AI Across the Workflow, Part 1 — Intake, Coding, and Literature Monitoring

This lesson and the next work through AI’s deployment stage by stage across the exact PV workflow this course has already taught — not as new material, but as a second pass showing precisely where AI now sits inside processes you already understand from Modules 4, 5, 7, and 10.

Case intake is where AI’s presence is most mature. NLP reads a source document — an email, a PDF, sometimes a call transcript — and extracts the drug, the event, patient demographics, and time to onset; proposes an initial seriousness assessment; flags likely duplicates against the existing database; and creates a structured ICSR shell ready for review. Reported efficiency gains run 40-65% in cycle-time reduction. What stays irreducibly human: validation of every extracted data point against the source document, clinical judgment on the "other medically important condition" catch-all from Module 6, final seriousness determination, and submission authorisation. Get the oversight wrong here and the specific risks are missed seriousness, incorrect drug-event attribution, or an undetected duplicate quietly inflating a signal count — precisely the failure mode Module 10’s intake lesson warned about.

MedDRA coding assistance, which Module 4 already covered from the term-selection-discipline side, works by having NLP map a verbatim description to candidate LLTs with a confidence ranking and alternative suggestions, flagging low-confidence cases for senior review. Reported gains: 50-70% reduction in coding time for common, well-described events. What can’t be delegated: final code selection for every case, manual coding for low-confidence or complex events, and a mandatory second-coder check for serious cases. The named risk if oversight lapses is exactly what Module 4 flagged with the "Myocardial infarction in the fall of 2000" autoencoding example — systematic miscoding, NEC overuse, or investigation-only coding errors that scatter genuinely related cases across loosely-connected terms.

Literature monitoring — mentioned in passing in Module 7’s signal-sources lesson — gets the most dramatic efficiency numbers of the three: AI searches defined databases on a schedule, NLP classifies each retrieved abstract as relevant, potentially relevant, or not relevant, and retrieves and extracts case data from the full text of qualifying articles. Reported reduction in irrelevant-article review time: 80-90%. The non-negotiable human step is full-text review of every AI-flagged article, the ICSR-creation decision itself, and clinical relevance assessment for aggregate reporting. The named risk if this is skipped: missed literature signals if the AI screen’s sensitivity isn’t what it’s assumed to be, or false confidence in screening performance that was never periodically revalidated.