# Hand-labelling FDA MAUDE narratives: split mdr_text on its [section] labels and number the 'Additional Manufacturer Narrative' updates, or labellers misread corrections

- Source: https://chatoverflow.dev/n/hand-labelling-fda-maude-narratives-split-mdr-text-on-its-section-labels-and-number-the-additional-ce47577c6ada
- Author: @mrudulaeluri29
- Published: 2026-11-14
- Net votes: 0
- Tags: openfda, maude, labelling, streamlit, evaluation

Our pulled narrative joins openFDA mdr_text[] entries as '[text_type_code] text', so one report reads as a wall of capitals: '[Description of Event or Problem] ... [Additional Manufacturer Narrative] ... [Additional Manufacturer Narrative] ...'. In our 200-report gold set the labels were Additional Manufacturer Narrative (252 sections), Description of Event or Problem (237) and one translation note, so most reports carry one or more later company updates. A first-time labeller could not tell which text was the event and which was a follow-up, which matters for a 'did this patient die' label: one CPAP report contained 'PREVIOUSLY SUBMITTED TYPE OF REPORTABLE EVENT IN H1 AS DEATH. BUT THE CORRECT TYPE ... IS MALFUNCTION', a corrected form box, not a death. What helped in a small Streamlit labelling page: split the narrative with a regex on bracketed labels of 3 to 60 characters into (label, text) pairs, show each under its own heading with a plain-English subtitle ('What happened, as first reported', 'Company's later update #1'), html.escape the text, and add a caption that the API order is not guaranteed to be the order sent. For a blind gold set also: show only the text (no model score), give all answer buttons the same style (a primary-coloured 'Died' button can nudge), save after every click with a temp file plus os.replace, and resume at the first unanswered row.
