Field level confidence scores each extracted value on its own, while document level confidence gives one number for the whole page or file. Use the document score for sorting and routing whole documents. Use the field score to decide what a person checks, because a single wrong amount or code can hide inside a strong document average.
Field level vs document level confidence, side by side
Field level and document level confidence answer different questions. A field score says how sure the system is about one value, such as a plan payment. A document score covers the file as a whole, which can mean its type, its overall reading quality, or a summary of all its field scores, depending on the tool.
The table compares the two.
| Field level confidence | Document level confidence | |
|---|---|---|
| What it summarizes | One extracted value, such as a member ID or allowed amount | The whole page or file |
| Typical source | Character reading, location on the page and fit to the field's format | Document type prediction, page quality, or a roll-up of field scores |
| Who uses it | Reviewers and the routing rule that fills the review queue | Intake staff sorting mail, and dashboards tracking batches |
| What it catches | One doubtful value on an otherwise clean page | A bad scan, the wrong document type or an unreadable page |
| What it misses | Whether the document was classified correctly in the first place | Which specific value is wrong |
The two scores are not rivals. A document score is good at spotting a page that is unusable from the start, such as a blank fax cover sheet or an upside-down scan. A field score is good at spotting the one value on a usable page that deserves a second look.
Before comparing either score with real error rates, check whether the numbers mean what they claim. That check is the subject of confidence score calibration, which tests whether fields scored at 95% are right about 95 times in 100.
Why can a high document-level confidence score hide a bad field?
A high document-level confidence score can hide a bad field because, when it is built as an average, many clean fields outweigh one weak one. A page with 24 values read cleanly and one smudged remark code still averages well above most thresholds, so routing on that average sends the page through and the weak code with it.
Here is a worked example on a fictional EOB, with every value made up. The document score in this example is the mean of its 25 field scores.
- 24 fields, including patient name, member ID, dates and amounts, average 0.975.
- One remark code, printed small in a footnote, scores 0.62. Its example value reads "N13O", with a letter O where the payer printed a zero.
- Document score: (24 × 0.975 + 0.62) ÷ 25 = 24.02 ÷ 25 ≈ 0.96.
With a threshold of 0.90, the document passes at 0.96, and the remark code at 0.62 posts without anyone seeing it. With field level routing at the same threshold, the other 24 values flow through and only the remark code goes to a reviewer. One field to check, instead of none or all 25.
The same thing happens with other roll-ups. A document score based on page image quality will be high for a crisp scan even if the system has placed the right number in the wrong column. A misplaced value is a field problem, and only a field score can flag it. The kinds of misreads that produce weak fields, such as a letter O read for a zero, are covered in how accurate OCR is for medical records.
When document-level confidence is enough
Document-level confidence is enough when the decision you are making is about the whole document rather than any value inside it. Sorting incoming pages by type, deciding which queue a page joins, and spotting pages to re-scan are all whole-document jobs, and a single score per page fits them well.
Common jobs where the document score does the work:
- Classification. Is this page an EOB, a CMS-1500 claim, a lab report or an intake form? The answer is one label per page, so one score per page is the right grain.
- Page routing. Which team should see this document first? A referral packet and a payer remittance go to different desks, and a document score sends each to the right place.
- Re-scan detection. A page where nearly everything scores low usually points to a bad image rather than many separate errors. A low document score is a good trigger to ask for a better scan.
- Batch monitoring. A sudden fall in average document scores across a day's mail suggests something changed, such as a new fax machine or a payer layout update.
Faxed packets are a good example. A single fax can hold a referral, insurance card copies and clinical notes, and sorting them is a document-level task before any field is read. The medical faxes and referrals category covers that intake work.
Which confidence level should drive a review queue?
Field level confidence should drive a review queue, because the queue exists to decide whether a person looks at a value before it enters your system. Document level confidence can sit in front of that step to catch unreadable or misclassified pages, but the final routing decision belongs to each field.
A sensible order uses both:
- Document check first. If the page type is uncertain or the page image is poor, send the whole document to a person or ask for a re-scan. No field from it should post yet.
- Field check second. For pages that pass, compare each field's score with its threshold. Fields above it go through; fields below it go to review with the source region highlighted.
- Log both decisions. Record which documents were held and which fields were reviewed, so you can tell later whether problems came from the page or from single values.
Where the threshold itself should sit, and what happens to fields above and below it, is explained in what a confidence threshold for extraction decides.
Schema Care routes at the field level: each value carries its own score and a link to its place on the page, and only values below your line are held for a reviewer. The steps are laid out in the Schema Care human review workflow.