How we measure drafts
Every MyRadAgent draft is reviewed, edited and signed by a radiologist. The figures on this page come from those signed reports: how much of each draft survived, how many were signed without any change, and how that has moved over time. The method and its limits are stated next to the numbers, and the page is recomputed from the database rather than written by hand.
The figures
| Draft-and-signed pairs | 5,238 |
|---|---|
| Radiologist accounts contributing | 2 |
| Date range | 2026-04-01 to 2026-10-31 |
| Draft retention (mean share of draft text kept) | 98.3% |
| Signed with no edit at all | 90.4% |
| Edited in some way | 9.6% (504 reports) |
Computed from the production database on 2026-11-01 19:56 UTC; refreshed about hourly.
How much drafts change
| How much of the draft changed | Pairs | Share |
|---|---|---|
| Signed exactly as drafted (0%) | 4,783 | 91.3% |
| Up to 5% changed | 138 | 2.6% |
| 5–20% changed | 179 | 3.4% |
| More than 20% changed | 138 | 2.6% |
By study type
| Study type | Pairs | Draft retention | Signed unchanged |
|---|---|---|---|
| chest xr | 1,316 | 98.3% | 89.6% |
| msk xr | 740 | 98.4% | 90.9% |
| pelvic us | 620 | 98.6% | 91.1% |
| vascular us | 549 | 98.5% | 90.7% |
| abdomen us | 484 | 98.9% | 91.1% |
| general | 451 | 98.7% | 93.1% |
| abdomen ct | 356 | 98.1% | 91.0% |
| abdomen mri | 117 | 96.4% | 87.2% |
| head ct | 103 | 98.2% | 88.3% |
| chest ct | 73 | 97.5% | 86.3% |
| abdomen xr | 66 | 95.7% | 84.8% |
| mammogram | 41 | 98.8% | 95.1% |
Last twelve months
| Month | Pairs | Draft retention | Signed unchanged |
|---|---|---|---|
| 2026-04 | 641 | 95.2% | 83.2% |
| 2026-05 | 1,064 | 98.2% | 87.1% |
| 2026-06 | 684 | 99.2% | 94.4% |
| 2026-07 | 675 | 99.3% | 94.5% |
| 2026-08 | 914 | 99.0% | 94.5% |
| 2026-09 | 720 | 98.9% | 91.7% |
| 2026-10 | 540 | 98.1% | 86.3% |
Method
What a pair is
One report with the AI draft and the text the radiologist actually signed. Reports that were never confirmed are not pairs and do not count either way.
What “changed” means
The share of the draft’s characters that differ from the signed text, whitespace-normalized, computed with a standard sequence-matching ratio and capped at 10,000 characters per text. A two-word fix on a 300-word draft is about a 1% change; retyping the impression is a large one. “Signed unchanged” is the stricter count: any edit at all puts a pair in the edited group.
What is excluded
Pairs where the input and output describe different studies (a swapped case), pairs where the radiologist discarded the draft and signed the pasted dictation instead, gold-standard reports entered by hand (no draft to compare), and system-generated re-drafts whose change is not a radiologist’s edit. Excluding the discarded drafts makes the retention figure higher than it would be with them; the edited count below is not affected by that choice.
What this is not
Not clinical accuracy. A draft the radiologist signed unchanged was judged acceptable by that radiologist; it was not checked against a reference standard, so nothing here is a sensitivity or specificity. It also says nothing about studies that were never run through the product.
Who the data comes from
Every account on the platform, but most of the volume comes from one practice. Per-study-type rows show only with at least 20 pairs. A group pilot measures its own reports; ask for one from the pricing page.
What would be stronger evidence
A group pilot that measures its own before-and-after edit rates, turnaround and claim denials on its own reports. That is the comparison a department chair should ask for, and it is what the group pilot is set up to produce. The documentation savings estimator shows the cost side of that case with every assumption visible.