Kayo Waki

Nutrient estimation from meal images by multimodal large language models can support diabetes self-management, but its robustness is limited by variability driven by both sensitivity to visual presentation and inherent model instability. Accounting for and mitigating this variability is essential for robust AI-assisted dietary assessment.

To reduce meal logging burden in diet interventions, we fine-tuned OpenAI’s GPT-4o on 1269 Japanese meal photographs (train/val/eval: 912/252/105) to estimate nutrients, using weighed food records or dietitian estimates as ground truth, and compared it with 27 non–fine-tuned models and a human dietitian. Non–fine-tuned models did poorly for fiber. Most models did well for carbohydrates, protein, and energy, while performance for salt and fat varied by model. GPT-5.1 (minimal reasoning) and non-fine-tuned GPT-4o models both provided strong accuracy, though not universally better than dietitian performance. The fine-tuned GPT-4o model’s accuracy exceeded that of the dietitian for all nutrients, with the intra-class correlation coefficient for fiber of 0.79 (95% CI 0.782-0.797) greatly exceeded the dietitian performance of 0.68, validating the accuracy of the model.

Introduction and objective: Short sleep duration is associated with worse hemoglobin A _1c (HbA _1c ), and treatment guidelines recommend sufficient sleep duration. However, it is unclear whether extending sleep duration of short-sleeping people with type 2 diabetes causes improved HbA _1c . We aimed to assess the effect on HbA _1c of increasing sleep duration. Methods: We developed a personalized digital intervention based on the theory of planned behavior and conducted a single-blind, two-arm randomized controlled trial with 70 short-sleeping (≤6 hours) people with type 2 diabetes in Japan. Both arms measured sleep duration using an actigraph and a sleep diary. The intervention group received bedtime advancement support, including achievable bedtime goal setting and feedback. The primary outcome was between-arm difference in HbA _1c change after 12 weeks. Results: The baseline average sleep duration and HbA _1c were 5.96 ± 1.26 (hours) and 8.30 ± 0.91 (%) in the intervention group, and 5.94 ± 0.96 (hours) and 8.09 ± 0.51 (%) in the control group, respectively. Sleep duration improved with statistical significance (between-arm difference in change: 32.8 minutes, P = .0042). However, HbA _1c did not reduce significantly (between-arm difference in change: –0.11%, P = .51), and increases in sleep duration were not significantly associated with HbA _1c reduction ( P = .69). Conclusion: In this randomized trial, an average sleep extension of approximately 30 minutes over 12 weeks did not result in a significant improvement in HbA _1c among short-sleeping adults with type 2 diabetes. Further studies with larger sample sizes and greater sleep extension are warranted.

Sharing research code in an open access version-controlled repository offers significant benefits for both science as a whole and for individual researchers. In this article, we focus on this practice, which is fully aligned with the NIH’s Gold Standard Science (GSS) program as well as FAIR (findable, accessible, interoperable, reusable) and TRUST (transparency, responsibility, user focus, sustainability, technology) principles. Gold Standard Science supports open science by emphasizing transparency, reproducibility, and the use of best practices that enable others to verify and extend research. Pairing a research article’s cited data snapshot with a versioned, environment-specific code release, deposited in a companion code repository, ensures that, upon submission to a medical journal, readers and reviewers can directly verify results. An executable and updatable companion code repository complements, rather than replaces, established research data repositories. When code underlying medical research results is made openly available, then other scientists can inspect, run, and validate analyses. These activities enhance reproducibility, which is a core aim of GSS. Shared code also facilitates collaborative innovation by allowing researchers to extend the utility of the code to new datasets and applications. For researchers, code sharing can increase visibility, credibility, and citation impact. Demonstrating transparency through shared executable and updatable code builds trust with journal readers, peer reviewers, funders, and peers. Shared code in an open access repository signals adherence to high standards of scientific integrity and attracts opportunities for collaboration. A researcher who shares code receives recognition as a leader in reproducible, trustworthy research consistent with NIH’s GSS principles.

Background: A high intake of dietary fiber has been shown to improve glycemic control and decrease hyperinsulinemia in people living with type 2 diabetes (T2D). T2D patients in Japan consume less than the recommended amount of fiber. Based on findings from a formative study, we developed an artificial intelligence (AI)-powered mobile health (mHealth) intervention, FiberMore, that uses the theory of planned behavior to help T2D patients increase their dietary fiber intake by enhancing their perceived behavioral control and attitude toward fiber consumption. Objective: We aimed to assess the efficacy of FiberMore in improving the dietary fiber intake of T2D patients by conducting a pilot randomized controlled trial. In addition, we want to explore the efficacy of FiberMore in reducing HbA _1c of T2D patients via improvement in dietary fiber intake. Methods: This is a randomized, single-blinded, multicenter study targeting 80 T2D patients from 3 institutions in Japan with a 2-week run-in, a 12-week intervention, and a 12-week observation. The intervention group is given access to FiberMore throughout the 12-week intervention period. A core feature of FiberMore is AI-powered meal photo logging using a fine-tuned GPT-4o (OpenAI) model, which analyzes the nutrient content of meals and delivers personalized, real-time feedback on fiber content. In addition, FiberMore provides personalized fiber goal setting and supports participants in identifying barriers to increasing fiber intake, along with corresponding coping strategies (labeled as “solutions” to the participant), through an AI chatbot. The AI chatbot also assesses participants’ emotional attitudes toward eating more fiber and delivers relevant educational content on dietary fiber. The control group receives a sham intervention focused on salt reduction, consisting of educational content delivered at 3 time points during the intervention period and records their daily efforts in salt reduction in a diary. The 12-week intervention period will be followed by a 12-week observational period to investigate the sustainability of the intervention’s effects. The primary outcome is between-group difference in the change of dietary fiber intake at 12 weeks. The secondary outcomes include HbA1c, other clinical measures, measurements of behavior changes, and assessment of participants’ satisfaction and perceived usefulness of the intervention. Results: Recruitment began on February 12, 2025, and ended on September 1, 2025. We anticipate that the intervention period will conclude in December 2025 and the observation period will conclude in March 2026. As of September 22, 2025, a total of 72 participants have been officially enrolled and randomized. Conclusions: There are currently no mHealth dietary interventions that specifically focus on increasing fiber intake in Japan, highlighting the novelty of this intervention. This trial will generate important evidence on the efficacy, feasibility, and safety of an AI-based mHealth intervention for enhancing dietary fiber intake and glycemic control in free-living individuals with T2D. Furthermore, as a pilot study, it will offer valuable insights into the development of AI as a promising tool for accurate, low-burden dietary assessment.

Background: Despite the importance of dietary fiber in regulating glycemic control, the reported intake among patients with type 2 diabetes (T2D) in Japan was around 12‐16 g, well below the local official recommended intake of 20 g and above. Recent data is lacking, with the most recent available estimates collected between 2014 and 2019. Most mHealth dietary intervention apps for T2D focus on calorie and carbohydrate outcomes, with limited evidence on fiber intake. Fiber data was collected in a recent 12-month trial of DialBetesPlus, a multimodal diabetes mHealth self-management system that supports dietary behavior change by allowing users to record their meals and provides timely and detailed information on users’ nutrient intake. Objective: This study aimed to assess the pattern of dietary fiber intake among DialBetesPlus intervention users to provide recent data. As a secondary objective, the study explored factors that may influence dietary fiber intake among participants. Methods: Meal records were extracted using the DialBetesPlus app developer dashboard. Dietary fiber intake was measured for all intervention participants. The analysis included only data from participants with complete (breakfast, lunch, and dinner) meal records for at least 7 days. We calculated the average dietary fiber intake and fiber density per day by first averaging across all participants for each day, then averaging these daily values over 1 year. We averaged fiber intake per meal type across all days with available data for each participant, without excluding incomplete meal record days. Results: Out of 66 participants from the intervention group who were assigned to the DialBetesPlus intervention, 47 (71.2%) had at least 7 days of complete meal records and were included in the analysis. A 1-year trend analysis revealed a slight upward trend of daily fiber intake with the rolling average consistently below 18 g. The average fiber intake was 17.1 g/day, with a corresponding mean fiber density of 10.5 g/1000 kcal. The overall mean fiber intake was 17.1 g/day. Separate analysis by meal types revealed that the highest fiber intake was during dinner (6.7 g), followed by lunch (4.8 g), breakfast (4.4 g), and snacks (1.5 g), while fiber density was lowest for snacks (7.8 g/1000 kcal), followed by dinner (10.2 g/1000 kcal), lunch (10.5 g/1000 kcal), and breakfast (10.8 g/1000 kcal). No significant correlations were observed between average fiber intake and participant characteristics such as age, sex, BMI, hemoglobin A _1c , blood pressure, and frequency of meal logging. Conclusions: Despite using a general diabetes self-management mHealth app (DiabetesPlus) that included dietary self-monitoring and basic nutritional feedback, users consumed less than the recommended 20 g/day of dietary fiber on average over a 1-year period. This study highlights the need to explore alternative mHealth strategies to further enhance dietary fiber intake among patients with T2D.

Background: Diabetes affects millions worldwide. Primary care physicians provide a significant portion of care, and they often struggle with selecting appropriate medications. Objective: This study aimed to develop a model that accurately predicts what drug an endocrinologist would prescribe based on the current measurements. The goal was to create a system that would assist nonspecialists in choosing medications, thereby potentially improving diabetes treatment outcomes. Based on the performance of previous studies, we set a performance target of achieving a receiver operating characteristic area under the curve (ROC-AUC) above 0.95. Methods: A transformer-based encoder-decoder model predicts whether 44 types of diabetes drugs will be prescribed. The model uses sequences of age, sex, history for 12 laboratory tests, and prescribed drug history as inputs. We assessed the model using the electronic health records from 7034 patients with diabetes seeing endocrinologists between 2012 and 2022 at the University of Tokyo Hospital. We assessed model performance trained on data subsets spanning different time periods (2, 5, and 10 years) using micro- and macro-averaged ROC-AUC on a hold-out test set comprising data solely from 2022. The model’s performance was compared against LightGBM. Results: The model trained on data from the past 5 years (2017‐2021) yielded the best predictive performance, achieving a microaverage (95% CI) ROC-AUC of 0.993 (0.992-0.994) and a macroaverage (95% CI) ROC-AUC of 0.988 (0.980-0.993). The model achieved an ROC-AUC above 0.95 for 43 out of 44 drugs. These results surpassed the predefined performance target and outperformed both previous studies and the LightGBM model’s microaverage ROC-AUC of 0.988 (0.985-0.990) in terms of prediction accuracy. Furthermore, training the model with short-term data from the past 5 years yielded high accuracy compared to using data from the past 10 years, suggesting that learning from more recent prescribing patterns might be advantageous. Conclusions: The proposed model demonstrates the feasibility of accurately predicting the next prescribed drugs. This model, trained from the past prescriptions of endocrinologists, has the potential to provide information that can assist nonspecialists in making diabetes-treatment decisions. Future studies will focus on incorporating important factors such as prescription contraindications and constraints to enhance safety, as well as leveraging large-scale clinical data across multiple hospitals to improve the generalizability of the model.

Digital therapeutics (DTx) offers a patient-facing method of treatment of type 2 diabetes (T2D). This paper is a broad survey of DTx for T2D, applying DTx definitions and exploring therapeutic and behaviour-change mechanisms. We searched English language sources for candidate DTx through July 2024. We reviewed 64 candidate apps/interventions and found 12 that are DTx or DTx-adjacent varying by mechanism (providing information, inducing lifestyle change and helping with medication) and implementation. Four (mySugr, Dario Health, BlueStar, and Aspyre) met the strictest definition for DTx. There were no approved T2D DTx in east Asia. DTx offers proven benefits, but there are significant gaps in the current state of T2D DTx. Interventions that are anchored in a clear theory of human behaviour are scarce and there is a need for improved patient engagement, perhaps via methods using artificial intelligence.

Background: A 12-week pilot of the StepAdd mobile health (mHealth) behavior change intervention based on social cognitive theory (SCT) saw an 86.7% increase in mean daily step counts among patients with type 2 diabetes. Due to the lack of exploration of theoretical implications in mHealth intervention studies, there is a need to understand the mechanism underlying the behavioral change to inform the future design of digital therapeutics. Objective: This study aimed to examine the SCT drivers underlying the mean increase in exercise among Japanese patients with type 2 diabetes who participated in the StepAdd intervention. Methods: This is a post hoc analysis of data collected in the single-arm pilot study of the 32 patients who completed the StepAdd intervention. The StepAdd app uses self-mastery and coping strategies to increase self-efficacy and thus increase walking. Self-mastery was measured by the goal completion (GC) rate, which is the percentage of days in which patients met these adapting goals. The use of coping strategies was measured by the strategy implementation (SI) rate, which is the percentage of days in which patients applied their selected coping strategies. We assessed correlations between GC, SI, and self-efficacy to increase walking via linear regression and analyzed relationships via structural equation modeling. Results: We found statistically significant support for the SCT approach, including a correlation coefficient (ρ) of 0.649 between step increase and GC rate (P<.001); a ρ of 0.497 between the coping SI rate and self-efficacy increase (P=.004); a ρ of 0.446 between GC rate and self-mastery increase (P=.01); and a ρ of 0.355 between self-regulation increase and step increase (P=.046), giving us insight into why the behavior intervention succeeded. We also found significant correlations between self-efficacy for barriers and self-efficacy for task-specific behavior (ρ=0.358; P=.04), as well as self-regulation and self-efficacy for task-specific behavior (ρ=0.583; P<.001). However, a cross-lagged panel modeling analysis found no significant evidence that changes in self-efficacy preceded behavior changes in line with SCT. Conclusions: Self-mastery and coping strategies contributed to the walking behavior change in StepAdd, supporting the SCT model of behavior change. Future research is needed to better understand the causal pathways proposed by SCT.

[This corrects the article DOI: 10.2196/42607.].