Patient perspectives on artificial intelligence in mammography interpretation: a comparative survey study of safety-net and academic hospital settings.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, University of Texas Southwestern Medical Center, Dallas, TX, USA.
- Department of Radiology, University of Texas Southwestern Medical Center, Dallas, TX, USA. [email protected].
- Department of Population and Data Sciences, University of Texas Southwestern Medical Center, Dallas, TX, USA.
Abstract
To evaluate and compare patient perceptions of artificial intelligence (AI) use in mammogram interpretation across academic and safety-net healthcare settings. We offered a 29-item survey to patients visiting our safety-net (SNH) and academic (ACH) hospital breast imaging clinics between 04/2024-06/2024 and 02/2023-08/2023, respectively. Demographic data was compared between populations using Chi-squared tests. We used ORs (95% CI) to estimate response odds by patient factors. Significant group differences were further analyzed via multivariable regression. A total of 924 [ACH: 518(56.1%), SNH: 406(43.9%)] surveys were collected. Participants from the ACH were older (≥ 70 years: 20%vs3.1%, p < 0.001), mostly identified as Non-Hispanic White (56%vs7.2%, p < 0.001), had higher income (≥ $100,000: 49%vs3.2%, p < 0.001), higher education (≥ college: 71%vs20%, p < 0.001) and higher self-reported knowledge of AI (68%vs56%, p < 0.001) compared to SNH. Use of AI alone or as a second reader was accepted by 74%, with SNH participants being less likely to accept [OR(95%CI): 0.71(0.53-0.96), p = 0.02]. SNH participants were more likely to request a reading by AI following radiologist-interpreted abnormalities [1.83(1.35-2.49), p < 0.001], rate AI as the same or better than a radiologist at detecting cancer [1.54(1.12-2.15), p = 0.01], and have higher concern regarding data privacy [1.87(1.22-2.93), p = 0.01]. Higher education [1.99(1.33-2.99), p < 0.001] and self-reported AI knowledge [1.98(1.38-2.83), p < 0.001] were associated with higher acceptance of AI use, while Non-Hispanic Black race [0.40(0.25-0.65), p < 0.001] was associated with lower acceptance when controlled for other covariates. Significant differences exist in patients' views of AI between the demographically distinct academic and safety-net populations. Our study revealed lower educational attainment and Non-Hispanic Black race as independent factors associated with lower acceptance of AI.