AI Adoption in US Cancer Centers: National Cross-Sectional Study of Institutional and Policy Determinants

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background AI is increasingly being integrated into cancer screening, treatment, and patient care. However, AI adoption across cancer centers varies, raising concerns about unequal access to AI-enabled cancer care. Objective This study examined publicly visible AI adoption among National Cancer Institute (NCI)–designated cancer centers in the United States and assessed whether adoption was associated with institutional characteristics, socioeconomic context, geographic distribution, and state-level political environment. Methods We assembled a national dataset of 75 NCI-designated cancer centers using publicly available sources. AI adoption was measured across 3 domains (screening, treatment, and patient care) and summarized as a composite index (0-3). Spatial clustering was evaluated using Moran I . Ordered logistic regression models examined associations between AI adoption and institutional factors, including physician workforce size, hospital beds, cancer center type, and contextual factors, including population characteristics, socioeconomic indicators, and state political environment. Results Among the 75 cancer centers, the mean AI adoption index was 1.37 (SD 0.86), indicating adoption in approximately 1 to 2 domains on average. Publicly visible AI adoption was most common for screening applications (mean 0.86, SD 0.35), followed by patient care applications (mean 0.50, SD 0.50), and treatment-related applications (mean 0.22, SD 0.42). Moran I showed no statistically significant spatial autocorrelation, suggesting that AI adoption did not follow a clear geographic clustering pattern. In regression models, institutional capacity measures, including physician workforce size and hospital bed capacity, showed positive but generally modest associations with AI adoption. State-level socioeconomic indicators, including income, education, and urbanicity, were not consistently associated with adoption. Political-context findings were mixed and should be interpreted cautiously; Republican Party control was associated with higher AI adoption in the primary adjusted model, while the exploratory interaction model suggested that cancer center type and governance context may jointly shape publicly visible AI adoption. Conclusions AI adoption among US cancer centers appears uneven across clinical domains and is more closely associated with institutional capacity than with surrounding community socioeconomic characteristics. The absence of statistically significant spatial clustering suggests that, if AI diffusion is occurring, it may be driven primarily through nonspatial channels, such as institutional resources, academic networks, vendor partnerships, or policy environments, rather than geographic proximity alone. Because the study relies on public-source reporting, the findings should be interpreted as patterns of publicly visible AI adoption rather than confirmed clinical integration. Future research should use longitudinal designs, direct institutional validation, and more detailed implementation measures to assess whether AI-enabled cancer care is diffusing equitably across institutions and populations.