AI-enabled Pharmacovigilance & Drug Safety

Two briefs. Real-shaped work.

1. Project 1

   LLM literature-screening pipeline for adverse events, validated against manual screening

   A PV team screens published literature every week for case reports involving its product. You build an LLM-assisted first-pass screener for a fictional long-marketed generic, orvasetan, on a frozen practice set, and prove how well it works. You write the screening criteria, create an independent gold set by manual double screening, run a fixed prompt, measure recall and precision, analyse every miss, improve once, and re-test on held-out abstracts. You finish with a validation record and a short SOP that keeps a qualified human responsible for every potential case. The pipeline is a training artefact; it is not used on real company data.

   You hand in

   • Screening criteria (one page) and the saved PubMed search string
   • Gold set of 300 abstracts labelled by you and a peer independently, with a reconciliation log
   • Prompt and model record (model name, access date, prompt version history)
   • Results sheet with a confusion matrix, recall, precision and workload cut, before and after one improvement
   • Held-out test on 50 fresh abstracts with the same metrics
   • False-negative analysis listing every missed abstract and its cause
   • Validation record and a one-page SOP for human review, routing of 'unclear' and re-validation triggers

   Data: A frozen Saarthi sandbox set of 350 fictional titles and abstracts about a fictional long-marketed generic, orvasetan (300 for the gold set, 50 held out). They are written for teaching, are not PubMed records and contain no patient data; your saved PubMed search string is still a real search. Reference labels from two Saarthi reviewers are held back for grading.. Reviewed by Saarthi against a published rubric, signed off by a mentor.

2. Project 2

   Synthetic ICSR narratives with QC and MedDRA coding checks

   You process 25 synthetic source packets the way a case processing team would. For each you check validity and seriousness, build the structured case, code events, history and indications from the practice MedDRA-style list, draft the narrative with AI from your structured fields, and QC it line by line against the source. You track every AI error in a discrepancy log and report the correction rate by error type. The output shows you can use AI for speed while keeping the case accurate. All data is synthetic.

   You hand in

   • 25 structured case forms with validity, seriousness and Day 0 recorded
   • Coding table (verbatim, LLT, PT, primary SOC, term-list version) for every event, history item and indication
   • AI draft and final narrative for each case
   • Completed QC checklist per case
   • Discrepancy log with problem types and a summary of corrections per case
   • A one-page reflection on where AI saved time and where it introduced risk

   Data: 25 synthetic source packets from the Saarthi sandbox (call notes, emails, discharge pages, lab reports, follow-ups, some in Hinglish), written to resemble real intake but describing no real person, plus the sandbox's MedDRA-style practice term list. Reference cases are held back for grading.. Reviewed by Saarthi against a published rubric, signed off by a mentor.