How to audit a teleradiology QA program

Audit a teleradiology QA program with four numbers: double read rate, written discrepancy definition, measured rate, and the workflow after a miss.

Published 15 November 2026

"What's your discrepancy rate?" Every buyer puts some version of that question to every teleradiology vendor, and every vendor has a number ready. The number alone settles nothing. Auditing a teleradiology QA program takes four disclosures: the share of reports independently double read, the written definition of a major discrepancy, the measured rate against that definition over a stated window, and the documented workflow after a miss. A vendor that produces all four is measuring quality. A vendor that offers one polished percentage is measuring complaint volume.

If you're still assembling the shortlist, start with the teleradiology buyer's guide. This page is its quality chapter, written as an audit you can run on paper.

Four numbers audit any teleradiology QA program

Four disclosures make a quality claim checkable: the sampling rate, the definition, the rate with its denominator, and what happens after a miss. Asking for them is not a buyer's invention. The ACR White Paper on Teleradiology Practice states that a teleradiology provider should regularly participate in an established quality assurance program including formal peer review. The white paper also sets out what such a program should do: address physician education and error reduction, enable longitudinal follow-up, provide a second opinion when local caregivers raise concern, and include a process of remediation for low-performing radiologists. Those four disclosures are how you check whether the program described to you is that program.

# The number A strong answer looks like A weak answer looks like
1 Double-read sampling rate A fixed percentage of all reads, selected at random, applied continuously "Difficult cases", "as needed", "when a clinician raises a question"
2 Definition of a major discrepancy A written definition tied to change in patient management, with the scoring scale named No definition offered, or "significant errors"
3 Measured major discrepancy rate A rate with its denominator stated and a trailing 12 month window A single accuracy percentage with no denominator
4 Post-discrepancy workflow A documented sequence with owners and time targets "We contact the ordering physician if needed"

Number three deserves the longest look, because it's where audits go wrong. Two vendors can report wildly different discrepancy rates while performing identically, since the rate is a fraction and the denominator is a choice. One vendor divides its misses by a random sample of everything it signed. The other divides by the handful of reports a clinician already complained about. The complaint denominator is small, self selected, and flattering, because it counts only the misses somebody downstream happened to notice, and the misses nobody noticed are precisely the ones a QA program exists to find. That second number will always look better, and it will always mean less. So push on the fraction itself: which reports, against which definition, over which period. A vendor that tracks this answers in a sentence, because the report already exists internally. A vendor that improvises will restate the headline percentage, and the restatement is itself an answer.

Case selection moves the measured rate by an order of magnitude

How cases enter peer review changes the resulting rate more than anything the radiologists do, and vendors are rarely asked to specify it. Random prospective sampling of live volume, reviewed blind, measures the true miss rate across ordinary work; it's the only figure you can compare between vendors. Unblinded sampling measures reviewer agreement, a softer thing. Complaint-triggered review measures only what a clinician already caught. Peer learning conferences without scoring produce education and no rate at all.

The published literature shows how large the effect of case selection is. A 2025 Academic Radiology meta-analysis of 616,918 on-call preliminary reports found pooled major discrepancy rates of 0.38% for ultrasound, 0.64% for radiographs, 1.35% for CT and 1.86% for MRI. A JACR systematic review and meta-analysis of 12,676 imaging secondary interpretations, covering studies that had been sent out for reinterpretation, found a 20.4% major discrepancy rate. Both measure something called a major discrepancy rate. They differ by more than a factor of ten because the cases entering the second one were selected for a reason, which is exactly the effect you are trying to see through when a vendor quotes you a single number.

Blinding is the detail most often skipped, and it's cheap to verify. Ask one question: does the second radiologist form an impression before seeing the original report and the original reader's name? A yes makes the two reads independent. Anything else anchors the reviewer on the first answer, and concurrence climbs for reasons unconnected to the images.

Most structured programs score on a named scale. The ACR RADPEER program scores on three points, where 1 is concurrence, 2 is a discrepancy not ordinarily expected to be made, and 3 is a discrepancy that should be made most of the time, with a modifier on the latter two for whether the difference is likely to be clinically significant. ACR also states that no scoring benchmarks have been established, so there is no published target rate to hold a vendor's number against, which is another reason the definition and the denominator carry the weight. The vendor doesn't need that exact scale. It needs to name whatever scale it uses and apply it the same way every month.

One last question closes the loop: is sampling applied continuously to live volume, or run as a periodic batch? Continuous sampling catches a staffing problem in the month it appears. A quarterly batch finds it a quarter late.

What happens after a radiologist misses something on a read?

A confirmed discrepancy should trigger correction, notification, feedback to the reader, and aggregation, each with an owner and a time target. Correction means an addendum on the final signed report, timestamped, with the original text preserved so the audit trail survives; if your workflow mixes preliminary and final reads, confirm how a correction propagates between them, which our guide to preliminary versus final reads covers. Notification means a human calls the ordering clinician inside a defined window and documents the contact in the record. Ask for that window in minutes when a finding changes management, and ask who gets called when the ordering clinician is off shift. Feedback means the radiologist who signed the original report sees the review; it sounds obvious, it's frequently missing where reads scatter across a large contractor pool, and without it a program counts misses without reducing them. Aggregation means discrepancies are trended by radiologist, modality, subspecialty, and shift, so the trend can drive assignment decisions. A program that only produces one company-wide percentage can't tell you whether the problem lives in one section or one overnight window.

Ask for the workflow document itself. It runs about a page, and a vendor that has one can send it the same day.

Make the vendor quantify subspecialty match

Subspecialty match becomes a quality control the moment it's expressed as a percentage of your studies. Nearly every vendor uses the word subspecialist; few will state the share of your studies read by the matching section, which is the only version of the claim that touches your patients. Ask three things: the percentage of studies in each of your modalities read by a section-matched subspecialist over the trailing 12 months; what that percentage looks like overnight and on weekends, when matched coverage is hardest to staff; and whether the discrepancy rate is broken out by matched versus unmatched reads. The third answer is the honest test. If the routing works, the vendor's own data will show it, and the vendor should be glad to show you.

AstraRad routes each study to a board-certified, fellowship-trained subspecialist, every one physically located in the United States and licensed in the state where your patients are located. Shift coverage runs 24/7/365 on scheduled US shifts, which is why a child's abdominal CT arriving at 3 a.m. reaches an awake pediatric radiologist. It is also what supports the STAT commitment on STAT CT reads, under 4 hours as standard with 2-hour and 1-hour STAT on request, measured from last-image arrival to radiologist signature. Coverage detail and SLA tiers live on the SLA page.

Send one disclosure request to every shortlisted vendor

Put the whole audit in a single written request, sent to every shortlisted vendor on the same day, with a five business day deadline and a mutual NDA offered up front so confidentiality can't become a reason to delay.

Please provide, in writing: (1) the percentage of reports independently double read, and whether those cases are selected at random or by another method; (2) whether the second read is blinded to the original report and reader; (3) your written definition of a major discrepancy and the scoring scale used; (4) your measured major discrepancy rate for the trailing 12 months, stating the denominator; (5) the documented workflow that follows a discrepancy, including who notifies the ordering clinician, within what time window, and how the finding reaches the radiologist who issued the original report; (6) the percentage of studies in our modality mix that would be read by a section-matched subspecialist, including overnight and weekend hours.

Six questions, one page of answers. Pair it with the turnaround disclosures in our guide to verifying vendor claims, and score the report output itself with the 12-point report scorecard. Together the three cover whether the finding was right, whether the report was usable, and whether it arrived in time.

Every teleradiology QA program answers in one of three patterns

The replies you get will be documented, unfalsifiable, or complaint driven. The documented reply attaches a fixed sampling rate, a written definition, a rate with denominator and period, the workflow document, and match percentages by modality; that program exists and is measured, and you can now compare vendors on one axis. The unfalsifiable reply is a single high accuracy percentage with references offered in place of method. It proves nothing in either direction, so push once more, in writing, specifically for the definition and the denominator, under NDA if that helps; vendors that track quality internally usually produce the detail once they see the headline number won't close the question. The complaint-driven reply describes peer review as happening "as needed" or "on flagged cases", with notification "if clinically relevant". There the rate is unknowable by construction, because the misses nobody caught downstream never enter the count.

One caveat before you score anyone harshly: if your own radiologists overread every outside study the next morning, that overread already functions as a 100% audit, and vendor QA depth should weigh less in your decision than turnaround and report usability.

Our own answers to the six questions are standing paperwork. That is what a teleradiology QA program looks like when it is written down: 1 in 20 reports independently double read, drawn at random from live volume; a written major discrepancy definition; and every major discrepancy reviewed and closed with the reader. Book a coverage consultation and you'll have the full set in writing. A rate without a denominator is an adjective. Ask for the fraction.

Questions, answered

Frequently asked questions

What is a normal major discrepancy rate for radiology reads?

There's no single industry number, because the rate depends on how a major discrepancy is defined and which reports sit in the denominator. A rate computed over a random sample of all signed reports is a measurement; a rate computed over the reports that drew a complaint is a smaller, friendlier number wearing the same name. AstraRad measures major discrepancies against a written definition on independently double-read studies drawn at random. Get every vendor's definition and denominator in writing before you compare one number to another.

What percentage of teleradiology reports should be double read?

Look for a fixed, prospective sampling rate. AstraRad independently double reads 1 in 20 reports, applied continuously to live volume. A vendor that double reads 'difficult cases' or 'as needed' has no sampling rate at all, and without a sampling rate it can't produce a defensible discrepancy rate either.

How is a major discrepancy defined in radiology peer review?

A major discrepancy is a difference between the original report and the reviewing radiologist's read that would reasonably have changed patient management. Differences in wording, measurement precision, or descriptive style are minor by definition. Most programs score on a four point peer review scale where the top two categories mean the finding should have been made, with a separate flag for clinical significance. What matters for a buyer is that the vendor has the definition written down and applies it the same way every month.

How do I ask a teleradiology vendor for its QA data?

Ask in writing during the RFP, before pricing is settled, for four specific things: the double-read sampling rate and how cases are selected, the written definition of a major discrepancy, the measured major discrepancy rate over the trailing 12 months with its denominator, and the documented workflow that follows a discrepancy. Give a deadline and offer a mutual NDA. A vendor that measures quality already produces this report internally and can send it within a week.

Does subspecialty reading reduce discrepancy rates?

The mechanism is plain: a radiologist who reads one anatomic area all day has seen more variants and more rare presentations in that area than a generalist has. You don't have to take the mechanism on faith. Ask the vendor to break its own discrepancy rate out by section-matched versus unmatched reads, and to state the percentage of your study types that would be matched. AstraRad routes each study to a fellowship-trained subspecialist in the matching section and will show you the match percentage for your modality mix, including overnight hours.

Should peer review be blinded, and does it matter for a vendor audit?

Yes, and it's one of the easiest things to check. In a blinded review the second radiologist forms an impression before seeing the original report or the original reader's name, so the two reads are independent. In an unblinded review the second reader anchors on what the first one wrote, and concurrence climbs for reasons unconnected to the images. Ask which one the vendor runs, and weight an unblinded rate accordingly.

Put a radiologist's name on your next read.

Tell us your modalities and monthly volume. A complete per-report rate card, with turnaround tiers and SLA terms in writing, lands in your inbox within one business day.