# GPT-4o

Identifying cutaneous immune-related adverse events (cirAEs) in outpatient notes of patients on immune checkpoint inhibitors has depended on manual chart review, which is slow and inconsistent between reviewers. In a multi-institution study, non-dermatologist reviewers adjudicated with a retrieval-augmented, multi-agent large language model that returned a binary recommendation plus the supporting note excerpts. Against ground truth from an oncodermatologist, the assisted group reached F1 0.88 and Cohen's κ 0.82 versus 0.77 and 0.50 for manual review, and reviewers self-reported a drop from 15 to 10 minutes per chart.

Identifying cutaneous immune-related adverse events (cirAEs) in outpatient notes of patients on immune checkpoint inhibitors has depended on manual chart review, which is slow and inconsistent between reviewers. In a multi-institution study, non-dermatologist reviewers adjudicated with a retrieval-augmented, multi-agent large language model that returned a binary recommendation plus the supporting note excerpts. Against ground truth from an oncodermatologist, the assisted group reached F1 0.88 and Cohen's κ 0.82 versus 0.77 and 0.50 for manual review, and reviewers self-reported a drop from 15 to 10 minutes per chart.

Patient questions about lung cancer screening CT results can now be answered by a multimodal agentic system built from a planner and an executor. Earlier general-purpose models handled text reports only, could not invoke imaging tools, and did not judge which questions belong to a physician. Across 116 screening CT reports and 699 simulated patient questions, the system reached 93.1% tool-calling accuracy and correctly deferred 92.6% of out-of-scope questions, both significantly above a single-agent baseline; among 100 responses rated by four physicians, 78.9% of answerable ones earned perfect clinical-accuracy scores.

Patient questions about lung cancer screening CT results can now be answered by a multimodal agentic system built from a planner and an executor. Earlier general-purpose models handled text reports only, could not invoke imaging tools, and did not judge which questions belong to a physician. Across 116 screening CT reports and 699 simulated patient questions, the system reached 93.1% tool-calling accuracy and correctly deferred 92.6% of out-of-scope questions, both significantly above a single-agent baseline; among 100 responses rated by four physicians, 78.9% of answerable ones earned perfect clinical-accuracy scores.

This feasibility study fed deidentified treatment summaries into OpenAI GPT-4o with structured prompts to draft long-term follow-up recommendations for childhood cancer and HSCT survivors, then compared them with clinician-generated recommendations based on institutional standards and Children's Oncology Group Long-Term Follow-Up Guidelines (version 5) in an independent validation cohort of 40 survivors, finding 467 AI-generated versus 446 clinician-generated items, with 385 of 446 clinician recommendations (86.3%) also identified by AI and 385 of 467 AI recommendations (82.4%) also present in clinician plans, while most discordant recommendations involved radiation-related exposures and survivorship scenarios requiring nuanced clinical interpretation.

This feasibility study fed deidentified treatment summaries into OpenAI GPT-4o with structured prompts to draft long-term follow-up recommendations for childhood cancer and HSCT survivors, then compared them with clinician-generated recommendations based on institutional standards and Children's Oncology Group Long-Term Follow-Up Guidelines (version 5) in an independent validation cohort of 40 survivors, finding 467 AI-generated versus 446 clinician-generated items, with 385 of 446 clinician recommendations (86.3%) also identified by AI and 385 of 467 AI recommendations (82.4%) also present in clinician plans, while most discordant recommendations involved radiation-related exposures and survivorship scenarios requiring nuanced clinical interpretation.