The guide · Part I of IV
Part I - Seeing the Data Layer
From Clinical Meaning in the Age of Health Data & AI - a guide by Dr Balsam Khalil, MD. First edition, 2026.
What this is. Part I of a four-part guide, published in full and without alteration. It sets out what happens to clinical meaning between the consultation room and the record.
Parts II and IV - the standards read clinically, and the physician's role - are not published here. Part III is, and follows this one.
Reference case · one patient, fourteen chapters
"One Patient. Fourteen Chapters." - The Running Clinical Case
This case appears at the end of each part. The same patient. The same clinical event. Seen through a different layer each time. Read it once here. Return to it as the guide develops.
Ahmed is 58. He presents with left-sided chest pain radiating to the jaw, onset two hours ago, diaphoresis, no prior cardiac history documented in the system. His wife mentions he had "a heart problem" three years ago in another country. There is no record. The physician suspects acute coronary syndrome. She documents: "Chest pain - query ACS. Previous cardiac history per family - unverified." She orders a troponin, an ECG, and a cardiology referral. She marks the problem as "suspected" in the structured field. She does not know that the system's coding engine will read "chest pain" from the narrative and ignore the structured field.
Part I
What the physician meant vs. what was documented. The pause before "unverified." The absent record from another country.
Part II
How "suspected ACS" travels - or fails to - through SNOMED CT, ICD-11, and FHIR. What the coding engine reads. What the modifier "suspected" requires to survive.
Part III
What a model trained on this institution's data inherits. Why "chest pain" becomes the signal, not "query ACS."
Part IV
What Ahmed's physician can ask. What she cannot change. Where the decision actually lives.
Ahmed is a composite. The documentation failure is not.
Chapter 1
The Clinical Encounter as an Information Event
A woman of 67 comes in about a rash. The consultation runs twenty minutes. The physician notices weight loss, an absent husband, a telling pause. The rash takes ninety seconds. Everything else takes the rest. Afterwards she writes eleven lines - a good note, competent by every professional standard.
And it is not what happened. It does not contain the pause, the absent husband, the four possibilities weighed and set aside. Documentation is not recording. It is selection under judgement - and it is the only part of the chain that leaves no trace.
The imagined reader has changed, and nobody announced it. The note is now read by a coding process, an extraction pipeline, a decision support rule, a model. None shares the physician's training or capacity to infer what was meant.
What You Should See Differently - by the end of this chapter:
- You should be able to say what a clinical encounter contains that no record can hold, and why that is a property of records rather than a failure of yours.
- You should be able to describe documentation as an act of selection under time pressure, and name what governs the selection when nobody has specified it.
- You should be able to identify, in a note you wrote yourself, at least one thing you knew and did not write - and say what made it fall outside.
- You should understand that the loss is not the problem. The problem is that everything downstream treats what remains as though nothing was left out.
Six Questions to Ask
- What in my documentation is determined by the template rather than by the patient?
- When I omit something I noticed, is it because it was clinically irrelevant, or because there was no field for it?
- Who reads my notes other than clinicians, and does what with them?
- Does anything in my institution measure what documentation omits, as distinct from what it contains?
- When I write under time pressure, which parts go first - and has anyone examined that pattern?
- If a system were trained on everything I have written this year, what would it conclude that I never meant?
Chapter 2
What Save Creates
One act of documentation produces nine objects - each with a separate future. The clinician chose which received which piece of meaning by where she typed, without knowing she was choosing.
- Narrative note text
- Coded problem-list entry
- Encounter diagnosis
- Medication order
- Laboratory / imaging order
- Billing and claims record
- Decision-support evaluation
- Audit-log entry
- Downstream export payload
One save
Creation fans out automatically
Within seconds, messages leave for pharmacy, regional repositories, decision support rules fire.
Correction does not
Correction repairs the source. It does not reach snapshots, undo triggered actions, or notify anyone downstream that something they hold is now known to be wrong.
What You Should See Differently - by the end of this chapter:
- You should be able to name what the act of saving creates that did not exist a moment before, and say how many separate objects it creates from a single note.
- You should be able to describe why those objects diverge over time, and why correcting one does not correct the others.
- You should understand that correction repairs the source and does not reach what the source has already produced - snapshots taken, actions triggered, exports sent.
- You should be able to say who, in your institution, is notified when something they hold is now known to be wrong. And to recognise what it means if the answer is nobody.
Six Questions to Ask
- When I correct a note, what else in the institution changes - and what does not?
- Which systems hold a copy of what I wrote, and when did they take it?
- If a diagnosis I entered was wrong and I amended it, would anyone downstream ever learn that?
- Does my institution distinguish between a record that was corrected and one that was always right?
- What is triggered automatically by the act of saving, before anyone reviews it?
- If I needed to know every place a statement I wrote had travelled to, could anyone tell me?
Part II follows in the full guide, and examines what each standard carries natively, what it carries only when configured, and what it drops without recording that anything was dropped. Part III, which is published here, takes up what happens when a model is trained on what remains.
Parts II and IV are not published here. They are part of the full guide, which will be available for purchase.