A Novel Natural Language Processing Model for Triaging Head and Neck Patient Appointments

Inaccurate patient triage contributes to suboptimal clinical capacity management and delays in patient care, which in cancer patients may significantly increase morbidity and mortality. We developed a natural language processing (NLP) model as an adjunctive tool for head and neck (H&N) patient triage workflows. This study assesses the model's ability to categorize and triage patient appointments based on available documentation.

A retrospective cohort study.

An academic institution.

A total of 83 new patients seeing an H&N surgeon from January to April 2024 with at least 1 referral record (clinic note, imaging, or pathology report) available were included in this study. Referral clinic, imaging, and pathology reports were entered into the NLP model to predict pathology type (non-endocrine H&N neoplasm, thyroid, parathyroid, and benign lesions), malignancy risk, and appointment urgency. The gold standard was the final diagnosis from pathology reports or surgeons' clinic notes.

The NLP model achieved an accuracy of 81.9% for pathology type and 86.8% for urgency level. Sensitivity was high for non-endocrine H&N neoplasms (88.9%), thyroid pathology (88.9%), and parathyroid pathology (100%), although lower for benign lesions (67.9%). Specificity was 86.8% for non-endocrine H&N neoplasms, 91.9% for thyroid pathology, 97.6% for parathyroid pathology, and 96.4% for benign lesions. Prediction of appointment urgency achieved a Matthews correlation coefficient of 0.698, reflecting strong predictive performance.

This novel NLP model demonstrated robust performance characteristics for predicting H&N diagnoses based on referring documents and excelled at identifying patients requiring urgent care based on malignancy risk. This tool may help H&N practice coordinators screen referrals, potentially optimizing patient care.