Home / Case studies / Intake
Intake

Admission paperwork scanned once and filed to the right chart

A start-of-care packet scanned at the office printer is read, matched to the right patient and split into one document per form, with completeness checked against the agency’s own admission policy.

Status
Built and verified, switched off on purpose
Result
42 pages in, 7 forms recognized, 18 blank backs dropped, patient pinned unanimously.

Built and verified, switched off on purpose

Verified on three real packets and wired in as a job. Filing is held until the agency approves the completeness rule; until then every packet routes to review, which is the safe default.

The situation

A California home health agency’s start-of-care packet has about seven different forms in mixed letter and legal sizes, printed double-sided and scanned up to about 100 pages per feed. Staff sorted it and filed it to each chart by hand.

The cost of leaving it alone

Misfiled admission paperwork is a PHI incident waiting to happen, and an incomplete packet found at survey time is a finding.

What we built

The office printer scans into a synced SharePoint library. A job on a local server reads each page with OCR, drops blank backs, and recognizes each form from a table. It identifies the patient by fuzzy-matching names across at least two pages against the EMR census, and refuses to file on any disagreement. It then writes one PDF per form into the patient’s chart in a separate, restricted clinical site.

Completeness is checked against the agency’s own admission policy: not whether a form is present, but whether the policy’s required documents are covered. Incomplete packets go to review, and re-scans after a paper jam are recognized so a packet is not filed twice.

Tesseract OCR on an encrypted local serverPython and poppler PDF toolsSharePointPrinter scan-to-share

Why AI, or why not

No. OCR plus deterministic matching does the job. Blank pages are found by a fixed grey level rather than by OCR text length, because OCR invents about two thousand characters from the speckle on a blank back.

The results

42 pages
in three real packets: 7 forms recognized in each, 24 content pages, 18 blank backs, nothing unrecognized
3 of 3
patients pinned unanimously across pages
4.5–14.4%
of pixels carry real print, against 0.000–0.005% on blank backs, so the threshold is not a judgment call
3 to 0
times the first test filed one packet; after the re-scan fix, duplicates were zero

How PHI was protected

  • Page images and the OCR working files are processed on agency-controlled hardware, on an encrypted volume. No cloud model is involved.
  • Emails and cards were verified to carry no patient name or record number.
  • Filing goes to a private clinical site hidden from the address list.
  • The OCR write cache was moved onto the encrypted volume after the default turned out to be plain disk.

What we would tell you to skip

Forms and policy documents are not one-to-one: one consent sheet can satisfy five documents on the policy list. Check coverage of the policy, not the presence of a form.

Reusable for

Agencies with paper admission packets and a multifunction scanner.

Tell us the one workflow that costs you the most time.

A working session is free of obligation: we look at one process, tell you plainly whether AI belongs in it, and what keeping it inside HIPAA would take.

Book a working session