Industry must fix PV data
An industry opinion argues that reusable, standardised data—not algorithms alone—is the bottleneck for AI to improve biopharma outcomes, putting data governance at the center of trustworthy AI. For pharmacovigilance, that means consistent source fields, traceable transformations and metadata discipline are prerequisites for reliable signal detection and defensible AI outputs. (biospace.com)
Drug safety teams are trying to use artificial intelligence on reports that still arrive like mismatched paper forms, spreadsheets, and copied text. The argument in biopharma this week is that the bottleneck is not smarter models first, but cleaner data first. (biospace.com) Pharmacovigilance is the part of the drug business that watches for side effects after a medicine is already on the market. It runs on individual case safety reports, which are structured records about one patient, one drug, and one suspected adverse event. (ema.europa.eu) Those reports only work at scale if everyone describes the same thing the same way. The International Council for Harmonisation’s E2B(R3) standard exists to make one company’s safety report readable by another company and by regulators. (database.ich.org) The United States Food and Drug Administration still warns that its Adverse Event Reporting System contains reports that can be incomplete, inaccurate, untimely, and unverified. If the raw reports are messy, any artificial intelligence system trained on them learns from the mess. (fda.gov) Signal detection is the job of spotting an unusual pattern early, like seeing three smoke alarms in the same hallway before anyone sees flames. The European Medicines Agency’s signal-management rules treat that process as a formal quality task, not a loose data-mining exercise. (ema.europa.eu) That is why the opinion piece keeps coming back to reusable data. If one system stores a dose in milligrams, another stores free text, and a third drops the date format, a model cannot reliably tell whether two cases are similar or different. (biospace.com) Traceability is the other half of the problem. In regulated drug safety, a company has to show where a field came from, what software changed it, and which version of the source was used, the same way accountants need an audit trail for every number in a filing. (biospace.com) The Food and Drug Administration’s current E2B(R3) materials show how specific that gets in practice. They include regional data elements, business rules, and technical specifications for sending safety reports into the Food and Drug Administration Adverse Event Reporting System in a machine-readable format. (fda.gov) So the fight is shifting from “which model should we buy” to “can our source fields survive reuse.” In pharmacovigilance, the companies that standardize field names, preserve metadata, and document every transformation are the ones most likely to get artificial intelligence outputs they can defend to regulators. (biospace.com)
