What High-Quality Data Is · the 4+5 Dimensions

This chapter is the “professional muscle” of the whole talk. It has to be driven home.

The traditional base: the 4 dimensions of the national EMR grading standard

The National Health Commission’s Application Maturity Grading Standard for EMR Systems defines four data-quality dimensions:

Dimension	Definition	Counter-example
Completeness	Every field that should be there is there; no key item missing	A note missing the chief complaint or the discharge diagnosis
Consistency	The same information is logically consistent across systems and tables	Patient sex is male in HIS, female in LIS
Integration	Data across systems and across business lines can be linked into a complete view	Lab orders can’t be linked to physician orders
Timeliness	Data is entered on schedule, not backfilled after the fact	Notes backfilled in bulk after the fact

⚠️ A tell of expertise: a lot of people mis-state “integration” as “correlation.” The original term is Integration — it stresses the inner logical closure that comes after cross-domain integration. Correlation is a statistical notion; it’s not the same thing. A one-word slip like this exposes how deep your expertise really goes.

The turn: the traditional 4 dimensions aren’t enough

The traditional 4 dimensions solve “the business can run,” not “AI can learn.”

Dimension	Traditional (operations view)	AI era (training / inference view)
Goal	Data is usable for care and management	Data lets a model learn, reason, and decide
Standard	Compliant, closed-loop, auditable	Compliant + diverse + knowledge-dense + aligned to human expert judgment
Fatal flaw	Missing, wrong	Bias, spurious correlation, shallow knowledge

The traditional 4 dimensions are the entry ticket. The age of AI redefines what “high quality” means.

An international reference: how to build a clinical-note quality dataset

Clinical note quality control isn’t just “did they finish writing it.” It’s whether the note is useful, accurate, concise, well-organized, and easy for the next physician to pick up.

There’s a mature international approach for this called PDQI-9 (Physician Documentation Quality Instrument). It wasn’t invented for the AI era — it’s an earlier rating scale for the quality of EMR notes, asking whether a note is up-to-date, accurate, thorough, useful, organized, comprehensible, succinct, synthesized, and internally consistent. In other words, a clinical-note quality dataset need not be a simple “typos / missing fields” dataset. It can be:

• An original note, or an AI-generated note
• The physician’s final signed version
• Scores from multiple clinical experts, by PDQI-9 or your hospital’s own QC rubric
• Structured annotation of the discrepancies: omissions, hallucinations, redundancy, logical inconsistency, missing care plan

This kind of dataset is enormously valuable, because it answers one question directly: can the AI-generated note be safely picked up and signed by a physician?

But note the caveat: PDQI-9 is not a universal yardstick. Quality standards differ across specialties and across the ER / outpatient / inpatient settings. The sturdier move is: treat PDQI-9 as the starting point, then layer on your hospital’s own QC rules and specialty-expert scoring. See: Research on PDQI-9 for EMR quality assessment, and the recent line of work using PDQI-9 to evaluate AI-generated clinical notes.

The age of AI adds five more dimensions

The base 4 dimensions (completeness, consistency, integration, timeliness) stay — they’re the entry ticket. The AI era adds:

5️⃣ Standardization & computability (Standardization)

• Terminology coding (ICD/SNOMED/LOINC/RxNorm), unit standards, ontology mapping
• Skip this step and all the data in the world is just a data ruin

6️⃣ Representativeness & diversity (Representativeness & Diversity)

• Population coverage (age, sex, geography, comorbidity)
• Sample density for rare diseases and long-tail scenarios
• You can’t let the AI only handle textbook cases and collapse the moment it hits an edge

7️⃣ Annotation & expert-consensus quality (Annotation Quality)

• Who annotated it (a resident vs. an attending vs. a multi-expert consensus)
• Inter-annotator agreement (a Kappa coefficient > 0.8 is the floor)
• In medicine this one all but sets the ceiling — especially for imaging, pathology, and genomic interpretation

8️⃣ Knowledge density & reasoning learnability (Knowledge Density)

• Not just “a lot of data,” but “how much expert reasoning each record carries”
• One note with a full differential-diagnosis line of thought ≫ a hundred notes with conclusions only
• This is why reasoning models like o1 / Mythos depend on expert reasoning-trace data
• This is a dimension added in 2026 — 90% of the people talking about data governance never mention it

9️⃣ Compliance, safety & lineage (Compliance & Lineage)

• Privacy de-identification, informed consent, ethics approval
• Traceable data lineage — the make-or-break for medical AI reaching market
• Aligned with the Data Security Law, the Personal Information Protection Law, and the Measures for the Management of Generative AI Services

The 9 dimensions, each with a concrete medical example

The abstract 9 dimensions go in one ear and out the other. Pair each with a concrete scenario and they land:

Dimension	Concrete medical example
Completeness	Chief complaint / HPI / discharge diagnosis are mandatory fields
Consistency	Patient sex matches across HIS / LIS / EMR
Integration	Lab orders ↔ physician orders ↔ notes are queryable end to end
Timeliness	Orders entered in real time; backfilling not allowed
Standardization	ICD-11, LOINC, SNOMED CT mapped hospital-wide
Representativeness	Rare-disease coverage; cross-province population distribution
Annotation quality	Three attendings blind-label imaging; Kappa > 0.8
Knowledge density	Notes must include the differential-diagnosis line of thought
Compliance & lineage	Training data’s lineage traces back to the original note

Extra dimensions for the legal industry

Beyond the universal 9 dimensions, the legal industry has a few of its own:

• Jurisdictional accuracy — mainland China / Hong Kong / the U.S. / the EU; one wrong word in a statute citation is an incident
• Currency — statutes get repealed and amended; citing an expired statute is a rookie error
• Adjudication tendency — the same case can lean differently across regions and courts
• Citability — you must be able to cite precisely to a specific statute and a specific judgment number

Together these four are the “legal industry + 4 dimensions,” stacked on top of the universal 9.

The core one-liners of this chapter

Traditional data quality solves “can it be used.” AI-era data quality solves “can the AI learn it right, judge it well, and dare to go to the bedside.”

The 4 dimensions are the entry ticket. The 9 dimensions are the AI-era maturity of a hospital’s data assets.

Next chapter: how high-quality data is formed — the traditional toolkit + the new AI-era meaning + three unsolved new problems.