oncology record a demanding case is that it holds all of them at once, and adds some of its own.

A Biomarker appears in three forms in the timeline: as a serum value (row 1), as tissue expression on a slide (row 3), and as the uptake of an imaging agent that binds the same target in the living patient (row 5). These are three observations of something that clinicians speak of as one thing, Target P expression, and they do not agree with one another in any simple way. The slide said "low to moderate, focal"; the scan said "intense." A biopsy samples a few cubic millimeters; a PET scan integrates over the lesion and over the whole body. The disagreement is informative, and it is only informative if the record knows that these three observations are about a common target and were made by three incommensurable methods.

A theranostic course adds the administered agent as a quantity with its own measurement history (row 9), post-therapy images that may or may not be read (row 10), and a response assessment that depends on which criteria the reader applies (rows 17 and 19). The imaging agent and the therapeutic agent bind the same target by design; the record must relate them without conflating them, which is the distinction the lexicon entry on Theranostics draws.

And the whole of it is longitudinal. The question that matters for the field this publication calls Precision nuclear oncology is not what any one row says but how the rows relate across time: whether the uptake in July is lower than in February on the same scanner with the same reconstruction; whether the hemoglobin in July is a real decline from March or a unit artifact; whether the variant that was uncertain in February and likely pathogenic in September should have changed anything in between; whether the four-month silence hides something. The monograph on Precision nuclear oncology develops the argument that this integration is the central informational problem of the field. This essay has tried to show, with one invented patient, how much of that integration depends on things that are routinely dropped: the second date, the unit, the version, the reason for the blank.

The proposed NucLex role

This publication exists because of a position its founders took in discussion: that a good decision for an individual patient can require distilling an enormous number of records from many domains, stitched through time, with medications and their interactions included, and that this requires ontology-supported integration. The timeline above is an attempt to make that position concrete without overstating what follows from it.

What NucLex proposes in this area is knowledge representation. Specifically, it proposes a shared vocabulary for the kinds of things in a longitudinal record (observation, interpretation, report, report version, addendum, specimen, image series, administration, measured quantity with unit and time, knowledge-base version, reason for absence) and for the relationships among them (an interpretation interprets an observation; an addendum revises a report; a report reports on an acquisition; a measured activity is measured at a time by an instrument; a classification is made under a named version of a knowledge base). With such a vocabulary, the twenty-two rows of Figure 1 could be recorded so that the conflicts are visible as conflicts, the gaps are typed by their reason, and the two clocks are kept apart. The aim is that a human reader, or a program working for one, can reconstruct what was known, by whom, when.

It is important to say what this is not. It is not a decision-support service, and NucLex does not propose to recommend treatment for Mr. Varga or anyone. It is not an implemented system; no NucLex vocabulary, record model, or integration tool exists in this release. It is not a claim that ontology-supported integration improves outcomes; such a claim would require clinical evidence that does not exist here and would need to be stated as a research question. And it is not a substitute for the standards already mentioned; the proposal is to build on FHIR, DICOM, PROV, and SNOMED CT where they already carry what is needed, and to supply vocabulary only where the nuclear oncology record has needs those standards do not yet meet. Where that boundary lies is itself a question for the monographs on Provenance and evidence and SNOMED CT and the NucLex niche.

Mr. Varga's story has no ending here because it is not a story about Mr. Varga. It is a story about what twenty-two entries in a record would have to carry so that the people reading them, months or years later, could tell what happened. The answer this essay offers is that an observation must travel with its time, its method, its unit, its source, its interpretations and their versions, and the reasons for what is missing around it. That is a representational proposal. Whether it can be built, and whether building it helps anyone, are the questions the rest of this section leaves open.

Limitations

This is an AI-assisted draft that has not been source-checked or reviewed by a domain expert or an editor. Its claims should be read accordingly.

The patient, dates, values, institutions, and findings are wholly invented. Agent A, Agent B, Target P, gene G, marker M1, and Drug Q are teaching labels that correspond to no product, target, gene, assay, or drug. The laboratory values and their flags are illustrative and are not offered as reference ranges. The sequence of events was constructed to display particular problems (two clocks, missingness, conflict) and does not represent a typical, recommended, or observed course of care. Nothing in the essay is a diagnosis, a treatment recommendation, a statement about any drug's indication, or a comparison of the efficacy of any interventions.

The essay paraphrases the general approach of published response criteria and variant-classification guidelines and does not reproduce their rules; readers who need the actual definitions should consult the cited documents. The description of FHIR Observation and Provenance elements and of DICOM date and time attributes follows the cited specifications as the writer understands them and must be verified against the editions in force at source check.

The claim that ontology-supported integration is needed for individualized theranostic care is the position recorded in the NucLex founding discussion. It is presented here as the project's proposal and motivation, not as an established finding. No evidence is offered that such integration improves clinical outcomes, and the essay makes no such claim.

Clinical and informatics reviewers should check the timeline for plausibility and for any unintended implication about care. The essay describes no implemented NucLex capability. No NucLex vocabulary, record model, integration tool, or decision-support service exists in this release.

Sources and further reading

• HL7 International. FHIR R5 Observation resource. https://www.hl7.org/fhir/R5/observation.html (access to be checked at source check).
• HL7 International. FHIR R5 Provenance resource. https://www.hl7.org/fhir/R5/provenance.html (access to be checked at source check).
• W3C Provenance Working Group. 2013. PROV-O: The PROV Ontology. W3C Recommendation, 30 April 2013. https://www.w3.org/TR/prov-o/
• DICOM Standards Committee. DICOM PS3.3 Information Object Definitions (current edition). https://www.dicomstandard.org/current
• Rector AL, Nowlan WA, Kay S. 1991. Foundations for an electronic medical record. Methods of Information in Medicine 30(3):179-186.
• Hripcsak G, Albers DJ. 2013. Next-generation phenotyping of electronic health records. Journal of the American Medical Informatics Association 20(1):117-121.
• Hripcsak G, Albers DJ, Perotte A. 2011. Exploiting time in electronic health record correlations. Journal of the American Medical Informatics Association 18(Suppl 1):i109-i115.
• Weiskopf NG, Weng C. 2013. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. Journal of the American Medical Informatics Association 20(1):144-151.
• Agniel D, Kohane IS, Weber GM. 2018. Biases in electronic health record data due to processes within the healthcare system: retrospective observational study. BMJ 361:k1479.
• Little RJA, Rubin DB. 2019. Statistical Analysis with Missing Data. 3rd ed. Wiley.
• Richards S, Aziz N, Bale S, et al. 2015. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine 17(5):405-424.
• Wahl RL, Jacene H, Kasamon Y, Lodge MA. 2009. From RECIST to PERCIST: evolving considerations for PET response criteria in solid tumors. Journal of Nuclear Medicine 50(Suppl 1):122S-150S.
• Eisenhauer EA, Therasse P, Bogaerts J, et al. 2009. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). European Journal of Cancer 45(2):228-247.
• Bowker GC, Star SL. 1999. Sorting Things Out: Classification and Its Consequences. MIT Press.