Why this tool makes you check every value

Automatic extraction from a document is good, not perfect. A decimal point read in the wrong place turns a routine result into an alarming one, or hides one that matters.

That is why nothing is explained here until every value has been confirmed against the report it came from. It is the least convenient part of the process and the part that makes the rest defensible.

Where extraction goes wrong

Columns that drift out of alignment on a scan can attach a unit to the wrong test, or a reference range to the row above. The result is a value that looks entirely plausible and belongs to something else.

A reference range split across two lines can be read as two ranges, or as none. Where no range is found, a comparison cannot be made at all — which is better than making the wrong one, but only if it is visible.

Photographs taken at an angle, in poor light, or of a folded page are the single most common cause of misreadings. Text that a person can read easily can still defeat automatic extraction, because the software has no expectation of what the number ought to be.

Test names are their own problem. Laboratories abbreviate differently, and the same marker appears as haemoglobin, hemoglobin, HGB and Hb across four reports from four providers.

What the checking step is for

Rows that were read with low confidence are marked and have to be confirmed one at a time rather than in bulk. The most common reason for low confidence is a missing reference range, which is exactly the case where a wrong comparison would be least obvious.

Any value can be corrected or removed. A corrected value is what gets explained and what appears in a download, so a correction made here follows through everywhere.

The page image sits beside the rows for the same reason. Checking a number against a memory of the report is not checking; checking it against the report is.

Why not simply trust the extraction

Because the cost of the two possible errors is not symmetrical. An explanation built on a correct number is useful. An explanation built on a wrong number is worse than no explanation, because it is fluent, specific and mistaken.

Automatic extraction offers no signal about which of those two situations you are in. Confidence scores help, but a confidently misread column is still confidently misread.

The person holding the report is the only participant who can see both the document and the extracted values. Putting the check there is not passing the work along; it is putting it where it can actually be done.

The shortcut that is not a shortcut

Typing values in by hand skips extraction entirely and produces exactly the same explanation. For a short panel it is often faster than checking an extraction, and it means no file is uploaded at all.

For a long panel, uploading and checking is usually quicker. Either way the values that get explained are values a person has looked at and agreed with.