A Comprehensive Risk Analysis Framework for Medical AI: A Mixed-Methods Approach

Artificial intelligence (AI) is set to transform traditional healthcare delivery and patient care. However, this transformation presents a range of challenges in medical AI decision-making. To address these challenges, this study aims to develop a risk analysis framework for medical AI. A mixed-methods approach is adopted, combining quantitative and qualitative analyses. First, Latent Dirichlet Allocation (LDA) topic modeling is used to analyze 1618 news articles from a leading health information website in China, and this identifies eight risk attributes (i.e., privacy and security, social bias, decision-making status, data quality, algorithm inference, system performance, liability attribution, and regulatory legislation). Second, to identify the risk attributes of medical AI from the patient's perspective, semi-structured interviews with 21 patients and analysis using grounded theory were employed, identifying six key risk attributes of concern to patients from the initial set of eight. Lastly, an online survey of 396 patients was conducted, followed by a Choice-Based Conjoint (CBC) analysis to assess patient preferences in relation to these six risk attributes. The results show that patients prioritize risks in the following order: Data quality (30.20%), privacy and security (29.50%), social bias (19.10%), system performance (13.70%), liability attribution (6.87%), and algorithm inference (0.59%). This study proposes a risk analysis framework that offers practical insights for healthcare policymakers, medical AI developers, and risk analysts.