Quality Assurance and Accuracy

"Ninety-nine per cent accurate" is the most quoted and least useful claim in the industry. Whether it means anything depends entirely on what was counted and who counted it.

The denominator problem

Accuracy percentages are usually computed over characters, because characters are easy to count. A one-thousand-character report with one wrong drug name is 99.9% character-accurate and clinically dangerous. A report with fifteen misplaced commas is less accurate by the same measure and entirely harmless.

Any accuracy figure that is not accompanied by an error classification is therefore close to meaningless. The correct question is not how accurate the work is but how frequently errors of each severity occur.

Classifying errors by consequence

Mature quality programmes classify by clinical and legal impact rather than by error type. The common structure has three tiers:

• Critical. Errors that change clinical meaning or could affect patient safety: a wrong medication, dose or laterality; an omitted or reversed negation; a wrong patient or encounter identifier; an omitted section that changes interpretation. The tolerance for these is effectively zero, and they are counted as events rather than as a rate.
• Major. Errors that change a discrete data point or require the document to be corrected and reissued but do not alter clinical meaning: a misspelled but unambiguous drug name, a wrong date that is obviously wrong, a misassigned heading.
• Minor. Style, punctuation, formatting and preference deviations. Worth tracking as a trend, not worth treating as a defect.

Two properties make this scheme work: it is applied against the original audio rather than against the reviewer's impression of what should have been said, and the boundaries are documented well enough that two reviewers reach the same classification. A scheme that different reviewers apply differently produces numbers that cannot be compared across time or staff.

Sampling

Reviewing everything is affordable in almost no operation and is not the goal. Risk-weighted sampling is:

• Full review during onboarding for new transcriptionists and editors, and for every new dictating clinician, until performance is established.
• Full or elevated review permanently for the highest-consequence document types — operative notes and discharge summaries in most settings.
• Routine random sampling across the rest, sized so that the resulting rates mean something rather than being sized to a convenient percentage.
• Triggered review after any critical error, after a complaint, and after any change to a template, a recognition model or a clinician's dictation setup.

Sampling should be blinded where practical. A reviewer who knows which transcriptionist produced a document, and who works alongside them, is not an independent check.

What changes under speech recognition

Recognition-assisted editing changes the error profile, not the error rate, and this is the single most important thing to understand about modern transcription quality. Typed errors tend to be visibly wrong — misspellings, transpositions, obvious nonsense — and they announce themselves. Recognition errors are grammatical, correctly spelled, contextually plausible words that happen to be the wrong ones.

Those are far harder to catch by reading, because reading is exactly the process that supplies the missing plausibility. An editor skimming a fluent draft will pass errors that a typist would never have produced. The defences are procedural: editing against the audio rather than reading the text, mandatory verification of high-risk elements — medications, doses, laterality, numeric values, negation — regardless of how the draft reads, and quality review that specifically targets those elements rather than sampling uniformly.

Feedback, or the difference between QA and proofreading

A quality function that finds and fixes errors is a proofreading function. A quality function that changes the rate at which errors occur is a quality function. The difference is what happens after the correction:

• Individual feedback and coaching, tracked over time rather than delivered once.
• Lexicon and reference updates when an error reflects a vocabulary gap — new drugs, new procedures, local clinician and place names.
• Template changes when errors cluster around a document structure rather than around a person.
• Dictation feedback to clinicians when errors cluster around a particular dictator, which is frequently the highest-yield intervention and the one most avoided.
• Recognition model tuning when errors cluster around a voice or a vocabulary.

What to ask a provider

1. The error classification scheme, in writing.
2. The sampling design, including what is reviewed at one hundred per cent and why.
3. Critical error events over a defined recent period, not a headline accuracy percentage.
4. Whether review is independent of production and how that independence is maintained.
5. How findings feed back into training, lexicons and templates.
6. Whether you may audit, and on what terms.

The professional body for the field, the Association for Healthcare Documentation Integrity, publishes the standard reference material on documentation quality and error classification, and its framework is the usual basis for the schemes described above.

See also the workflow and provider evaluation.