The MIRROR study compared automated extraction from hospital records with manually completed clinical-trial forms for 113 breast-cancer patients across 11 trials at one hospital in Seville. Each extracted variable retained a link to its source text. The authors report strong performance across most data domains and identify unstructured extraction as an area for improvement.[1]
MIRROR study design
| Measure | Value | Context |
|---|---|---|
| Patients | 113 | Breast-cancer patients participating in clinical trials.Source 1 |
| Trials represented | 11 | Multiple protocols but one hospital site.Source 1 |
| Clinical setting | Virgen del Rocío University Hospital | Seville, Spain.Source 1 |
| Comparison | Automated EHR extraction versus manual eCRF capture | The manual trial record served as the comparison dataset.Source 1 |
| Evidence link | Original source text attached to each variable | Designed to make verification faster and more direct.Source 1 |
The result is field-specific, not one universal accuracy score
The accessible article abstract reports accuracy, precision, recall and F1 qualitatively but does not give the domain-by-domain figures needed to compare fields. This account therefore does not assign MIRROR a single accuracy percentage. The sample size and traceability design are clear; the size of the performance differences between variables is not established here.[1]
Why the source passage matters
Consider this invented note: “Previous anaemia; haemoglobin 10.2 g/dL on 12 June; repeat test pending.” A useful extraction preserves the result, unit and date, and lets a reviewer see the original wording. Recording only “anaemia” loses whether the note concerns a past condition, a current measurement or an unresolved follow-up.
Illustrative extraction from that note
| Field | Proposed value | Evidence retained |
|---|---|---|
| Haemoglobin | 10.2 g/dL | Numeric result and unit in the note |
| Measurement date | 12 June; year not stated | Do not supply a year without another source |
| Repeat test | Pending | No new test result inferred |
The paper identifies eCRF standardisation across sites as a future priority. If every sponsor or hospital asks for the same concept in a different format, the extraction system must repeatedly translate definitions. Standard fields and controlled terminology would reduce mapping work and make performance comparable between institutions.[1]
This was a retrospective data-capture comparison, not a prospective test of patient treatment or autonomous trial management. The next operational question is how much reviewer time source-linked extraction saves after corrections are included, and whether that benefit survives different hospitals and trial forms.
Sources & further reading
- Artificial intelligence for clinical data extraction from EHRs in breast cancer trials: the MIRROR study npj Breast Cancer · 7 Oct 2026 · Accessed 14 Oct 2026