A locally deployed large language model for pathology-informed and nurse-reviewed communication support in bladder cancer immunotherapy

A study evaluated the safety and clinical usability of a locally deployed Qwen2-7B language model to support nurses in caring for bladder cancer patients treated with immunotherapy. The model was tested in five areas: general immunotherapy education, pathology-informed communication, biomarker communication, symptom triage, and risk-based triage. The model achieved a score of 3.83 for general patient education and 3.57 for pathology-informed communication, but only 3.26 for biomarker communication. In symptom triage, the model agreed with expert triage in 75.9% of cases and correctly recognized warning signs in 71.1% of cases. Only 3.7% of outputs required no revision, while 31.4% required substantial revision or withholding. The study concluded that the model can support nurse-reviewed communication, but biomarker communication and symptom triage require nurse oversight and clinician backup.