GSK's pharmacovigilance teams developed and prospectively tested a machine learning tool, Machine Learning for Intelligent Triage (MLIT), to help safety reviewers work through statistical safety alerts for vaccine and drug portfolios. In two four-month prospective validation studies covering more than 20 products and over 2,300 alerts, 92% of vaccine alerts were closed in line with the model's top-ranked prediction and 98% within its top three predictions. Reviewers reported a self-estimated collective saving of about 83 hours, roughly 24% of the triage time spent during the validation.
The work was published in Drug Safety by GSK authors in August 2026. The company's signal management platform generated about 27,000 alerts across drug and vaccine portfolios in 2023, and 99.9% were closed with a documented rationale. The authors describe this as a sign of system sensitivity rather than missed signals, but also as an inefficiency: manual triage is time-consuming, varies between reviewers, and records its clinical reasoning only partly in free-text comments, which limits auditability and traceability.
Alerts come from disproportionality analysis, time-to-onset tests, designated medical events and aberration checks. For each one, a reviewer either closes the alert with one of 13 predefined decision rationales or starts further investigation. MLIT treats this as a multiclass classification problem. Separate models were built for vaccines and drugs, using alert statistics and case-level attributes from individual case safety reports, several chosen with the Clinical Utility Score for Prioritisation. XGBoost was the best of several algorithms tested, trained and tested on a 75/25 split, and refined iteratively with Shapley Additive Explanations analyses and expert review so that the model's logic matched reviewers' reasoning. The first vaccine model used alerts through December 2022 for 17 marketed vaccines; the refined one was retrained on all 36 marketed vaccines with alerts through July 2024 and reached a weighted F1 score of 0.81, up from 0.70, with an accuracy of 0.79. Drug model performance was similar.
In the first prospective validation, from November 2024 to February 2025, reviewers were shown the refined model's top three predictions and the features behind each one and recorded whether they agreed. For the 881 vaccine alerts in that study, 810 (92%) were closed in line with the top-ranked prediction, 41 with the second and 13 with the third. The tool also exposed inconsistencies and human errors in manual triage, suggesting a quality-control role. Across both prospective validations, more than 2,300 alerts in total, only two investigations took place where the decision to investigate was not among the top three predictions; both were precautionary, produced no validated signal and had low prediction confidence.
The authors keep humans in the loop: reviewers see the top predictions with their probabilities and the features that explain them, and can override the model. Misclassifications clustered among alerts with low or closely matched top probabilities, which the authors say supports using probability thresholds to separate confident predictions from those that need closer human scrutiny; first-time alerts, which lack a previous-decision feature, performed worse and should stay prioritised for manual review. They also note that the strong influence of the previous-decision feature risks anchoring the model to historical triage patterns. Before routine use, MLIT will undergo formal computerised system validation followed by a parallel phase of at least six months. Planned enhancements include automatically generating closure rationales and adding external triggers such as health authority requests.