FDA RECORD · METHOD

How the record is built.

Each decision is one case. A case is built by running the kgx pipeline over the decision's FDA documents: every page is read, split into sentences with their page numbers, and turned into a knowledge graph. No language model is used anywhere in it.

An editor picks the sentences that state the agency's position on each thread. The exporter then finds each sentence again on the PDF page it came from and copies it from there. If the words are not on that page, the build stops. Dates, committee votes, the basis of approval and the confirmatory terms are read from the documents the same way, never typed in.

Each position is compared to a named earlier position on the same thread and marked with one status. The first position on a thread is its origin.

Origin
The first agency position on this thread in the record.
Consistent
Holds to the earlier position it is compared to.
Conditional
Holds to the earlier position only under a condition the agency states.
Departs
Moves away from the earlier position it is compared to.
Not addressed
The decision's loaded documents do not address this thread.

Scope: the list of every FDA-approved Alzheimer's drug is read from openFDA's drug labels (any current label whose indication names Alzheimer's disease, plus discontinued brands of the same ingredients) and the Drugs@FDA database, refreshed weekly; its indication quotes are checked against the label text word for word. The in-depth record covers anti-amyloid antibody decisions. A decision whose documents are not loaded yet shows as record incomplete, with what is missing.

Citations give the PDF page number, which can differ from the page number printed on the document.

What this record cannot show

The record is only as complete as what FDA publishes. Where FDA is silent, the page says so rather than filling the gap.