Cardiology AI scribe 2026: echo reads, cath lab notes, EP procedures, and the specialty vocabulary problem

May 8, 2026 · 8 min read

General-purpose ambient AI scribes are tuned for primary care vocabulary — "diabetes," "hypertension," "follow up in three months." Cardiology vocabulary breaks them. "PISA radius 0.7," "RVOT VTI 18 cm," "LBBB with QRS 158," "ablated PVC focus at posterior tricuspid annulus" — on these phrases the typical AI scribe has 2-3× the error rate it has on a primary-care SOAP note.

For a cardiology practice, that error rate is not cosmetic. Mis-transcribed echo measurements drive treatment decisions. Mis-named ablation targets land in EP procedure reports. The DIY cardiology AI scribe stack in 2026 has to be built on top of the specialty vocabulary — not bolted on after.

Where general scribes break in cardiology

The pattern is clear: numbers, abbreviations, and anatomical terms collide with general English. Whisper and most ambient scribes default to the more common interpretation when in doubt. Cardiology's "more common" interpretation is almost always wrong.

The three cardiology scribe workflows

Each has different latency and accuracy requirements:

The cardiology-aware AI scribe pipeline

1. Specialty-tuned ASR. Whisper-class with a custom language model bias file. Boost weights for: cardiology abbreviations (LBBB, RBBB, PVC, NSVT, V-tach, A-fib, A-flutter, AVNRT, AVRT), anatomical terms (RVOT, LVOT, PISA, RA, LA, RV, LV, IVC, SVC, posterior leaflet, anterior mitral leaflet), and measurement units (mmHg, cm, ms, beats per minute, mL/min). Whisper supports prompt biasing — build a 200-term cardiology vocabulary primer.
2. Numerical extraction pass. A second LLM pass with the structured cardiology schema: ejection fraction, chamber dimensions, valve gradients, PA pressure, etc. Produces structured fields (not freeform text) so the EHR can populate the report template directly.
3. Procedure-note structure. For cath/EP, use a procedure-note schema: indication, access, devices used, anatomical findings, ablation/intervention details, complications, post-procedure plan. The LLM populates each section with cited transcript lines, never freeform.
4. Audit trail. Original audio + transcript + structured output retained per state medical board rules (typically 7-10 years for adult patients).

Vendor matrix — cardiology AI scribes 2026

For a 1-3 cardiologist practice without IDN backing, the DIY stack is cost-competitive (~$60-150/provider/mo all-in) and the specialty primer is a one-time build. For groups of 5+ cardiologists, vendor pricing starts to look reasonable in exchange for the integration savings.

The cardiology Whisper primer (200 terms)

Build a text file with cardiology vocabulary — abbreviations, anatomical terms, drug names, measurement units. Pass it as Whisper's initial_prompt parameter. This biases the language model toward cardiology interpretations on ambiguous tokens.

# cardiology-primer.txt — pass via Whisper initial_prompt
LBBB RBBB LAFB LPFB IVCD QRS QT QTc PR ST T-wave
PVC PAC NSVT VT V-fib A-fib A-flutter AVNRT AVRT WPW
EF LVEF RVEF LVEDV LVESV LVH RVH PISA EROA
RVOT VTI LVOT VTI MAPSE TAPSE TDI E/A E/e prime
mmHg ms bpm cm meters per second mL per minute
RA LA RV LV LAA RAA IVC SVC TR PR MR AR AS PS MS
HFrEF HFpEF HFmrEF NYHA Class I II III IV
DOAC apixaban rivaroxaban dabigatran edoxaban warfarin
amiodarone sotalol dofetilide flecainide propafenone metoprolol
ablation cryoballoon RF catheter mapping CARTO Ensite
CRT-D ICD pacemaker leadless Micra LBBP His-bundle
CrCl GFR eGFR cystatin troponin BNP NT-proBNP
… (extend to 200 terms covering your subspecialty)

The structured cardiology note prompt

You are documenting a cardiology encounter. Extract structured data from the transcript.

INPUT:
- Encounter audio transcript (verbatim, with cardiology vocabulary)
- Patient profile: age, sex, prior cardiac history, current meds

OUTPUT structured note with these sections:
1. Chief complaint and HPI (narrative)
2. Cardiac exam: HR, BP, JVD, murmurs, S3/S4, edema, peripheral pulses
3. Echo / imaging review (if discussed): EF, chamber sizes, valve function, RV function, PA pressure
4. Rhythm (sinus, A-fib, A-flutter, AV block, etc.) with rate and any ectopy
5. Functional status: NYHA class, exercise tolerance, recent decompensation
6. Assessment: structured by problem (CHF / CAD / arrhythmia / valvular / HTN / lipids)
7. Plan: meds, device, procedure, follow-up timing, lab orders
8. Pertinent ICD-10 codes with v28 specificity (e.g., I50.22 not I50.9)

Cite transcript lines for each numerical value and clinical decision. Flag any value in the transcript that seems implausibly out of range — do not auto-correct, surface for clinician review.

The cath / EP procedure note differences

Procedure notes need stricter structure than office visits. The DIY pipeline for procedure notes adds:

• Live transcription with anatomical-term biasing (separate primer for EP electroanatomical terms)
• Device catalog — the LLM has access to a device-name list (Medtronic Arctic Front Advance, Abbott Tactiflex, Biosense Webster CARTO 3, etc.) so it doesn't paraphrase a model name into a generic
• Complications log — auto-extract any phrase matching a complications dictionary (vascular access bleeding, tamponade, AV block, PV stenosis, etc.) and surface to the structured note even if mentioned only briefly
• Time-stamped events — the audio timestamp lets the note correlate to fluoroscopy time and total procedure time

When the DIY stack pays off vs vendor

For a solo cardiologist or 2-3 person group: DIY pays off. The primer is a few hours of work, the LLM cost per encounter is < $1, and the accuracy on specialty vocabulary beats every general scribe in the price range.

For a 5-10 cardiologist group with shared EHR templates: split the difference — use Heidi or Suki with custom cardiology templates, plus a DIY post-pass for the procedure notes.

For an IDN-employed cardiology service line: Ambience or Abridge enterprise integration likely wins on EHR-native workflow even at a cost premium.

What's coming in the next 6 months

Whisper v4-class open weights expected in 2026 H2 with native medical-vocabulary tuning. Once that lands, the DIY cardiology stack becomes 10-30% more accurate without adding cost. The current 6-month strategy: build the primer, run the DIY stack, plan to swap the underlying ASR when the upgrade ships.