Abstract
BACKGROUND: In neuro-oncology, imaging findings and a patient’s clinical status do not always correlate. The Neurologic Assessment in Neuro-Oncology (NANO) scale is a clinician-reported outcome assessment tool consisting of nine domains to measure neurologic function routinely assessed by clinicians. METHODS: Using GPT4DFCI, Dana-Farber Cancer Institute’s (DFCI) HIPAA-secure endpoint to GPT-4o, we ran free text physical exam documentation through a secure Application Programming Interface to assess whether a large language model (LLM) can accurately generate NANO scores. We used an iterative process for prompt development. Temperature, a parameter that controls the randomness of LLM output, was set to zero. Cox proportional hazards analysis was used to predict whether NANO score at the time of bevacizumab initiation (last line therapy) could predict overall survival (OS). RESULTS: We evaluated 287 patients with progressive grade 4 gliomas seen at least twice at DFCI between 1/1/2018 to 12/31/2023 from the time of bevacizumab initiation. LLM-generated scores were grouped into 0 or 1 (none to mild) versus 2 or 3 (moderate to severe) for each NANO domain. Accuracy between clinician-reported and LLM-extracted NANO scores was high (mean = 92.5%, range 76.6-100%, n=56). Average patient age at bevacizumab initiation was 61 years, 36% were female, 11% were IDH-mutant, and 41% had MGMT-unmethylated tumors. Median total NANO at initiation of bevacizumab was 1 (range 0-11). Cox proportional hazards analysis revealed that deficits in gait at bevacizumab initiation predicting worse OS (hazard ratio 2.86, 95% CI: 1.21, 6.74; p=0.02), independent of age, MGMT-status, sex, and IDH-status. CONCLUSIONS: LLMs can generate NANO scores from clinical notes. Moderate to severe gait impairment by NANO was predictive of worse OS at bevacizumab initiation. LLMs offer the ability to scale NANO to understand how NANO may affect OS at different disease points for different disease groups.