Medical data can be categorized into structured, unstructured, time-series, image/video, and multimodal data, each of which imposes distinct constraints and offers specific opportunities for artificial intelligence (AI) model design, learning strategies, and validation methods. Recent studies have demonstrated that multimodal approaches significantly enhance predictive performance and clinical usability compared with single-modality models, while emphasizing the critical importance of data quality, governance, and interoperability. This review summarizes the major learning paradigms employed in medical AI, including supervised, unsupervised, semi-supervised, self-supervised, transfer, and reinforcement learning, from the perspectives of data characteristics and clinical context. Labeling costs, data bias, and the need for explainability play a decisive role in the selection of appropriate learning strategies. In particular, self-supervised learning and large-scale language models are increasingly recognized as central drivers of future developments in medical AI. The roles of medical AI are categorized into expertimitating AI, which supports and reproduces expert clinical judgment, and exploratory (emergent) AI, which transcends human empirical limitations by deriving optimized data-driven policies. Moreover, medical AI is expanding beyond diagnostic and therapeutic support into healthcare administration, education, and research, and this review proposes that its ultimate evolution should be toward augmented intelligence grounded in human–AI collaboration.
Hun‐Sung Kim (Mon,) studied this question.