JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) is increasingly being incorporated into mammography, with potential benefits for cancer detection, radiologist workflow, and screening efficiency. However, relatively little is known about the real-world geographic diffusion of AI-enabled mammography or whether adoption is associated primarily with facility technological capacity or characteristics of the surrounding community.

Objective:

This study examined the geographic distribution of AI adoption among mammography facilities across Texas and evaluated facility and community characteristics associated with adoption before and after accounting for spatial dependence.

Methods:

We conducted a statewide cross-sectional geospatial study of 625 mammography facilities in Texas. We obtained facility characteristics, including AI adoption, 3-dimensional (3D) mammography capability, onsite mammography, and mobile mammography services, from publicly available facility information. We linked community socioeconomic and demographic characteristics using 2020-2024 American Community Survey 5-year estimates at the ZIP Code Tabulation Area level. Global Moran's I statistics using 25-, 50-, 75-, and 100-km distance bands were used to evaluate spatial clustering of AI adoption. Conventional multivariable logistic regression and spatial logistic regression with a Matérn Gaussian random field were used to examine facility and community factors associated with AI adoption. The primary regression analysis included 609 facilities with complete covariate data.

Results:

Among 625 mammography facilities, 214 [34.2%] had adopted AI-assisted mammography. AI adoption demonstrated significant positive spatial autocorrelation across all evaluated distance bands [Moran's I range 0.085-0.121; all P<.001]. In the conventional multivariable model, 3D mammography was strongly associated with AI adoption [adjusted odds ratio [aOR] 31.92, 95% CI 4.35-233.95; P<.001]. Higher median household income [aOR 1.32 per 1-SD increase, 95% CI 1.02-1.71; P=.035], a higher proportion of women aged 50 years or older [aOR 1.04 per 1-percentage-point increase, 95% CI 1.00-1.07; P=.026], and a higher proportion of Hispanic residents [aOR 1.02, 95% CI 1.00-1.03; P=.031] were also associated with AI adoption. After accounting for spatial dependence, 3D mammography remained strongly associated with AI adoption [aOR 39.48, 95% CI 5.25-296.88; P<.001], whereas none of the measured community socioeconomic or demographic characteristics remained statistically significant. The estimated Matérn spatial range was 90.36 km, with a spatial SD of 1.16. Sensitivity analyses using women aged 40 years or older and additionally adjusting for onsite and mobile mammography services produced similar overall findings.

Conclusions:

AI-assisted mammography adoption in Texas was geographically clustered, with residual spatial dependence extending across approximately 90 km. Although several community characteristics were associated with adoption in conventional regression, these associations were attenuated after accounting for geographic dependence. In contrast, 3D mammography remained strongly associated with AI adoption across conventional, spatial, and sensitivity analyses. These findings suggest that pre-existing facility technological capacity may be a more consistent correlate of AI implementation than the measured socioeconomic and demographic characteristics of surrounding communities and highlight the importance of regional technology diffusion in shaping the implementation of AI-enabled breast imaging.