Glossary of clinical documentation AI terms

Software that converts spoken audio into text.

Automatic speech recognition converts an audio signal into a text transcript. It produces words, not meaning: an ASR engine returns what was said, in order, optionally with timings, speaker labels and confidence values. It does not decide what belongs in a clinical note.

Why it matters: Confusing ASR with clinical documentation is the most expensive mistake in this category. An ASR API is roughly 20% of the work of a documentation feature; the structuring, template adherence, clinical review workflow and chart write-back are the other 80%.

Medical speech-to-text API guideAI speech to text in healthcareMedical speech-to-text rankings

The standard accuracy metric for transcription: substitutions plus insertions plus deletions, over reference words.

Word error rate is the edit distance between a produced transcript and a human reference transcript, divided by the number of reference words. A 5% WER means one word in twenty differs from the reference.

Why it matters: WER is only comparable when the corpus, audio conditions, microphone setup and reference-transcript protocol are identical. A vendor WER claim with no published methodology tells you nothing, and WER says nothing at all about whether the resulting note is clinically correct — a transcript can be 98% accurate and still drop the one number that mattered.

How we score accuracyEvaluation criteria frameworkVendor Transparency Index

Determining who spoke which words in multi-speaker audio.

Diarization segments audio by speaker and assigns labels, so a transcript reads as a conversation rather than a single stream. In clinical audio it must also survive interruptions, overlapping speech, a third party in the room and a clinician who steps away from the microphone.

Why it matters: Without reliable diarization, a scribe can attribute a patient's speculation to the clinician's assessment. Attribution errors are the most clinically dangerous class of documentation error in ambient capture.

AI scribe API guideMedical speech-to-text API guide

Whether audio is transcribed live as it arrives or submitted as a completed file.

Streaming transcription returns partial results over a persistent connection while audio is still being captured. Batch transcription accepts a finished recording and returns a result asynchronously, usually more cheaply and often slightly more accurately because the model sees the whole utterance.

Why it matters: This one choice determines your architecture. Interactive experiences — live captions, in-visit prompts — require streaming. Asynchronous note drafting does not, and paying for streaming you never use is a common waste.

Medical speech-to-text API guideCost calculator

The delay between speech and usable output.

For streaming, the interval between an utterance and its transcript appearing, typically measured to first partial and to finalised text. For documentation, the wall-clock time from encounter end to a reviewable note.

Why it matters: The two are frequently conflated in vendor material. Decide which one your product actually needs before comparing numbers.

Evaluation criteria frameworkMedical speech-to-text API guide

A system that listens to a clinical encounter and drafts the note from the conversation.

An ambient scribe captures the natural conversation between clinician and patient — with consent — and produces a structured draft note without dictation. The clinician reviews, edits and signs. 'Ambient' refers to the capture mode: nobody speaks to the software, they speak to each other.

Why it matters: Ambient capture is what makes recording-consent law a product requirement rather than a legal footnote, and it sets a higher bar for diarization and hallucination control than dictation does.

AI scribe API guideRecording consent by stateAI scribe API rankings

An API that takes encounter audio and returns a structured clinical note, not just a transcript.

A scribe API absorbs the whole documentation pipeline behind one interface: transcription, speaker attribution, clinical summarisation, template adherence and structured output ready for a chart. The integrator supplies audio and context and receives a note.

Why it matters: For a software team, the scribe-API boundary is the buy/build line. Above it you own product; below it you own an ML and clinical-safety programme.

AI scribe API guideBuild vs buy frameworkClinical documentation API guide

The standard clinical note structure: Subjective, Objective, Assessment, Plan.

SOAP organises a clinical encounter into what the patient reports (Subjective), what was measured or observed (Objective), the clinician's conclusion (Assessment) and the intended actions (Plan). Most generated notes target SOAP or a specialty variant of it.

Why it matters: Whether you can define the note template — and how tightly a vendor adheres to yours — is a product-differentiation question, not a formatting preference.

Clinical documentation API guideAI scribe API guide

Content in a generated note that was never present in the encounter.

A fabricated finding, measurement, medication or negative that the model produced from prior expectation rather than from the audio. In clinical documentation the dangerous form is a plausible normal — 'lungs clear to auscultation' in an encounter where the chest was never examined.

Why it matters: Word-level accuracy metrics do not detect it. Evaluating a documentation vendor means specifically testing for content that appears without a source in the transcript.

AI scribe API guideEvaluation criteria frameworkClinician review requirements

A clinician speaking a note directly to software, rather than the software listening to a conversation.

In dictation, the clinician is the only speaker and is addressing the system deliberately, often with commands and formatting instructions. It is the older workflow and remains the fallback when ambient capture is not consented to or not practical.

Why it matters: Dictation is far easier technically — single speaker, intentional speech, no consent complexity — and remains the right answer for procedural and inpatient workflows.

AI speech to text in healthcareMedical speech-to-text API guide

The modern healthcare data exchange standard, built on REST and JSON resources.

Fast Healthcare Interoperability Resources defines discrete resources — Patient, Encounter, DocumentReference, Composition, Condition, MedicationRequest — exchanged over a RESTful API. It is how a modern integration reads context from and writes results back to an EHR.

Why it matters: A documentation vendor that emits FHIR-shaped output removes weeks of mapping work. One that emits prose only leaves the hardest half of the integration with you.

EHR integration guideEHR write-back matrix

The FHIR resource that carries a clinical document, such as a generated note, into a chart.

DocumentReference wraps a document with its metadata — type, author, subject, encounter, status and the content itself or a pointer to it. It is the usual destination for a finished note in a FHIR-based write-back.

Why it matters: Knowing your target resource before you choose a vendor tells you exactly how much transformation the vendor's output needs.

EHR integration guideEHR write-back matrix

The standard way to launch a third-party app inside an EHR with authorised context.

SMART App Launch layers OAuth 2.0 authorisation and launch context onto FHIR, so an external application can open inside the EHR already knowing the patient, encounter and user, with scoped permissions.

Why it matters: If your documentation feature must appear inside the clinician's EHR session rather than your own UI, SMART is the route — and it constrains which vendors can help you.

EHR integration guideEvaluation criteria framework

Machine-readable clinical data alongside or instead of narrative note text.

Discrete fields — problems, medications, orders, vitals, coding candidates — emitted as data rather than prose, ideally mapped to FHIR resources or a documented schema.

Why it matters: Prose can only be pasted into a chart. Structured output can populate one, and it is what makes downstream features like coding support or order suggestion possible.

Clinical documentation API guideEHR integration guide

The system of record for a patient's clinical chart.

The clinical system where notes, problems, orders, results and medications live. For a documentation vendor it is both the destination for output and, frequently, the gatekeeper of the integration.

Why it matters: If you are an EHR, documentation is a feature you must own. If you integrate with EHRs, an EHR-agnostic vendor keeps your options open.

EHR integration guideEHR write-back matrix

Individually identifiable health information regulated under HIPAA.

PHI is health information that identifies an individual, or could reasonably be used to, held or transmitted by a covered entity or business associate. Encounter audio is PHI. So is a transcript, and so is a generated note.

Why it matters: Every architectural choice — where audio is processed, how long it is retained, whether it can train a model — is a PHI-handling decision that lands in your customer's security review.

HIPAA-compliant speech-to-text guideCompliance checklist

The contract that lets a vendor process PHI on a covered entity's behalf.

A business associate agreement binds a vendor to HIPAA obligations around safeguards, permitted use, subcontractors, breach notification and return or destruction of PHI. Without one, sending PHI to that vendor is a violation regardless of how secure the vendor is.

Why it matters: For an integrator the chain matters as much as the document: your customer signs with you, and you must have a BAA with every downstream vendor touching the audio.

HIPAA-compliant speech-to-text guideCompliance checklistVendor Transparency Index

An audit report on whether a vendor's controls operated effectively over a period of time.

SOC 2 Type II reports an independent auditor's opinion on the design and operating effectiveness of controls against the trust services criteria across a review window, typically six to twelve months. Type I covers design at a point in time only.

Why it matters: Enterprise security reviews ask for the report, not the badge. Read the scope section: a report that excludes the service you are buying is worth nothing to you.

HIPAA-compliant speech-to-text guideHow we score compliance

Whether a vendor may use your customers' clinical data to improve its models.

A contractual position, not a technical one: whether audio, transcripts or notes derived from your customers' encounters can be used to train or fine-tune the vendor's models, with what de-identification, and whether you can opt out.

Why it matters: This is the single most common blocking question in health-system security review, and the answer must be in the contract. A marketing page saying 'we respect your data' is not an answer.

HIPAA-compliant speech-to-text guideRFP templateVendor Transparency Index

A legal regime requiring everyone in a conversation to consent to it being recorded.

In an all-party consent jurisdiction, recording a conversation is lawful only if every participant consents. In a one-party jurisdiction, one participant's consent — potentially the clinician's — is sufficient. Healthcare-specific rules and professional obligations can be stricter than the general wiretap statute.

Why it matters: Ambient capture is recording. If your product operates across state lines, consent capture must be a feature with an audit trail, not a line in a training deck.

Recording consent index by stateCompliance checklist

Removing identifiers so information is no longer PHI under HIPAA.

HIPAA recognises two routes: Safe Harbor removal of eighteen identifier categories, or expert determination that re-identification risk is very small. Redaction of a transcript is not automatically de-identification.

Why it matters: Vendors frequently justify model training on 'de-identified' data. Ask which route, who determined it, and whether audio — which carries voice as a biometric — is in scope.

HIPAA-compliant speech-to-text guideRFP template

Software that offers clinical recommendations rather than only recording what happened.

CDS provides assessments or recommendations intended to inform clinical decisions. Documentation tools sit adjacent to it, and a feature that starts suggesting diagnoses or treatment can cross into regulated territory.

Why it matters: The line matters for regulatory exposure. Documenting what a clinician said is not CDS; proposing what they should do next may be.

Clinical documentation API guideCompliance checklist

Reselling or embedding a vendor's capability under your own brand.

White-label terms permit the documentation capability to appear as a feature of your product, with no vendor branding surfaced to the end user, and with you owning the customer relationship, support and pricing.

Why it matters: Without white-label rights, you are distributing a partner's brand inside your product and inviting them into your customer relationship.

AI scribe API guideCommercial terms scoringRFP template

Usage-based pricing charged against minutes of processed audio.

A metered model in which cost scales with audio duration rather than seats. Per-encounter and per-clinician-per-month models are the common alternatives; each pushes different risk onto the buyer.

Why it matters: Per-minute is the only model that lets you compute a marginal cost per encounter and therefore price your own feature with a known margin.

Cost calculatorBuild vs buy framework

The cost of moving off a vendor once you depend on it.

In this category lock-in shows up as proprietary output formats, vendor-owned templates, non-portable prompt or configuration logic, minimum commitments, and end-user brand exposure that makes replacement visible to your customers.

Why it matters: Documentation is a core workflow. Assume you will change vendors at least once and evaluate the exit before signing.

RFP templateBuild vs buy framework