Clinical evidence map
A review workspace for faxed prior-authorization packets. It reads the fax, removes names and dates, pulls out the clinical facts and shows the exact source words for each coverage criterion. The nurse makes the decision.
decision
A utilization-management nurse decides, criterion by criterion, whether the packet shows evidence. The tool proposes a status with its source quote. The nurse accepts, corrects or flags it, and makes the coverage call.
baseline
The same services run on the clean source text without OCR, keeping the same privacy step. That gets 180 of 250 criterion statuses right, against 124 of 250 from the page images, which shows how much reading the image costs. Neither is a comparison with manual nurse review.
review
Anything uncertain goes to the nurse: low OCR confidence, conflicting statements and missing dates. Every accept, correct or flag is written to an append-only, hash-chained log. The log records the review. It does not change the answer key or approve anything.
data
Synthetic packets with fictional names, rendered as fax images and read by local OCR. No real patient data, no model API and no network service. The app has no authentication or multi-user isolation, so it must not be used with real data.
platform
Six services with typed inputs and outputs: OCR, redaction, extraction, evidence linking, citation checking and audit. The redaction service is the redaction gate project, reused as a library. A care-management summary reuses redaction and extraction as a second consumer.
limits
Draft labels pending my pharmacist review, one author, no test set kept apart from development, and English templates. The extraction and criterion targets in the project brief are not met yet. No clinical validation, time saving or coverage decision is claimed.
Evaluation
Draft results, pending pharmacist review. 50 synthetic fax-style packets across 3 fictional criteria sets. All 50 packets are development data. No test set was kept apart from development yet, and there is no comparison with manual nurse review.
What was measured: Every finding points to its source | 728 of 728 extracted values, evidence links and summary citations reproduce their quote from the redacted page
What was measured: Marked as having evidence | 97 of 250 criteria, and none of those 97 contradict the draft labels
What was measured: Criterion status matches the draft exactly | 124 of 250
What was measured: Called missing although the draft labels show it or ask for review | 38 of 240 (25 on the five pages with simulated skew, 13 where OCR read A1c as Alc)
What was measured: Clinical facts pulled out | 356 of 362 extracted values correct, 356 of 600 labeled values found
What was measured: OCR character error rate | 1.21% (418 edits in 34,599 characters)
What was measured: Unsupported summary sentences caught | 150 of 150 injected claims, by an exact-quotation rule
By page condition: criterion statuses matching the draft: Clean render | 5 | 11 of 25
By page condition: criterion statuses matching the draft: Simulated 7-degree skew | 5 | 0 of 25
By page condition: criterion statuses matching the draft: Handwriting-style font, not handwriting | 5 | 10 of 25
By page condition: criterion statuses matching the draft: Missing lab | 5 | 16 of 25
By page condition: criterion statuses matching the draft: Conflicting dose | 5 | 15 of 25
- Synthetic packets built from English templates, so scores may look better than they would on varied real faxes.
- The pages are images rendered from text, with simulated skew and noise. None were printed, faxed or scanned, and the handwriting group uses a handwriting-style font. Drawn handwriting comes next.
- Every label is a draft pending my pharmacist review. No second independent reviewer has checked them.
- All 50 packets are development data. No test set was kept apart from development yet, and there is no confidence interval.
- The privacy step removes all dates, so time-window and trial-length criteria always need review. It also mistakes the first word of the condition for a name, such as Type in Type 2 diabetes, in 48 of 50 request letters.
- The exact-quote check flags any quote that does not appear on the page. It cannot tell whether a quote supports the finding, or whether it is true.
- No authentication or multi-user isolation, and the redaction step is not certified de-identification. It makes no coverage decision and claims no time saved.