Emi Furukawa

Digital behavior change interventions (DBCIs), including smartphone apps, wearable devices, and AI-driven adaptive systems, have become a dominant approach for promoting changes in health-related behavior. Although meta-analyses have consistently demonstrated short-term effectiveness, a recurring finding is that these effects are attenuated once user engagement declines or the intervention ends. We describe this phenomenon as the sustainability paradox: the same design features that maximize behavioral success during use, such as automated guidance, real-time feedback, and gamified rewards, may undermine the conditions necessary for the behavior to persist once the intervention is terminated. This viewpoint introduces narrative ownership as a conceptual framework for understanding and addressing this paradox. Narrative ownership refers to the extent to which individuals experience behavior change as an ongoing, interpretable, and self-authored process and comprises 3 components: process framing, narrative repair, and authorship. We argue that many contemporary DBCI architectures unintentionally displace these experiential processes by automating lapse correction, redirecting behaviors without interpretation, and externalizing the attribution of success. We map these tendencies onto 3 structural mechanisms—the erosion of process framing, the foreclosure of narrative repair, and the hollowing of authorship—and distinguish narrative ownership from adjacent constructs such as engagement, self-efficacy, and autonomy support. We also note that narrative ownership is double-edged: the same reflective processes can consolidate negative self-attributions, and its expression varies across cultural contexts, taking a familial or communal rather than strictly individual form in relational settings. Therefore, to translate this framework into practice, we outline 3 design principles for sustainable digital interventions: reflective pauses that support process framing, lapse reinterpretation features that enable narrative repair, and meaning-linking interfaces that strengthen authorship. For each principle, we derive testable propositions—for example, that reflective pauses will improve behavioral maintenance after the intervention ends and that meaning-linking interfaces will sustain behavior after app use is discontinued—along with candidate outcome measures and directions for empirical validation. Together, these principles reposition the user not merely as the recipient of behavioral guidance but also as the author of their own story of change in digital health interventions.

Background: The ability to travel is one of the most valued aspects of quality of life among people undergoing maintenance hemodialysis. However, little is known about how people undergoing hemodialysis experience travel. This study aimed to qualitatively explore the significance of travel from this group's perspective by analyzing posts on an online forum, and to provide insights to support patient-centered dialysis care. Methods: We collected posts made between April 1, 2015, and March 31, 2025, on Japan's largest Q&A social media platform, covering hemodialysis and travel. Using reflexive thematic analysis, we inductively coded the content and generated themes and subthemes about the travel experiences of people undergoing hemodialysis. Results: In total, 150 threads with 744 posts were included in the analysis. We identified two overarching contexts: asking questions about living with hemodialysis, and seeking travel-related information. Travel experiences were mainly categorized as either once-in-a-lifetime or an extension of everyday life, with an intermediate theme of travel as a way to reclaim life and freedom. The feasibility of travel was determined by two key factors: the vulnerability of people undergoing hemodialysis and acceptance and cooperation from healthcare providers. Conclusions: For people undergoing hemodialysis, travel is a symbolic opportunity to reclaim autonomy and meaning in their lives. Clinicians should recognize travel as a critical element of quality of life and respect patients' aspirations and independence, while ensuring their safety through active collaboration and coordination among healthcare providers.

Effective heatstroke prevention depends not only on access to information but also on whether online health materials support understanding and actionable decision-making. However, evidence comparing the quality and presentation of online heatstroke information across countries remains limited. We conducted a quantitative content analysis of 120 publicly available heatstroke-related webpages from Japan, the United States, the United Kingdom, and Australia (30 per country), with one World Health Organization webpage included as a reference. Readability, information volume, credibility, and information presentation were assessed using established metrics, including the Patient Education Materials Assessment Tool for Print Materials. Cross-country comparisons and sensitivity analyses focusing on webpages targeting the general public were performed. UK webpages showed higher readability, whereas Japanese webpages more frequently incorporated visual aids and tools that facilitate preventive actions, resulting in higher understandability and actionability scores. However, actionability did not reach adequate levels in any country. Online heatstroke information differs substantially across countries. While readability improves accessibility, greater integration of visual and action-supporting design elements is needed to better promote preventive action.

Background: The accuracy of the latest reasoning-enhanced large language models on national medical licensing examinations remains unknown, which is crucial for determining how close they are to serving as effective knowledge sources for medical education. This study aimed to evaluate the performance of four reasoning-enhanced large language models (LLMs)—GPT-5, Grok-4, Claude Opus 4.1, and Gemini 2.5 Pro—on the Japanese National Medical Examination (JNME), providing insights into their potential as educational resources and their future applicability in medical practice. Methods: We evaluated LLM performance using the 2019 and 2025 JNME ( n = 793). Questions were entered into each model with chain-of-thought prompting enabled. Accuracy was assessed overall and by question type. Incorrect responses were qualitatively reviewed by a licensed physician and a medical student. Results: From highest to lowest, the overall accuracies of the four LLMs were 97.2% for Gemini 2.5 Pro, 96.3% for GPT-5, 96.1% for Claude Opus 4.1, and 95.6% for Grok-4, with no significant pairwise differences. For image-based and non-image-based items, Gemini 2.5 Pro achieved the highest accuracy of 96.1% and 97.6%, with no significant difference, whereas accuracy was significantly lower on image-based items for the other three LLMs. Across difficulty levels, Gemini 2.5 Pro again achieved the highest accuracy (98.4% for easy, 97.3% for moderate, and 93.2% for difficult items). Within each LLM, accuracy on difficult questions was significantly lower than on easy questions. Common error patterns included providing unnecessary additional options in single-choice questions, misdiagnosis of X-ray or computed tomography images (primarily due to confusion regarding left–right laterality), and difficulties in prioritizing appropriate actions in clinical questions with complex contextual information. Conclusions: Four LLMs released in 2025 surpassed the 95% benchmark on the JNME, and their near-perfect (approximately 99%) performance on basic medical knowledge questions highlights substantial potential for use as learning resources in foundational medical education. Gemini 2.5 Pro demonstrated the most consistent performance across question types, while Grok-4 showed greater variability. The concentration of incorrectness in clinical questions indicates that LLMs still require substantial refinement and validation before their use can be extended to clinical reasoning or patient care.

Background: Mpox (monkeypox) outbreaks since 2022 have emphasized the importance of accessible health education materials. However, many Japanese online resources on mpox are difficult to understand, creating barriers for public health communication. Recent advances in artificial intelligence (AI) such as ChatGPT-4o show promise in generating more comprehensible and actionable health education content. Objective: The aim of this study was to evaluate the comprehensibility, actionability, and readability of Japanese health education materials on mpox compared with texts generated by ChatGPT-4o. Methods: A cross-sectional study was conducted using systematic quantitative content analysis. A total of 119 publicly available Japanese health education materials on mpox were compared with 30 texts generated by ChatGPT-4o. Websites containing videos, social media posts, academic papers, and non-Japanese language content were excluded. For generating ChatGPT-4o texts, we used 3 separate prompts with 3 different keywords. For each keyword, text generation was repeated 10 times, with prompt history deleted each time to prevent previous outputs from influencing subsequent generations and to account for output variability. The Patient Education Materials Assessment Tool for Printable Materials (PEMAT-P) was used to assess the understandability and actionability of the generated text, while the Japanese Readability Measurement System (jReadability) was used to evaluate readability. The Journal of the American Medical Association benchmark criteria were applied to evaluate the quality of the materials. Results: A total of 119 Japanese mpox-related health education web pages and 30 ChatGPT-4o–generated texts were analyzed. AI-generated texts significantly outperformed web pages in understandability, with 80% (24/30) scoring ≥70% in PEMAT-P (P<.001). Readability scores for AI texts (mean 2.9, SD 0.4) were also higher than those for web pages (mean 2.4, SD 1.0; P=.009). However, web pages included more visual aids and actionable guidance such as practical instructions, which were largely absent in AI-generated content. Government agencies authored 90 (75.6%) out of 119 web pages, but only 31 (26.1%) included proper attribution. Most web pages (117/119, 98.3%) disclosed sponsorship and ownership. Conclusions: AI-generated texts were easier to understand and read than traditional web-based materials. However, web-based texts provided more visual aids and practical guidance. Combining AI-generated texts with traditional web-based materials may enhance the effectiveness of health education materials and improve accessibility to a broader audience. Further research is needed to explore the integration of AI-generated content into public health communication strategies and policies to optimize information delivery during health crises such as the mpox outbreak.

Study aims and objectives This study aimed to evaluate the performance of four reasoning-enhanced large language models (LLMs)—GPT-5, Grok-4, Claude Opus 4.1, and Gemini 2.5 Pro—on the Japanese National Medical Examination (JNME). Methods We evaluated LLM performance using the 2019 and 2025 JNME (n = 793). Questions were entered into each model with chain-of-thought prompting enabled. Accuracy was assessed overall as well as by question type, content domain, and difficulty. Incorrect responses were qualitatively reviewed by a licensed physician and a medical student. Results From highest to lowest, the overall accuracies of the four LLMs were 97.2% for Gemini 2.5 Pro, 96.3% for GPT-5, 96.1% for Claude Opus 4.1, and 95.6% for Grok-4, with no significant pairwise differences observed. All four LLMs reached the threshold generally regarded as sufficient to serve as reliable medical knowledge sources. Question type (e.g., image-based, clinically oriented, and difficult items) still influenced LLM performance, but the performance gaps were much smaller than in earlier generations of LLMs. Notably, Gemini 2.5 Pro consistently achieved the highest performance, including 96.1% on image-based questions and 97.0% on clinical questions. Common error patterns included providing extra response options and misinterpreting laterality when analyzing X-ray images or computed tomography (CT). Conclusions Advanced LLMs released in 2025 achieved textbook-level accuracy on the JNME, surpassing the 95% benchmark for reliable knowledge sources. Gemini 2.5 Pro achieved the highest accuracy across all question types and demonstrated the greatest stability, while Grok-4 showed more variability. These findings highlight a milestone in which LLMs have achieved the level necessary to be considered educational resources and decision-support tools. Statements and Declarations This work was supported by JSPS KAKENHI Grant Number 24KJ0830.

Background: AI tools are becoming primary information sources for patients with chronic kidney disease (CKD). However, as AI sometimes generates factual or inaccurate information, the reliability of information must be assessed. Methods: This study assessed the AI-generated responses to frequently asked questions on CKD. We entered Japanese prompts with top CKD-related keywords into ChatGPT, Copilot, and Gemini. The Quality Analysis of Medical Artificial Intelligence (QAMAI) tool was used to evaluate the reliability of the information. Results: We included 207 AI responses from 23 prompts. The AI tools generated reliable information, with a median QAMAI score of 23 (interquartile range: 7) out of 30. However, information accuracy and resource availability varied (median (IQR): ChatGPT versus Copilot versus Gemini = 18 (2) versus 25 (3) versus 24 (5), p < 0.01). Among AI tools, ChatGPT provided the least accurate information and did not provide any resources. Conclusion: The quality of AI responses on CKD was generally acceptable. While most information provided was reliable and comprehensive, some information lacked accuracy and references.