ML Predicts Severe AEs
- A JCO study used FDA's FAERS database with machine learning to predict severe adverse events in oncology patients. (x.com) - The model illustrates how large spontaneous-report datasets can be used for signal detection and risk prediction. (x.com) - Deploying such tools will require governance and traceability consistent with emerging FDA–EMA AI expectations. (x.com) (appliedclinicaltrialsonline.com)
A machine-learning model trained on the Food and Drug Administration’s adverse-event database sorted cancer drug reports by the risk of severe harm. (asco.org) The model used reports in the FDA Adverse Event Reporting System, or FAERS, from 2012 Q4 through 2024 Q3 that listed cancer as an indication. Severe adverse events were defined as death, life-threatening events, disability, hospitalization, or congenital anomaly or birth defect. (asco.org, fda.gov) FAERS is the FDA’s postmarketing safety repository for marketed drugs and biologics, built from reports submitted by manufacturers and from voluntary reports by clinicians, patients, and consumers. The database covers January 2004 to the present and is updated quarterly. (fda.gov) The Johns Hopkins-led team said it started with 2.28 million oncology-related reports after excluding non-cancer and incomplete entries. About 44% of those reports met the study’s severe-event definition. (asco.org) On a test set of about 450,000 reports, the calibrated LightGBM model posted 75% accuracy, 73.7% precision, 86.3% recall, and an area under the receiver operating characteristic curve of about 0.82. A logistic-regression baseline reached 73% accuracy, 72.9% precision, 81.5% recall, and an F1 score of 0.77. (asco.org) In plain terms, the system works like a triage filter for a very large inbox: it looks across age, drug combinations, prior event history, and other report features to estimate which cases are more likely to involve severe toxicity. Partial SHAP analysis, a method for ranking which inputs pushed the prediction, pointed to age 65 and older, multi-agent chemotherapy, and prior adverse-event histories as leading predictors. (asco.org) That approach fits the basic job of FAERS, which regulators use for postmarketing signal detection after drugs are already on the market. The FDA says the database supports safety surveillance rather than proving that a drug caused a reported event. (fda.gov) The study was presented through the American Society of Clinical Oncology and described by its authors as the largest FAERS-based oncology severe-adverse-event analysis they had run. The abstract said there are still no well-established predictive tools for identifying which cancer patients are at highest risk of severe adverse events. (asco.org) Any move from conference model to regulated workflow would land in a stricter policy environment than even a year ago. On January 14, 2026, the FDA and European Medicines Agency jointly released 10 principles for good artificial-intelligence practice in drug development. (fda.gov, ema.europa.eu) Those principles call for a risk-based approach, clear context of use, multidisciplinary expertise, data governance and documentation, life-cycle management, and clear essential information. The FDA and EMA said the framework applies across the medicines lifecycle, including safety monitoring. (fda.gov, ema.europa.eu) So the immediate takeaway is narrower than “AI can predict toxicity” and more concrete: a model built on 2.28 million FAERS cancer reports beat logistic regression on recall while keeping similar precision, and any real-world rollout would need the data trail and controls regulators now expect. (asco.org, fda.gov)
