ASCERTAINING PROVIDER-LEVEL IMPLICIT BIAS IN THE TREATMENT OF PROSTATE CANCER UTILIZING NATURAL LANGUAGE PROCESSING TECHNIQUES

• Published in: Urologic oncology 2024-03, Vol.42, p.S86-S86
• Main Authors: Ramaswamy, Ashwin, Hung, Michael, Pelt, Joe, Iranmahboub, Parsa, Calderon, Lina Posada, Hu, Jim C, Barbieri, Christopher, Scherr, Douglas S.
• Format: Article
• Language: English
• ISSN: 1078-1439; 1873-2496
• DOI: 10.1016/j.urolonc.2024.01.241

Implicit biases in medicine are personal attitudes about race, ethnicity, gender, disability, and other characteristics that may lead to discriminatory patterns of care. Presently, there is no consensus on whether implicit bias represents a true predictor of differential care as there is a near-total absence of studies using real-world data. Unstructured clinical narrative data (e.g., progress notes) represents a unique opportunity to capture implicit bias as it is the most common way providers document granular details of the patient encounter. Therefore, leveraging recent advances in natural language processing (NLP), we assessed the relationship between providers’ implicit bias and their respective treatment parity in prostate cancer, a condition with well-documented disparities in care. All note-level data from patients with very low-, low-, or favorable intermediate-risk prostate cancer followed by three urologic oncologists were extracted from electronic health records (EHRs) between January 2010 and December 2021. NLP-screened human annotation using validated regular expression phrasal keyword libraries was used to evaluate provider's care on four quality indicators: (1) utilization of active surveillance (AS); (2) utilization of prognostic (e.g., Decipher) biomarkers; (3) discussion of patients’ sexual function; and (4) discussion of patients’ urinary function. Additionally, the Race Implicit Association Test was administered to all providers. A total of 1,299 patients were included for analysis. Stratified by the highest volume providers demonstrated a higher proportion on AS were White (75% vs 66%) in one of the three practices (Provider C, Table 1). A higher proportion of patients who had biomarkers were White;(75.4% vs 64%). On multivariable analysis, older age was associated with a higher odds of AS (OR1.05, 95%CI 1.03-1.06) and non-White patients were less likely to receive biomarkers (OR0.5, 95%CI 0.35-0.70) (Figure 1). A discussion about either EF or UF was less likely for non-White patients (OR 0.73, 95%CI 0.55-0.96 and OR0.74, 95%CI0.56-0.96, respectively). The Race IAT demonstrated that Provider C had a slight preference for White vs. Black persons which was statistically associated with a recommendation of AS to White vs. Black men (p=0.011). Overall clinical practice patterns as measured by the four quality indicators were associated with both patient race and implicit provider racial preferences among urologic oncologists in the manag