ANSM's AI-Assisted Pharmacovigilance — France's Machine-Learning Pre-Coding of COVID-19 Vaccine Adverse-Event Reports

On 6 April 2021, France's national medicines safety agency, the Agence nationale de sécurité du médicament et des produits de santé (ANSM), launched a new national pharmacovigilance application with a built-in artificial-intelligence module to categorise and pre-code adverse-event reports. The tool was introduced specifically to handle the surge in COVID-19 vaccine side-effect reports, mobilising a network of ten regional pharmacovigilance centres (CRPV) so that human experts could focus their time on the most complex or serious cases rather than routine coding.
The underlying machine-learning pipeline, developed by a consortium of INSERM's pharmacoepidemiology team, Bordeaux University Hospital and the health-tech company Synapse Medicine, was validated in a peer-reviewed study (Martin et al., 2021) that tested several algorithms — including logistic regression, random forests, LightGBM and deep-learning models (FastText, LSTM, CNN) — on 2,058 patient-submitted adverse-reaction reports (10,675 coded reactions across 125 MedDRA terms). The best-performing LightGBM model reached an AUC of 0.93 and an F-measure of 0.72 on internal validation.
When tested on an external, independent set of 187 cases from a different pharmacovigilance centre, performance dropped to an AUC of 0.91 and an F-measure of 0.58 — a reminder that the model generalises imperfectly across sites and that human review of AI-suggested codes remains part of the workflow. ANSM has not published a formal post-deployment evaluation of how the tool affected report-processing times or EU submission compliance since the 2021 rollout.

Read the full analysis: https://ansm.sante.fr/actualites/lancement-de-la-nouvelle-application-nationale-de-pharmacovigilance