The rapid expansion of artificial intelligence (AI) in healthcare has led to increasing adoption of AI-based software as a medical device (SaMD). This paper reviews the current regulatory and approval framework for AI-based SaMD in Japan and discusses emerging challenges associated with generative and adaptive AI technologies. Under the Pharmaceuticals and Medical Devices Act (PMD Act), software intended for diagnosis, treatment, or prevention is regulated as a medical device when classified as Class II or higher, and its clinical utility, performance, and safety are evaluated. While the number of approved AI-based SaMDs has increased, most existing products are task-specific systems supporting clinical decision-making within defined scopes. Recent advances in generative AI introduce novel regulatory issues, including difficulties in defining intended use, evaluating reliability of natural language outputs, and managing continuously evolving performance after market entry. These characteristics challenge conventional regulatory paradigms based on fixed product specifications. In light of ongoing international regulatory developments, key issues include clarifying scope of regulated functions, strengthening lifecycle and change management approaches, enhancing transparency, and improving user literacy. Developing adaptive regulatory frameworks that balance innovation, patient safety, and regulatory clarity will be essential for responsible integration of generative AI into healthcare.

The Pharmaceuticals and Medical Devices Agency (PMDA) continues to face increasing operational demands stemming from growing regulatory complexity, expanding data volumes, and evolving scientific and societal expectations. In this context, the appropriate adoption of generative artificial intelligence has emerged as a potential approach for enhancing operational efficiency while reinforcing scientific rigor and accountability. This article describes the current status of generative artificial intelligence utilization at PMDA, outlines its governance framework, and discusses future perspectives for its sustainable application based on institutional experience, internal policy development, and planned/ongoing proof-of-concept activities conducted within PMDA. We summarize a phased implementation strategy that combines commercially available generative artificial intelligence tools for administrative support with the exploration of large language models in secure internal environments for scientifically specialized tasks. Central to this approach is a governance framework that emphasizes human-in-the-loop decision-making, staged evaluation, information governance, and staff capacity building. We also present practical use cases across information collection, analysis and evaluation, and dissemination activities to illustrate how generative artificial intelligence may support regulatory work without replacing human judgment. In conclusion, PMDA's experience suggests that proactive yet cautious adoption of generative artificial intelligence, grounded in robust governance and organizational learning, can improve productivity and enhance scientific capacity within regulatory authorities while maintaining public trust and institutional accountability.

Medical Adaptive Machine Learning Systems (MAMLS) that continuously update their models using clinical data blur the conventional boundary between therapy and research, prompting the argument that their use should be classified as research and governed by informed consent requirements. Although informed consent remains normatively and legally important, this paper contends that consent-centered ethics faces two structural limitations in the context of MAMLS. First, the irreversibility inherent in deep learning models substantially undermines withdrawability—an important ancillary right of consent—thereby suggesting that consent may be transformed from an instrument of ongoing self-determination into a form of delegation to institutions. Second, the problem of data representativeness and bias shifts the unit of ethical analysis from the individual to the population, creating an "autonomy dilemma" in which respect for individual consent can paradoxically undermine the protection of autonomy at the collective level. Under these conditions, ethical justification must be complemented by, and in some contexts repositioned toward, public trust in institutions. The paper concludes that the ethical challenges surrounding MAMLS cannot be adequately addressed within the framework of research ethics alone, but must instead be taken up within the broader framework of public health ethics, with particular attention to transparency, accountability, and participatory governance.

This article examines the ethical governance of artificial intelligence (AI) use in drug development through joint principles of good AI practice issued by the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA). It argues that the significance of the principles lies in moving beyond AI exceptionalism: AI should neither be uniformly prohibited nor uniformly permitted but assessed in a risk-based manner according to context, purpose, and potential impact across the drug lifecycle. Among the ethical and governance risks associated with AI, this study focuses on two organizational risks that are particularly relevant to implementation. The first is shadow use, in which AI involvement remains insufficiently visible, documented, or reviewed. The second is reliance management. Once AI is integrated into research and regulatory workflows, some degree of reliance is inevitable; however, such reliance must remain conscious, proportionate, reviewable, and supported by meaningful human oversight. Overreliance and deskilling are risks associated with poorly managed reliance. Ethical governance should therefore make AI use visible and reviewable while preserving the practical ability to question, verify, escalate, or set aside AI-assisted outputs.

Artificial intelligence (AI) has advanced rapidly across clinical domains, generating both a growing evidence base and dedicated regulatory frameworks for AI-based software as a medical device (SaMD). This review provides a comprehensive assessment of clinical AI across five major domains—diagnostic imaging, gastrointestinal endoscopy, cardiology and remote patient monitoring, diagnosis of infectious diseases, and an AI-ready data infrastructure—examining Japan's regulatory framework, approved device portfolio, and research contributions in an international context. We reviewed literature published between 2019 and 2026, using Japan's regulatory trajectory, approved device portfolio, and domain-specific research output as the primary lens for international comparison and prioritizing prospective studies, multicenter trials, and real-world implementation reports. The state of evidence varies markedly across the domains examined: endoscopy AI has the strongest randomized trial base, while diagnostic imaging AI has seen a systematic decline in real-world performance despite large-scale regulatory approval. Across the three dimensions examined, Japan has a distinctive profile: its strengths are a regulatory and clinical deployment infrastructure—evince by an established program medical device pathway and among the world's highest densities of diagnostic imaging systems and endoscopy volumes—while the data infrastructure lags, constrained by limited open-access resources relative to programs such as The Cancer Imaging Archive and the European Health Data Space. Large language models and generative AI, falling largely outside existing SaMD frameworks, carry the risk of hallucinations and gaps in oversight that healthcare systems in Japan and abroad are only beginning to address. Japan's established program medical device regulatory pathway, high-volume clinical deployment infrastructure, and proven regulatory-approval-to-reimbursement pathway provide a strong foundation for clinical AI adoption; post-approval change management frameworks and clinical accountability mechanisms need to be strengthened, AI-ready data accessibility needs to be expanded, and validated tools need to be embedded within reimbursed clinical workflows to translate this foundation into internationally competitive AI development and deployment.

Clinical data management (CDM) is central to the quality of clinical research. In Japan, CDM faces a shortage of qualified personnel, particularly in academic research organizations (AROs), as well as increasing data volume and complexity. Rapid advances in artificial intelligence (AI), especially large language models, have therefore attracted attention as a way to support CDM. This review summarizes domestic and international examples of AI utilization in CDM-related tasks, including data cleaning, medical coding, and query generation. Across the cases reviewed, a common implementation principle emerged: a human-in-the-loop design in which AI performs initial processing or detection, while final judgment remains with human personnel. This design is especially relevant to AROs, where high data quality must be maintained with limited CDM human resources. Regulatory frameworks, including ICH E6 (R3) and the FDA-EMA Guiding Principles, are beginning to address AI use, but how AI-aided processes should be handled under Good Clinical Practice remains under discussion. Comprehensive risk mitigation is therefore essential. AI and data are interdependent: better data improve AI performance, and better AI can further improve data quality. The shift from manual processes to human-AI collaborative workflows is likely to accelerate, and CDM must develop the technical, regulatory, and risk-management frameworks needed to support that transition.

Artificial intelligence (AI) in dermatology has moved beyond the early paradigm of single-image classification. Dermatological diagnosis is achieved based on morphology, distribution, symptoms, tactile findings, temporal evolution, patient history, histopathology, and treatment response. Clinically important differentials, such as eczema versus psoriasis, cutaneous T-cell lymphoma versus chronic dermatitis, drug eruption versus viral exanthem, lupus erythematosus versus dermatomyositis, and melanoma versus atypical nevus, are rarely resolved with one photograph alone. This review therefore frames AI-assisted dermatology around a central argument: the field must progress from lesion recognition to dermatology-specific multimodal clinical reasoning. We summarize major advances in convolutional neural networks, dermoscopic benchmarks, clinical-image datasets, large language models, vision-language systems, and dermatology foundation models. We also analyze challenges that are particularly relevant to dermatology, including morphologic overlap, skin-tone bias, reduced erythema visibility on darker skin, dataset imbalance, variable smartphone imaging, imperfect reference standards, and the gap between benchmark performance and clinical deployment. Special attention is given to fairness, regulatory oversight, software as a medical device, human-AI collaboration, prognosis prediction, biologic-response modeling, longitudinal monitoring, and treatment optimization. Finally, we discuss future directions, including skin-tone-aware foundation models, lesion-level and body-site grounding, pathology-genomics integration, dermatology copilots, post-marketing surveillance, and prospective clinical trials. By prioritizing dermatological reasoning rather than generic AI architecture, this review outlines a clinically grounded pathway for building safe, interpretable, equitable, and useful AI systems for skin disease management.

The objective of this study was to develop an automated deep learning-based method for the assessment of shoulder balance in adolescent idiopathic scoliosis (AIS) patients using X-ray images in order to provide a reliable and efficient alternative to manual measurements. A total of 940 AIS radiographs were screened; 937 cases were included in the model-development cohort after quality control and were annotated for precise identification and segmentation of the T1 vertebra, both clavicles, and both coracoids. A deep learning neural network was used to segment these structures. Landmarks were extracted based on morphological image processing, and shoulder balance parameters including the clavicle angle (CA), coracoid height difference (CHD), clavicle tilt angle difference (CTAD), radiological shoulder height (RSH), and T1 tilting angle (T1TA) were calculated. The accuracy of the automated measurements was validated using an external dataset (n = 70) assessed by three senior spinal surgeons. The deep learning neural network achieved reliable segmentation performance for foreground anatomical structures, with macro-average intersection over union (IoU) values of 0.77 and 0.73 and Dice coefficients of 0.87 and 0.84 in the internal and external validation datasets, respectively. In the external dataset, the automated measurements displayed a high level of agreement with observer-averaged measurements, with intraclass correlation coefficients ranging from 0.964 to 0.994. Bland–Altman analysis revealed small mean biases across the five shoulder balance parameters, and 90.0 to 98.6% of automated measurements were within the range of interobserver variability. The proposed method provides an efficient and reproducible approach for radiographic shoulder balance assessment and may help reduce observer-dependent measurement variability.

Loneliness and social isolation are important psychosocial concerns in dementia care, but they are difficult to address through pharmacological treatment or episodic social activities alone. Companion technologies, including socially assistive robots, humanoid and conversational robots, avatars, virtual agents, and artificial intelligence (AI)-enabled companions, are increasingly discussed as possible aids for engagement and interaction. However, current evidence does not justify treating these technologies as standalone interventions that directly reduce loneliness in people with dementia. We argue that these technologies may be more appropriately considered not only as devices or interventions, but also as potential components of social care infrastructure: tools that can encourage individual engagement and mediate human relationships but that need to be integrated into care ecosystems. We propose the multi-level care with social technologies (MCST) model, which consists of three connected layers: individual engagement, relational interaction, and the care ecosystem. The model emphasizes that loneliness-related support is produced through the interaction between technology, human facilitation, care workflows, ethical governance, and feedback-based adjustment. This approach is especially relevant as dementia care systems face workforce constraints and an increasing need for home-based care, while psychosocial needs remain and may be overlooked in routine care. AI-enabled and large-language-model-based companions may expand possibilities for personalization and continuous engagement, but dementia-specific evidence remains preliminary and safety concerns are substantial. Future research should validate the MCST model in real-world dementia care and establish evaluation frameworks that address psychosocial outcomes, sustained use, safety, privacy, human oversight, and accountability.

Primary liver cancer presents substantial management challenges across the surgical trajectory, including high recurrence rates, prolonged rehabilitation, and fragmented post-discharge care. This correspondence presents a multidisciplinary physician–nurse co-led, artificial intelligence (AI)-assisted full-course case management model grounded in just-in-time adaptive intervention (JITAI) theory. The model spans four phases—peri-admission, perioperative, post-discharge home-based care, and long-term follow-up—supported by an intelligent platform enabling automated decision triggering, symptom monitoring, tailored health education, and intervention matching. A pilot study with 25 patients was conducted from March to May 2026 at a tertiary cancer hospital in Tianjin, China. Preliminary results revealed improvements in antiviral medication adherence (80–96%), targeted therapy adherence (96–100%), and satisfaction (99.8%), with reductions in missed follow-ups and symptom reporting delays. Multicenter controlled studies need to be conducted to evaluate this model's effectiveness, cost-effectiveness, and long-term sustainability.

Unplanned removal of life-sustaining tubes in intensive care units (ICUs) poses serious risks, yet existing monitoring methods relying on physical restraints have ethical and clinical drawbacks. Here we applied artificial intelligence (AI)-based pose estimation using MediaPipe to analyze ICU surveillance videos, extracting skeletal coordinates to detect movements associated with tube removal. Using Singular Spectrum Transformation for change-point detection, we identified movement changes corresponding to tube-removal behaviors in three consented cases, achieving average precision values substantially above chance. These preliminary results demonstrate that AI-driven, contactless motion analysis can capture clinically relevant signals from existing ICU infrastructure without additional patient burden. Although limited by sample size and environmental factors, this approach holds promise for real-time, non-invasive monitoring to reduce reliance on physical restraints and enhance patient safety in critical care settings.