OBJECTIVES: To compare healthcare resource utilization (HCRU) between natural language processing (NLP)-identified hormone receptor (HR)-positive and negative breast cancer patients in the real-world setting.

METHODS: Electronic health record data for patients having a diagnostic code for breast cancer (ICD-10: C50.*) from a hospital system in the OMNY Health database was used. HR status (for progesterone and estrogen receptors) was extracted from unstructured clinical notes using a NLP t pipeline that consisted of pretrained oncological named entity recognition and relation extraction models (Spark NLP for Healthcare; John Snow Labs). Patients were divided into HR-positive (patients with a positive HR status for at least one receptor type) and HR-negative (patients with no positive HR status for either receptor type) groups. Descriptive statistics (median, quartiles 1 and 3 [Q1, Q3]) for total number of encounters and total gross charges per year of follow-up time were generated to assess HCRU by HR status.

RESULTS: Data from 2,982 breast cancer patients was analyzed using the NLP pipeline. HR status was resolved by the NLP pipeline for 399 patients (13.4%), comprising 298 HR-positive and 101 HR-negative patients). The HR-positive group had a median of 15.0 (Q1-Q3: 7.2-35.2) encounters per year, compared to 12.6 (Q1-Q3: 6.5-28.9) encounters for the HR-negative group. For total charges per year, a median of $202,039 (Q1-Q3: $81,927-$667,693) was observed for the HR-positive group, while a median of $216,283 (Q1-Q3: $78,659-$638,535) was observed for the HR-negative group.

CONCLUSIONS: These results suggest that HR-positive patients may experience slightly more encounters per year and similar total charges per year compared to HR-negative patients in the real-world setting. Earlier mortality in HR-negative patients and other comorbidities may be reasons for the observed results. Further research is required to elucidate care differences between breast cancer patients having various cancer subtypes.