AI Across the Workflow, Part 2 — Duplicates, Signals, Narratives, and Reports

Continuing the stage-by-stage pass: duplicate detection, which Module 10 covered as a high-stakes early step, now commonly runs on ML semantic similarity — comparing patient profile, drug, event, and date range across both within-database and cross-database sources, producing a confidence-scored match list rather than a binary yes/no. Reported results: near-elimination of undetected within-database duplicates. What has to stay human: confirmation before any merge is executed, clinical judgment on partial-match cases, and investigation of any cross-database duplicate the system flags. The specific risk of over-trusting this tool cuts both ways — spurious merges of genuinely distinct cases, or false negatives leaving real duplicates in the database, quietly inflating a signal count exactly the way Module 10 warned.

Signal detection, Module 7’s entire subject, now commonly runs continuous disproportionality monitoring rather than periodic batch analysis, with ML clustering across demographics, geography, and indication identifying multi-dimensional patterns that pairwise PRR/ROR analysis alone would miss. The non-negotiable human step is exactly what Module 7’s Lesson 7.6 already taught in depth: every candidate signal a system surfaces gets validated by a qualified medical reviewer, and every signal conclusion is a documented human decision — AI output is a candidate, never a validated finding on its own. Over-action on statistical noise, or missed signals if human review capacity doesn’t scale with AI-generated candidate volume, are the specific named risks.

Narrative generation, Module 5’s subject, is where GenAI drafts a structured narrative directly from populated case data fields — referencing the product’s SmPC for context, incorporating dechallenge/rechallenge information straight from database fields — producing a first draft in minutes against the 20-40 minutes a manual first draft typically takes. This is most valuable for high-volume, well-characterised case types, and least reliable for exactly the situations Module 5 flagged as needing the most careful writing: ambiguous causality, unusual presentations, special case types like death or pregnancy. Medical review and editing of every narrative before submission, and physician sign-off specifically for serious cases, remain mandatory — the AI draft is a starting point, never the final product, and a narrative that’s clinically thin or factually incomplete despite reading fluently is exactly the risk Module 5’s AI-drafting lesson already named.

Aggregate report data assembly — Module 8’s territory — sees AI query the safety database directly, populate PBRER tabulation tables, flag discrepancies between cumulative totals and interval data, and generate a signal-status table from the tracking system, delivering a significant reduction in table-preparation time along with automated consistency checking across report sections. What can’t be automated away: medical reviewer confirmation of each table’s clinical scope, and the physician writing or substantially editing the benefit-risk narrative and executive summary — exactly the structural centrepiece Module 8’s Lesson 8.7 identified as requiring the most integrative human judgment in the entire report.