Researchers have developed an artificial intelligence framework that improves the identification of skin-related complications associated with cancer immunotherapy, potentially making clinical safety monitoring faster and more reliable.
Published in npj Precision Oncology, the study evaluates a human-in-the-loop system combining large language models, retrieval-augmented generation and multiple AI agents to analyze clinical notes.
The AI-assisted reviewers achieved an F1 accuracy score of 0.88, compared with 0.77 for reviewers working without AI assistance. Agreement between reviewers also improved, while average review time fell by approximately one-third.
Immune-related skin reactions can complicate cancer treatment, making accurate identification important for patient monitoring and research.
The study contributes to wider efforts to improve AI-assisted cancer diagnosis and medical image interpretation, alongside advances in personalized cancer immunotherapy.
The findings demonstrate improved extraction of adverse-event information from clinical records, rather than direct clinical diagnosis or treatment. Further evaluation will be needed before wider deployment.