Natural language processing identification of documented mental health symptoms associated with risk of mental health disorders in patients with cancer.

Abstract

1561 Background: Delayed diagnosis and care of mental health disorders (MHD) is a significant challenge in the care for patients with cancer. The objective of this study was to use natural language processing (NLP) to identify words related to mental health documented in clinical notes surrounding the time of cancer diagnosis and assess their predictive ability of future, new MHD. Methods: This single institution cohort study consisted of patients diagnosed with cancer between January 2012 and November 2022. Cancer and MHD were identified based on ICD-10 codes obtained from deidentified electronic health record data. MHD included psychotic disorders (F20-29), mood disorders (F30-39), and anxiety disorders (F40-48). The clinical Text Analysis Knowledge Extraction System was applied to deidentified clinical notes, and symptoms mapped to SNOMED concepts relevant to mental health were identified. These mental health symptoms were aggregated in the 15 days preceding and 15 days following a first cancer diagnosis and analyzed across MHD status. Patient characteristics including sex, age, race, cancer, and insurance were also analyzed. Results: This cohort consisted of 64,010 patients with cancer who had no documented MHD prior to cancer diagnosis, with a majority being 40-64 years old (45.8%) or 65+ (43.7%) and identifying as male (53.0%) or white (60.2%). Most patients had prostate (12.5%), hematologic (10.8%), or breast (10.3%) cancer and private insurance (46.2%). 9,825 (15.3%) patients developed a newly documented MHD, with a median time of 139 days (IQR: 40-466) from cancer diagnosis. The top five most common mental health documented symptoms for all patients were normal mood (23.3%), mental state finding (17.9%), worried (10.2%), feeling content (9.9%), and cognitive function finding (6.6%). Those who had a future MHD had higher documented rates across all mental health symptoms. Multivariate cox proportional hazards model identified 18-39 years old, female, white, and Medicaid or Medicare insurance as independent factors associated with an increased risk of a future, new MHD. Prostate cancer was associated with lower risk of a future MHD. Panic (OR 2.1 [95% CI 1.8-2.4]), feeling nervous (1.9 [1.5-2.4]), feeling guilt (1.9 [1.4-2.5]), mild anxiety (1.8 [1.4-2.4]), and feeling frustrated (1.4 [1.2-1.6]) were identified as the symptoms most strongly associated with an increased risk of a future MHD. Conclusions: NLP extracted mental health symptoms documented in clinical notes correlated with an increased risk of documented MHD. Computational approaches may be tools for improving the timely diagnosis of MHD and referral to specialty services. Further work is needed to investigate potential disparities in documentation and management of care for patients with cancer who develop MHD, including delays between documentation and eventual diagnosis.