Agentic AI helps human reviewers identify immunotherapy adverse events in clinical notes
Published in Artificial Intelligence, Immunology, Journals, Tools, Toxicity.
Published October 30, 2026.
A Nature Cancer study published October 29 evaluates an AI system that helps reviewers identify immune-related adverse events in oncology notes. It extracts event presence, timing, severity, treatment attribution and diagnostic certainty across six toxicity types.
Retrospective testing produced a macro-averaged F1 score of 0.92 for current-event detection, but only 0.66 for severity grading. During three months of prospective silent deployment, binary F1 fell to 0.72–0.79, highlighting the importance of testing beyond development data.
In a randomized crossover study involving clinical trial staff, AI assistance shortened annotation by 2.71 minutes, improved complete-match accuracy and increased agreement between reviewers. The system was evaluated as assistance coupled with human verification.
These findings support more efficient chart review and research data collection. They do not establish autonomous diagnosis, prove improved patient outcomes or show that clinical oversight can be removed. F1 scores combine precision and recall and should not be interpreted as the percentage of patients correctly diagnosed.
Source: Gallifant et al., Nature Cancer. Published October 29, 2026. DOI: 10.1038/s43018-026-01264-6.