基于大语言模型的血液/肿瘤科患者消息分诊：性能、安全性与临床意义

Background: Patient portal messages are a major source of real-time symptom reports in Hematology/Oncology, but manual review is labor-intensive and may delay identification of urgent issues. Large language models (LLMs) can process free text at scale, yet their safety and performance for triage in cancer care are uncertain. Methods: We conducted a retrospective study of adult patients with hematologic malignancies receiving care at a single center within Southern California Permanente Medical Group, the care delivery arm of Kaiser Permanente, from November 2024 to November 2025. All patient-initiated messages to the Hematology/Oncology clinical team were extracted from the electronic health record portal. A stratified sample of messages was annotated by clinicians for urgent and non-urgent categories. Messages labeled as urgent reflected symptoms warranting prioritized clinician review within 24-48 hours. We developed an LLM-based classifier using few-shot learning that generates a label populating the “Topic” column in the receiving clinician's inbox. Primary outcomes included discriminative performance metrics such AUROC, sensitivity, specificity, and F1 score for classifying urgent messages. Results: A total of 10,683 messages from 2,287 unique patients were processed by our model; 433 messages were labeled as urgent. For urgent message classification, the few-shot LLM achieved an AUROC, sensitivity, specificity, and F1 score of 91.05% (95% CI, 85.69%-96.42%), 97.62% (95% CI, 92.50%-99.89%), 84.48% (95% CI, 74.60%-93.34%), and 89.13% (95% CI, 81.82%-95.74%) respectively. Conclusions: An LLM-based triage system reliably categorized patient messages into urgent and non-urgent categories and has the potential to facilitate timelier action for urgent symptoms within a Hematology/Oncology department. The system's practical utility highlights its relevance for departments seeking to leverage NLP to improve message management in cancer care.