Chapter 38 — Key Takeaways

What CDI is

Clinical documentation integrity — making the record describe the patient accurately, completely, and in language the classification systems can read, corrected only by the clinician who owns the statement.

It is not a revenue program · not a second coding department · not an audit. The difference from an audit is timing: CDI reads unfinished work that can still be completed by the person who made it.

A program that only ever moves in one direction is telling you about itself, not about its patients.

Copy-forward — the operational answer

You will not win a ban and should not want one. Attack the fields, not the feature.

Documented negatives

A sentence whose only purpose is to record that something did not happen. It converts silence — which proves nothing — into evidence, which decides a code.

• "No imaging guidance used" → excludes 20611 on the physician's own statement; 20610 stands
• "No aspirate obtained" → injection, not aspiration-and-injection
• Silence is not a documented negative. It is an absence, and an absence has to be asked about
• An auditor notices absent negatives (Ch. 37 §37.11)
• Install the three or four that matter in your specialty as required fields — never as default values, because a templated negative is a configuration making an assertion nobody chose

The concurrent review

Concurrent = while the encounter is still open. A clarification on day 2 is a progress note; the same clarification after discharge is an addendum dated when made (Ch. 4 §4.5).

• Working DRG = what the record would group to right now, mid-stay. A review tool, never a bill.
• Reconciliation between working and final coded DRG needs a written procedure and an adjudicator who is neither reviewer
• Worklist legs: severity-split records without a documented CC/MCC · known-ambiguity conditions · clinical indicator without a diagnosis · a random sample, which keeps the other three honest
• Outside the hospital: pre-visit review is care; post-visit harvesting is a different activity (Ch. 36 §36.8)

The compliant query — seven elements

  1  IDENTIFICATION  patient, encounter, date, who asks, how to respond
  2  THE EVIDENCE    the clinical indicators FROM THIS RECORD, and only those
  3  THE QUESTION    open; about a clinical fact; never about a code or a dollar
  4  THE OPTIONS     clinically reasonable, INCLUDING "clinically undetermined,"
                     "other," and at least one that adds nothing
  5  THE DISCLAIMER  no particular response is intended or preferred
  6  THE LANDING     the response goes in the MEDICAL RECORD, not the form
  7  THE RECORD      query + response retained per written policy, retrievable

Missing element 4 = leading. And an option list can be leading with five options if every reasonable answer adds something. The test: is the record's most likely answer — including "the assessment as written is what I intend" — on the list with equal weight?

Yes/no: permitted narrowly, to confirm or specify a diagnosis already in the record. Never to introduce one. Does a "yes" tell you something new about the record, or about the patient?

Verbal: legitimate — and documented. Who asked, whom, when, what was asked, what was answered.

The five markers of a leading query

  1  SUPPLIES THE ANSWER          "Can you document acute blood loss anemia?"
  2  NAMES THE CODE/CLASS         "...so it groups correctly" / "...a level 4"
  3  STATES OR IMPLIES THE MONEY  "...so the account isn't underpaid"
  4  NO REAL ALTERNATIVE          one option; or five that all add something
  5  SIGNALS THE PREFERENCE       "Please confirm..." / re-querying after a "no"

The exposure is not a coding error. It is manufacturing documentation to support a code — a false-record theory (Ch. 5 §5.1), extrapolated more cleanly the more consistent it is (Ch. 37 §37.6). The claim is the organization's. The query has an individual's name on it.

Retention — two models, three rules

Model 1: part of the legal health record. Model 2: a separate business record. Pick one, write it down, apply it consistently. Then, under either:

1. The response must land in the medical record. A code is never assigned from a query form
2. Every query and response must be retrievable — "show me the queries behind the shift" (Ch. 33 §33.6)
3. Retention runs at least as long as the audit look-back

The two metrics that corrupt

The second is the important one: a program can become a leading program without a single non-compliant query. The selection is the leading act and no review of query text detects it. Ch. 6 §6.3 predicted it: a program that lets "will this change the DRG?" drive "is this true?" has become the leading-query problem at institutional scale.

Defensible: prioritizing which records get reviewed by materiality. Not defensible: deciding whether to send a clinically warranted query based on what — or which direction — the answer would be worth.

Publish adds and removals as separate lines, "clinically undetermined" as its own line, and never let compensation depend on the direction of a finding.

The engine

CAC pipeline: ingest → segment → extract → qualify → map → score → present with the source text. That last step is the most valuable and least discussed output: an audit trail no manual process ever had.

Four qualifiers, where nearly all NLP errors live:

Section detection makes an engine MORE section-sensitive than a human, not less — which turns Ch. 22 §22.6's right-fact-wrong-section finding into a systematic miss.

An engine cannot flag the absence of a sentence nobody wrote. It can flag a clinical-indicator pattern with no corresponding diagnosis — which is how CDI worklists are built, and which carries the selection problem into a rule set if every rule fires one way.

Put the engine on the assertion register (Ch. 37 §37.10): what it asserts · where it lands (a claim is correctable; a note over a signature is not) · owner by name · evidence test run against a transmitted claim, not a setup screen.

Autonomy — the four-property test

  1  STRUCTURE     template-produced source document, code-deciding facts
                   in named fields rather than in prose?
  2  CODE SPACE    tens of candidate codes, not thousands?
  3  VARIABILITY   two clinicians produce substantially the same document?
  4  FEEDBACK      fast, unambiguous signal when wrong - days, not an audit?

  Four yeses: probably.  Three: assisted.  Two or fewer: it's a suggestion tool.

Passes: radiology, pathology, screening and routine laboratory. Fails outright: E/M leveling, surgical coding from an operative report, and modifier 25 — an argument, not a pattern.

Property 4 is the quiet one: denial feedback teaches a system to avoid the errors that get caught, which is not the same as being right (Ch. 28 §28.8).

Nobody's software signs anything. Scope · thresholds · sampling · stop conditions · the register entry · the October 1 / January 1 / quarterly calendar — all the organization's, in writing. This book names no date, for any of it.

Coder-in-the-loop

Automation bias — accepting a plausible suggestion and stopping looking. The engine's accuracy causes it.

Three design decisions decide whether the loop is real:

1. What the reviewer sees first — the note, or the answer (an answer first is an anchor)
2. Whether disagreement is cheap — if rejecting costs six clicks, the agreement rate measures the interface
3. Whether the productivity standard was reset for review work — the loop is a staffing decision before it is a technology decision

Audit the ACCEPTS, not the rejects. Score against the record, never against the engine.

Precision, recall, and the asymmetry

Precision = supported ÷ suggested. A failure is a false positive. Recall = found ÷ supported. A failure is a false negative.

   suggested 1,905 · supported 1,840 · both 1,712
     PRECISION 1,712/1,905 = 89.9%     RECALL 1,712/1,840 = 93.0%

   after the loop:  1,808 supported on claims + 22 unsupported = 1,830
     ACCURACY 98.8%     RECALL 98.3%     [constructed]

22 is the smallest number in the table and the only one with a False Claims Act attached to it.

Both errors are real; neither is safe — downcoding is not the conservative option (Ch. 5 §5.8). But a missed code costs revenue and truth; an unsupported code costs revenue, truth, and legal exposure that compounds with volume and consistency. So: review effort on the accept side · the confidence threshold is a compliance decision with a name on it · and an organization that checks only accepts will let recall rot silently, because nothing about a missed code ever announces itself.

Model drift — three drivers: the code sets move on a calendar · the documentation changes · and the definition of "supported" changes while every note reads the same (Ch. 28's CS1 mechanism: correct when written, decayed, no internal event to signal it).

What the human is for

1. The judgment that is an argument, not a pattern — modifier 25, on four elements, three of which have nothing to do with the knee (Ch. 14 §14.4)
2. The rules that are not in the note — payer policy, the edit, the contract, the calendar
3. The absence — the sentence nobody wrote (Ch. 37 §37.11)
4. The accountability — a certification requires a certifier, and "confidence 0.93" is not a path
5. The query — only a person may ask a clinician a question about a patient

Monday morning

You should be able to: state what CDI is and is not in one sentence each; write a copy-forward policy and name the three fields you would make uncopyable in your own specialty; explain a documented negative to a physician in thirty seconds using a real clause from a real note; draft a compliant query and check it against seven elements; spot all five markers of a leading one; explain to a director why agreement rate must never be a target; read an engine's suggestion panel and dispose of every line with a reason; apply the four-property autonomy test to your own work; compute precision and recall and say which error carries which exposure; and answer — with a list rather than a sentiment — the question someone will eventually ask you at a family dinner about whether a machine is going to do your job.