The pursuit of medicine-informed artificial intelligence (AI) is undermined by current methods lacking the capacity for continual adaptation to changing operational contexts. Such self-evolving intelligence is crucial in rapidly changing domains like healthcare. AI systems accessing real-time clinical pipelines can dynamically modify their knowledge base or data processing, incorporating newly available data categories, features, relations, or concepts. A self-evolutionary architecture incorporating AI in medicine-informed continual learning is outlined, with supporting infrastructures for data quality, data governance, privacy risk, and benchmarking. Completed short-term research constitutes a first instantiation for real-time risk stratification of patients in acute medical care with guaranteed performance. These developments contribute to the establishment of healthcare AI as continuously self-adapting, incrementally sound, and clinically reliable.

AI applied in healthcare consistently presents with an established and salient conceptual gap, despite ambitious and sophisticated application, deployment, and development efforts. Most operational AI models in healthcare are not continuously self-evolving. Thus they invariably become increasingly misaligned with information-rich, time-variant clinical environments and operations, gradually losing medical relevance and becoming actively misleading—a risk given the considerable and growing influence of such inferences on patient care—under a quasi-Boydian–Hyesque risk governance concept. Assurance of continually adaptive AI in both architecture and clinical performance necessitates an evolutionary impulse for AI.