Factors Influencing Adoption of Large Language Models in Health Care: Multicenter Cross-Sectional Mixed Methods Observational Study
Xiongwen Yang, Yi Xiao, Di Liu, Huiyin Deng, Jian Huang, Yubin Zhou, Maoli Liang, Longyan Dong, Zihao Yuan, Jing Yao, Wankai Guo, Chuan Xu
Abstract
Background: Large language models (LLMs) such as ChatGPT are transforming how health information is accessed, communicated, and applied. However, their adoption in health care remains limited by uncertainties surrounding trust, privacy, and digital readiness, particularly in low- and middle-income contexts. Objective: This study aimed to examine how trust, information behavior, and sociotechnical readiness influence the willingness of health care professionals (HCPs) and patients or caregivers (PCs) in China to adopt LLMs for medical information and decision support. Methods: We conducted a multicenter, cross-sectional mixed methods observational study across five tertiary hospitals, combining surveys of 240 HCPs and 480 PCs with semistructured interviews (n=30). Quantitative analyses included logistic regression (LR), random forest (RF), and extreme gradient boosting models with Shapley additive explanations-based interpretability. Qualitative data were thematically analyzed to capture role-specific concerns and expectations. Unlabelled: Among HCPs, mean age 39.9 (SD 6.5 years; 159/240, 66.2% physicians), 69.2% (166/240) were aware of LLMs and 36.7% (88/240) had previous experience. Among PCs (mean age 50.1, SD 12.6 years; 242/480, 50.4% male), only 26% (125/480) had previous exposure. Trust, perceived usefulness, and digital readiness were the strongest facilitators of adoption. Multivariable models identified trust as the dominant predictor for both groups (HCPs: odds ratio [OR] 3.78, 95% CI 2.15-6.63; PCs: OR 36.34, 95% CI 18.41-71.74; P<.001). For HCPs, previous use (OR 5.61, 95% CI 3.02-10.44; P<.001) and legal clarity (OR 1.56, 95% CI 1.07-2.27; P=.02) increased willingness, while privacy concerns reduced it (OR 0.72, 95% CI 0.53-0.97; P=.03). Among PCs, perceived usefulness (OR 2.01, 95% CI 1.52-2.67; P<.001), education, and digital tool use were positive predictors. Model performance was high (area under the receiver operating characteristic curve [AUC] 0.83-0.85 for HCPs and 0.93-0.96 for PCs). Qualitative findings identified 11 themes: HCPs stressed workflow integration and accountability, while PCs emphasized comprehensibility, reassurance, and equitable access; trust consistently linked technical credibility with social legitimacy. Conclusions: Adoption of LLMs in health care depends less on algorithmic performance than on the management of trust, literacy, and institutional readiness. Trust functions as a multidimensional construct rooted in transparency, reliability, and contextual validation. Theoretically, this study extends technology adoption frameworks by embedding ethical trust, digital literacy, and institutional support within a unified sociotechnical readiness model, advancing information management theory beyond performance-centric paradigms. Empirically, trust and perceived usefulness outweighed demographic or structural factors, with predictive accuracy exceeding 0.9 across user groups. Practically, these findings offer actionable guidance for the design and governance of artificial intelligence systems, emphasizing role-sensitive interfaces, plain-language communication, and transparent accountability mechanisms to promote equitable and trustworthy adoption.