Portraying Ethical Risks of Medical AI: Mixed Methods Study From Connotation Definition to a Survey on Physicians' Cognition.

Background: Ethical risks of medical artificial intelligence (AI) are a global concern, but existing understanding remains fragmented without an integrated framework, and physicians' awareness of these ethical risks is unclear. Objective: This study aimed to construct a multidimensional ethical risk framework for medical AI in the Chinese context, assess physicians' perceptions of these risks, and provide theoretical support for AI risk governance. Methods: In the first phase, we conducted semistructured interviews with 36 experts (102,000-word transcript), analyzed via grounded theory (NVivo 11 [Lumivero]), yielding 5 main risk categories and 15 subcategories. In the second phase, a 21-item questionnaire based on this framework was administered to 600 physicians across 19 Chinese provinces. After the reliability and validity of the questionnaire, descriptive statistics, and multiple linear regression identified risk perceptions and influencing factors. Results: The framework includes physiological risks (eg, diagnostic error and improper treatment), psychological risks (eg, physicians' technical anxiety and patient's psychological anxiety), data and privacy risks (eg, privacy leakage and data security), social risks (eg, trust crisis, occupational impact, unclear liability, and autonomy erosion), and economic and sustainability risks (eg, increased financial burden, resource waste, environmental pollution, and energy consumption). Physicians (n=600) showed the highest concern for data and privacy risks, ambiguous accountability, and a physician-patient trust crisis. Economic and sustainability risks received the lowest agreement. Multiple linear regression identified significant predictors for risk perception. Specialized AI training was positively associated with perceptions of misdiagnosis risks (β=0.230, 95% CI 0.031-0.429; P=.02), privacy leaks (β=0.220, 95% CI 0.041-0.399; P=.02), and unclear liability (β=0.285, 95% CI 0.110-0.460; P=.002). The establishment of medical institution AI ethics review procedures was positively associated with perceptions of diagnostic errors (β=0.355, 95% CI 0.141-0.569; P=.001) and unclear liability (β=0.390, 95% CI 0.200-0.580; P<.001), while AI unfamiliarity was negatively associated with trust crisis (β=-0.260, 95% CI -0.450 to -0.070, P=.006). Conclusions: This study proposes a contextualized ethical risk framework for medical AI in China to guide targeted governance. It is recommended that future efforts should focus on enhancing the ethical training of medical professionals, improving the ethical review mechanisms for AI in health care institutions, and clarifying the division of liabilities and accountability. These measures will promote the robust development of medical AI within an ethically compliant framework.