TL;DR: Modern healthcare-trained AI extraction on cardiology faxes typically reaches the mid-90s on field-level accuracy for clean, first-generation documents. Document-level accuracy — every field on a page correct — runs meaningfully lower, because errors compound across fields. Both numbers degrade on fax-of-a-fax copies, handwritten annotations from referring offices, and device reports from Medtronic, Boston Scientific, and Abbott that each format differently. A cardiology fax triage with OCR extraction tool is only worth buying if it publishes both metrics and routes low-confidence documents to a human instead of filing them silently.
Four different things get called "accuracy"
When a vendor says "99% accurate," ask which of these four they mean, because they are not close to the same number.
Character accuracy is the percentage of individual characters transcribed correctly. It's the highest number a vendor can quote and the least useful one you can act on. A device report can hit 99% character accuracy and still put the wrong lead impedance value in the chart.
Field accuracy is the percentage of individual data points pulled correctly — patient DOB, referring provider NPI, device model, battery longevity, the ordering physician's name. This is the number that maps to real work.
Document accuracy — sometimes called the straight-through rate — is the percentage of documents where every field came out right, so no human had to touch it. One wrong field kills the whole document.
Classification accuracy is whether the system correctly labeled the fax as a referral versus a device report versus a records request. It's usually the highest of the four and it gets cited constantly, because sorting is a much easier problem than reading.
Here's why the gap between field and document accuracy matters. Take a device report with 12 fields you care about and 95% field accuracy. If those errors were independent, document accuracy would land near 54%. Real errors cluster — a smudged page ruins several fields at once — so the true number sits well above that floor. But it is nowhere near 95%, and no arithmetic gets it there.
Why cardiology faxes break traditional OCR
Cardiology has a worse document problem than most specialties, for four specific reasons.
Fax-of-a-fax degradation. The ITU-T T.4 Group 3 fax standard tops out around 204 × 196 dpi in fine mode. OCR engines are built for 300 dpi. Every retransmission re-quantizes an already-marginal image, and a referral that's been faxed from the PCP to a hospital to your office is a third-generation copy. Character strokes bleed together. Decimal points vanish. A "3" becomes an "8."
Device vendor report layouts. Medtronic CareLink, Boston Scientific LATITUDE, and Abbott Merlin each print interrogation summaries in their own layout, with their own labels for the same clinical concept. One prints "Battery Status," another prints "Estimated Longevity." Column positions differ. A tool trained on one manufacturer's output does not transfer cleanly to another's, which is exactly why a single blended accuracy number across "device reports" hides so much.
Handwritten annotations. Referring offices write on faxes. A circled date, "pt needs echo first" in the margin, a callback number scrawled at the top. These are often the most operationally important marks on the page, and they're the hardest thing on it to read.
Stamps, signatures, and overlays. A "RECEIVED" stamp landing across the DOB field, or a signature crossing the provider name line, breaks character segmentation in ways that clean-document benchmarks never surface.
So what accuracy should you actually expect?
Published clinical research is more sober than vendor marketing, and it's worth anchoring to.
A 2024 JAMIA Open study built an OCR-plus-generative-AI pipeline for scanned outside medical records and reported F1 scores of 0.95 for document segmentation and 0.96 for classification — but 0.90 for date extraction. Date is one field, in a controlled study, on scanned PDFs rather than degraded faxes, and it still came in as the weakest result. That gap between "sorting the document" and "reading a specific value out of it" is the whole story.
Older clinical OCR work is a useful floor. A real-world clinical OCR evaluation measuring vitals capture from an encounter form reported a digit recognition rate of 92.4% (95% CI 91.6–93.2) — on a form designed to be machine-read.
Mayo Clinic researchers have specifically studied OCR and NLP applied to digitally faxed outside medical records for patients transferring into cardiovascular medicine, which tells you this is a recognized problem in cardiology and not a vendor invention.
Directionally, on a real cardiology queue:
- Clean, first-generation typed referrals: field-level accuracy in the mid-90s is realistic for a healthcare-trained model.
- Device manufacturer reports, first generation: high-80s to mid-90s, depending entirely on whether the tool has been trained on that manufacturer's layout.
- Second- or third-generation fax copies: measurably worse, and on the affected fields it can fall into the 70s or 80s.
- Handwritten annotations: the lowest of any category, with a wide enough range that a single number is misleading.
Anyone quoting one accuracy figure for "cardiology faxes" without splitting it by document class and fax generation is quoting a blended average that describes their sample, not your queue.
How vendors pick the flattering metric
Five tells, all of them common:
- A single blended number. A vendor reporting one accuracy figure is reporting the one that flatters them. Ask for all four, broken out.
- Classification presented as extraction. "We're 97% accurate on cardiology faxes" often means the system correctly labeled the document type 97% of the time. That's a different, easier task.
- Benchmarks on vendor-supplied samples. Their sample documents are clean because their sample documents were chosen. Testing on them reproduces their result and tells you nothing.
- Accuracy measured after human review. Post-correction accuracy approaches 100% by definition. It's the pre-review number that determines your staffing.
- Abstention that doesn't count against the score. If the system declines to process 15% of documents and reports accuracy only on the remaining 85%, the headline number is inflated by exactly the documents you most needed help with.
The counter-question is short: field-level and document-level accuracy, on my documents, broken out by document type, with the exception rate included in the denominator. A vendor who can't answer that in a pilot either hasn't measured it or doesn't want to.
How to pilot on your own fax queue
Vendor demos are theater. A real pilot takes about two weeks and settles the question.
Pull 300–500 consecutive faxes from your own queue. Consecutive matters more than volume. The moment someone curates the sample, you've rebuilt the vendor's benchmark. Include the third-generation copies, the 40-page records dumps, and the ones with coffee on them.
Stratify and label by class: new referral, device interrogation report by manufacturer, records request, prior auth response, lab and imaging results. You'll want per-class accuracy, and you can't get it without this step.
Define 8–15 fields that actually matter to your workflow. Not every field on the page — the ones that determine routing, scheduling, and billing.
Build ground truth by hand. This is the expensive part. Budget 8–15 hours of a staff member's time to key the correct values for your sample. It's also the only way to get an honest denominator.
Score field-level and document-level separately, per class. Then score one more thing most pilots skip: confidence calibration. When the tool reports 98% confidence, is it right 98% of the time? A well-calibrated 91% system beats a poorly-calibrated 96% one on every operational measure that matters.
What confidence scoring and exception routing should look like
The property you're buying isn't accuracy. It's the system knowing when it's wrong.
A tool at 91% field accuracy that reliably flags its own uncertain 9% is safer than a tool at 96% that files everything silently — because the second one drops 4% of errors into charts with no trail and no one looking. In cardiology, a misread battery-longevity value or a transposed MRN on a device report isn't a data-quality footnote.
What good looks like in practice:
- Per-field confidence, not per-document. A page can have a perfectly legible MRN and an illegible referring provider. One score for the page throws that information away.
- Thresholds your team sets, per field and per document type. MRN and DOB should be strict. "Reason for referral" can be looser.
- An exception queue that shows the document image beside the extracted values, so a staff member resolves it in seconds instead of reopening the fax.
- Corrections that feed back, so the same layout doesn't generate the same exception every week.
- A full audit trail of what was auto-filed, what was flagged, who accepted it, and when.
Honey Health's Fax Triage agent is built on that shape — per-field confidence scores, thresholds your team tunes by document type, and an exception queue for anything that falls below the line. We'd rather tell you a document is uncertain than tell you we're 99% accurate.
How much manual review should you plan to staff?
Plan for 10–25% of documents touching a human in the first 90 days, trending down as the system sees more of your specific referring offices and device layouts. A mixed cardiology queue with a heavy device-report component sits at the higher end of that range early on.
For context on what you're improving against: the 2023 HRS/EHRA/APHRS/LAHRS expert consensus statement on remote device clinic management estimates 3.0 FTEs are needed to support 1,000 CIED patients, with combined remote-monitoring and in-person follow-up workload ranging from 2.1 hours per patient per year for a pacemaker to 9.3 hours per patient per year for an implantable loop recorder. That's the baseline. Extraction automation takes a bite out of it — it does not zero it out.
If a vendor promises a sub-5% exception rate in month one on a mixed cardiology fax queue, they're describing a state you might reach in year two, on a tuned system, with a stable set of referring sources. Staff for the first number and be pleasantly surprised by the second.
Frequently asked questions
What is a realistic OCR accuracy rate for cardiology faxes?
Field-level accuracy in the mid-90s is realistic for clean, first-generation typed documents processed by a healthcare-trained AI model. Device manufacturer reports and multi-generation fax copies run lower. Document-level accuracy — every field correct on a page — is always meaningfully below the field-level number, and any vendor quoting one figure for both is blending them.
Why is document-level accuracy lower than field-level accuracy?
Because errors compound. A document with 12 fields at 95% field accuracy can't reach 95% document accuracy — every field has to land correctly for the document to count. Real errors cluster on bad pages rather than scattering randomly, which helps, but the document-level number still sits well below the headline field-level figure.
Can AI read handwritten notes on a cardiology fax?
Sometimes, and unreliably. Handwritten annotations from referring offices are the lowest-accuracy category on any fax, and accuracy varies enormously by handwriting. A well-built system should recognize that a handwritten region exists, extract what it can, flag it as low confidence, and route it for human review rather than guess.
Do device reports from different manufacturers extract differently?
Yes, substantially. Medtronic, Boston Scientific, and Abbott each print interrogation summaries in distinct layouts with different labels for the same values. A tool trained heavily on one manufacturer's format performs worse on the others. Ask any vendor for accuracy broken out by device manufacturer, not averaged across all device reports.
How many faxes should I test before trusting the numbers?
Three hundred to 500 consecutive faxes pulled from your own live queue is enough to get stable per-class numbers. Consecutive is the key word — the sample has to include the degraded copies and the odd document types. Curated samples reproduce the vendor's benchmark instead of predicting your results.
Should low-confidence extractions ever auto-file?
No. Any extraction below your confidence threshold should land in an exception queue with the source image attached, not in the chart. Silent auto-filing of uncertain data is the failure mode that made operators distrust legacy OCR in the first place, and it's worse than a slightly higher exception rate.
