Clinical governance of artificial intelligence in internal medicine: a literature-informed five-pillar framework for...

Most artificial intelligence (AI) governance frameworks in healthcare address model development, reporting standards, or regulation in broad terms, but do not adequately translate these principles into the operational realities of internal medicine. Although multimorbidity, frailty, polypharmacy, diagnostic uncertainty, incomplete data, and transitions of care are not exclusive to internal medicine, this specialty is characterized by its frequent and simultaneous convergence within the same clinical decision-making process. We developed a literature-informed conceptual framework through a targeted narrative review and interpretive synthesis of methodological, regulatory, implementation, and clinical literature on AI in healthcare. We examined five source domains: reporting and evaluation standards; risk-of-bias and model-appraisal tools; ethical and regulatory guidance; implementation, workflow, and electronic medical record literature; and internal medicine-specific literature on complexity, longitudinal care, and transitions of care. These domains were selected because they correspond to recurrent governance functions required for safe AI use: transparent evaluation, critical appraisal, accountability, workflow integration, data stewardship, and clinical-contextual interpretation. The synthesis highlighted a persistent gap between cross-cutting AI governance instruments and the discipline-oriented governance needs of internal medicine. This gap does not imply that existing frameworks are inadequate, but that their principles require translation into departmental governance structures capable of addressing complex, longitudinal, and multimorbid care. In particular, existing frameworks insufficiently address the interaction between algorithmic tools and the clinical complexity of multimorbid patients, the care-continuum nature of internist work, the role of electronic medical records as both data sources and clinical interfaces, and the growing role of AI in mediating access to scientific evidence. To address this gap, we propose a five-pillar framework comprising: (1) clinical oversight and human-in-the-loop decision-making; (2) data integration oriented to clinical complexity; (3) organizational embedding across the care continuum; (4) ethical, legal, and regulatory governance; and (5) governance of scientific knowledge and AI-mediated evidence access. We also provide an operational checklist to support local implementation readiness. AI should not be introduced into internal medicine as an isolated technological layer, but should be governed as a complex clinical intervention. A discipline-oriented five-pillar model may help departments, hospitals, and scientific societies assess not only whether an AI tool performs well, but whether the surrounding clinical, organizational, data, regulatory, and epistemic conditions are mature enough to support safe use.