Journal of Medical Internet Research

The Impact of Chatbot Type and Normative Messaging on Chatbot Usage Intention Based on the Health Technology Acceptance Model: Randomized Controlled Trial

2026-09-15T17:30:13-04:00

Background: Digital health tools, such as health chatbots, may improve access to scalable health support, but adoption remains inconsistent. Existing models do not fully integrate technology acceptance factors with health motivation factors relevant to digital health use. Objective: This study proposed and tested the health technology acceptance model and examined whether normative message framing and chatbot type were associated with health motivation, technology acceptance, and intention to use a health chatbot. Methods: In October 2025, we conducted a 4 × 2 between-participants online experiment with 1000 US adults recruited from a nationally representative YouGov panel. Participants were randomized to 1 of 8 conditions varying norm message type (self-oriented, peer-oriented, expert-oriented, or family-oriented) and chatbot type (AI-powered or rule-based) in a cancer prevention and genetic risk information scenario. Outcomes included descriptive norms, injunctive norms, perceived susceptibility, perceived severity, perceived benefits, self-efficacy, perceived ease of use, trust, privacy concerns, and usage intention. Data were analyzed using a multivariate ANOVA with Bonferroni-adjusted post hoc tests and multiple linear regression. Results: Peer-oriented and family-oriented messages produced higher usage intention than expert-oriented messages, and peer-oriented messages also increased descriptive norms, injunctive norms, self-efficacy, and trust. AI-powered chatbots were associated with higher usage intention (=.02) and greater trust (=.008) than rule-based chatbots. In regression analyses, the model explained 50.8% of the variance in usage intention. Usage intention was positively associated with descriptive norms (β=0.087; =.003), injunctive norms (β=0.078; =.009), perceived susceptibility (β=0.051; =.03), perceived benefits (β=0.253; <.001), and trust (β=0.33; <.001), and negatively associated with perceived severity (β=–0.047; =.049) and privacy concerns (β=−0.11; <.001). Perceived ease of use and self-efficacy were not significant predictors. Conclusions: The health technology acceptance model was a useful framework for explaining the intention to use a health chatbot by combining technology acceptance and health motivation constructs. Both social design features and chatbot design features shaped adoption-related beliefs, with peer-oriented and family-oriented framing and AI-powered chatbots showing particular promise. Trust and privacy concerns remained central determinants of intended use.

Study of Human Papillomavirus Vaccine–Related Social Media Content Across Eight Social Media Platforms: Scoping Review of Communication Content, Information Sources, and Analytical Approaches

2026-09-15T17:00:19-04:00

Background: Social media has become an important channel for health information dissemination and public discussion of human papillomavirus (HPV) vaccination. Previous reviews have examined social media and HPV vaccination, often focusing on single platforms, broader HPV-related topics, or the potential effects on knowledge, leaving limited cross-platform synthesis specifically focused on HPV vaccine–related communication content, information sources, and analytical approaches. Objective: This scoping review aimed to map and synthesize peer-reviewed studies examining HPV vaccine–related social media content, with particular attention to communication content, information sources, and methodological characteristics. Methods: Following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines, we searched PubMed, Web of Science, MEDLINE, Embase, and Scopus on June 10, 2026. Backward and forward citation searches were completed on July 17, 2026. We included peer-reviewed English-language original studies published between 2006 and 2026 that analyzed HPV vaccine–related communication content on social media. Data were charted using a structured extraction form and synthesized descriptively. No formal risk-of-bias assessment was conducted. Results: A total of 50 studies published between 2012 and 2026 met the eligibility criteria. Eight social media platforms were represented, with X (formerly Twitter; X Corp) most frequently studied, followed by Weibo (Weibo Corporation), and YouTube (Google LLC). HPV vaccine–related communication content was synthesized into 5 categories: risk- and concern-oriented, benefit- and prevention-oriented, misinformation- and distrust-related, experiential and narrative, and practical and policy-related communication. Across studies, prevention and benefit messages coexisted with safety concerns, side effects, access barriers, misinformation, distrust, and personal narratives. Personal accounts were often associated with experiential narratives or negative sentiment, whereas professional sources were less common but generally provided more informative or higher-quality content. Analytical approaches included content analysis, sentiment analysis, topic modeling, and network analysis, but theoretical frameworks were applied in only a minority of studies. Conclusions: The evidence was limited by heterogeneity in platforms, sampling strategies, data collection periods, coding schemes, and analytical methods, limiting direct comparison across studies. Because most studies analyzed publicly available social media content, this review could not determine individual-level exposure, interpretation, vaccination intention, or vaccine uptake. Overall, HPV vaccine discourse cannot be adequately understood through sentiment polarity alone. This review provides a cross-platform, content-focused synthesis of how HPV vaccine–related social media communication has been studied. Unlike previous reviews, this review clarifies how HPV vaccine–related content has been described, who generates it, and which methods have been used to analyze it. These findings support theory-informed, platform-sensitive research and may inform public health communication strategies, including evidence-based education, transparent risk communication, credible source engagement, narrative-informed messaging, misinformation monitoring, and improved information quality. Trial Registration: OSF Registries 10.17605/OSF.IO/B89FR; https://osf.io/b89fr/overview

Trust in Digital Sexual Health Information Among Chinese College Students: Qualitative Study

2026-09-15T16:30:34-04:00

Background: Digital platforms are primary sources of sexual health information for young adults, yet they are widely perceived as unreliable. This tension between distrust and reliance remains undertheorized. Objective: This study aimed to examine how Chinese college students use online sources for sexual health information under conditions of uncertainty and to explore how they navigate the tension between distrust and reliance. Methods: We conducted semistructured in-depth interviews with 22 Chinese undergraduate students (aged 18-25 years) recruited from 8 universities across China using a purposive sampling strategy with maximum variation, complemented by snowball sampling. Interviews were conducted in January 2026. Telephone interviews explored participants’ experiences with online sexual health information seeking, including their perceptions of reliability, encounters with contradictory content, credibility evaluation strategies, management of uncertainty, the tension between distrust and continued reliance, and the role of cultural factors in shaping information-seeking behaviors. Data were analyzed using reflexive thematic analysis. Results: Participants reported frequent reliance on online sexual health information despite widespread concerns about its reliability. Four interconnected themes were identified: (1) a fragmented and unreliable information landscape characterized by abundant but superficial, contradictory, and difficult-to-apply content; (2) incomplete trust yet continued reliance, driven by precautionary risk logic and limited offline alternatives; (3) strategies for constructing confidence through cross-platform verification, experiential validation, and platform-specific evaluation; and (4) hidden costs of navigation, including confusion, information overload, and search abandonment. Conclusions: Based on these findings, we propose functional trust as an interpretive conceptualization describing how users act on information with sufficient confidence despite persistent uncertainty—as an interpretive lens for understanding how users navigate digital health information. Our findings suggest that the use of online sexual health information does not depend on trust in the conventional sense, but may involve the construction of functional trust under conditions of persistent uncertainty. Addressing challenges in digital sexual health communication requires not only improving information quality but also enhancing its usability and reducing the cognitive burden of navigation within structurally constrained environments.

Quality of Life of People Living With Dementia Residing in Nursing Homes: Secondary Analysis of Observational Data

2026-09-15T15:30:14-04:00

Background: Quality of life (QoL) plays a crucial role in dementia care; however, QoL and its dynamic, context-dependent nature can be difficult to capture among people living with dementia due to challenges in memory and communication, and limitations of self-reported QoL instruments. Observational tools such as the Maastricht Electronic Daily Life Observation (MEDLO) provide narrative descriptions of the daily life of people living with dementia in nursing homes. However, the MEDLO tool was not developed to assess QoL specifically, and it remains unclear to what extent its narrative descriptions reflect aspects of QoL. Analyzing these narrative descriptions is labor-intensive and time-consuming. Recent advances in natural language processing, including large language models (LLMs), offer the potential to analyze these narrative descriptions at scale. Objective: The study aims to explore whether an LLM can be used to structure existing MEDLO narrative data into interpretable QoL-relevant patterns in people living with dementia. Specifically, this study examines whether N-gram analysis, sentiment analysis, and LLM-based topic modeling can identify recurring language patterns, emotional tone, and semantic clusters that can be mapped to Lawton QoL domains. Methods: This study conducted a secondary analysis of existing MEDLO observational data from 151 people living with dementia residing in Dutch long-term care. Narrative data had been documented by trained observers, describing activities, interactions, settings, and emotional expressions. For analysis, a local secure pipeline was developed in which GPT-4o-mini was deployed. The pipeline comprised three analytical steps: (1) N-gram frequency analysis, (2) sentiment analysis, and (3) topic modeling. Prompts were iteratively refined through prompt engineering. Coauthors and domain experts reviewed outputs for coherence, contextual plausibility, and relevance to long-term care practice. Results: A total of 5622 narratives (50,106 words) from 151 people living with dementia were analyzed. The narratives were short, averaging 10.5 (SD 5.80) words per narrative. N-gram frequency analysis identified the frequent documentation of passive activities () in limited indoor settings (). Emotional well-being was often described in positive terms ( and ), whereas explicitly negative expressions ( and ) occurred less frequently. Weighted sentiment analysis showed that, although fewer in number, negative expressions carried a stronger intensity, resulting in an overall predominance of negative sentiment across all QoL domains. Topic modeling identified 8 coherent clusters, most of which mapped onto multiple QoL domains, underscoring QoL’s multidimensionality. Conclusions: LLM-based analyses identified predominantly passive activities with little variation in indoor settings, while people living with dementia were often described as having positive affect. This exploratory study suggests that LLM-based analyses may help structure observational narratives into QoL-relevant patterns, but further validation is needed before such outputs can inform person-centered care practice.

Evaluation of a National Health Service Machine-Learning Model for Hypertension Case-Finding: Retrospective Cohort Study

2026-09-15T15:30:14-04:00

Background: Hypertension is a leading preventable cause of cardiovascular disease, yet a substantial proportion of adults remain undiagnosed, limiting opportunities for early intervention. A predictive model was commissioned by the North West London (NWL) Integrated Care Board to identify undiagnosed hypertension. The model was developed using health records from the Whole Systems Integrated Care (WSIC) database. Objective: We aimed to independently evaluate the predictive performance of the model as it would be encountered in deployment, how performance varied by demographic characteristics, and practical utility. Methods: To evaluate the predictive model, we conducted a retrospective cohort study of 1,802,920 individuals aged 16 years or older, registered with a general practice in NWL, and with no prior diagnosis of hypertension from May 2023 to May 2024. We assessed the model’s predictions against recorded hypertension status using medical diagnoses and blood pressure records. Logistic regression models were used to assess the sensitivity and specificity of the model’s predictions by sociodemographic groups. We also compared the model’s performance against a more interpretable regression approach. Results: The model yielded an overall sensitivity of 62.7% (95% CI 62.5-62.8) and specificity of 60.7% (95% CI 60.5-60.8). Positive predictive value ranged from 31.5% (95% CI 31.2-31.8) to 42.9% (95% CI 42.5-43.2), and negative predictive value ranged from 77.6% (95% CI 77.2-77.9) to 84.9% (95% CI 84.7-85.2). Sensitivity was higher in older adults and Black patients; specificity was higher in younger adults, female patients, and White patients. Overall, sensitivity was higher for those living in areas of higher socioeconomic deprivation, while specificity was lower. These effects plateaued in the 2 least deprived quintiles of deprivation, which were comparable in both sensitivity and specificity. Predictions varied by age, with 96.2% (58,951/61,281) of those aged 70 to 79 predicted to have hypertension, whereas 0.08% of those aged 20 to 39 were predicted to have the condition. The model’s performance was comparable with a more interpretable logistic regression model. Conclusions: Despite the model’s relatively good performance for those without hypertension, the positive predictive value was low, and a significant proportion of true cases remained undetected. Furthermore, there was considerable variation in performance associated with demographic characteristics, suggesting tailored approaches to case-finding may be beneficial in ensuring equity across demographic groups. Especially given the importance of understanding possible biases in predictive models, we recommend that, where there is no loss in performance, more parsimonious, transparent models be selected for prediction in health care settings. The findings of this evaluation can guide the practical application of the model, inform enhancements, direct targeted screening initiatives, and support cost-benefit analyses for broader implementation to improve hypertension management.

AI Health Services and Health Satisfaction Across Socioeconomic Groups in South Korea: National Cross-Sectional Study Using an Instrumental Variable Approach

2026-09-15T11:15:10-04:00

Background: AI-enabled digital health services are rapidly expanding within health care systems and are expected to improve health management and access to health information. However, rigorous empirical evidence on whether AI health service use is associated with individual health satisfaction remains limited, particularly regarding whether these potential benefits differ across socioeconomic groups. Objective: This study aims to examine the relationship between AI health service use and health satisfaction and to assess whether this association varies across socioeconomic groups. Methods: Nationally representative data from the 2024 Digital Divide Survey in South Korea (N=15,000) were analyzed. To address potential endogeneity arising from self-selection and reverse causality, a 2-stage least squares instrumental variable approach was used. Robustness analyses using an alternative sample restriction, an alternative estimation method, and alternative instrumental variable specifications were conducted to assess the robustness of the findings. Subgroup analyses and interaction tests were conducted to assess socioeconomic heterogeneity. Results: AI health service use was positively and significantly associated with health satisfaction (β=0.739, 95% CI 0.391‐1.088; <.001). The positive association was stronger among men, individuals living outside the capital area, those with lower income, people living alone, and individuals with disabilities. Conclusions: AI health service use was associated with higher levels of health satisfaction, and this association appeared stronger among several socioeconomically disadvantaged groups. These findings suggest that AI health services may function as complementary health resources and highlight the importance of considering socioeconomic heterogeneity when developing and evaluating AI-enabled health policies.

Attitudes Toward Large Language Models in Health Care and Preferences for Their Adoption and Oversight Among Health Care Professionals: Cross-Sectional Survey

2026-09-15T10:15:03-04:00

Background: Large language models (LLMs) are rapidly entering health care, but limited empirical data exist on health care professionals’ perceptions. Understanding health care professionals’ attitudes is essential for responsible implementation as LLMs transition from experimental to routine tools. Objective: To characterize health care professionals’ perspectives on LLM use in health care, including exposure, knowledge, perceived clinical utility, safety and bias concerns, and oversight preferences. Methods: This cross-sectional survey was distributed online through a health care news platform mailing list. A total of 335 health care professionals responded, including attending physicians (n=230, 68.7%), residents or fellows, nurse practitioners, physician assistants, and researchers. Most were aged 30 to 59 years (n=243, 72.5%) and practiced in the Northeast United States (n=261, 77.9%). Outcomes included LLM use patterns, knowledge levels, perceived applications, safety and bias concerns, and preferences for regulatory oversight. Analyses included descriptive statistics, Wilcoxon rank-sum tests, ² tests, and Spearman correlations. Results: Of 335 participants, 62.7% (n=210) reported current or contemplated LLM use. Users reported significantly higher self-reported knowledge than nonusers (<.001). Age was not associated with knowledge (ρ=–0.072; =.19). Participants identified literature review (n=246, 73.4%), decision support (n=191, 57%), and patient communication (n=184, 54.9%) as the most valuable applications. Concerns included decision errors (n=253, 75.5%) and algorithmic bias (n=245, 73.1%); nearly all respondents (n=323, 96.4%) expressed concern about bias, and those who had observed bias reported higher concern levels (<.001). Participants favored regulation by professional associations (n=219, 65.4%) over technology companies (n=97, 29%), with 87.8% (n=294) supporting professional guidelines. Confidence in existing oversight was low, with 66.6% (n=223) reporting none. Conclusions: In this exploratory convenience sample, health care professionals reported early adoption of LLMs for lower-risk tasks while expressing concerns about safety, bias, and governance. Given the low response rate and recruitment through a health care innovation–focused mailing list, these findings may not reflect the views of the broader health care professional population. Respondents preferred professional organizations over industry for oversight and suggested that successful integration of LLMs into health care will require careful planning, human supervision, transparent disclosure, and auditing. Future studies using more representative sampling methods are needed to better characterize health care professionals’ attitudes toward LLMs.

Temporal Analysis of Patient-Centered Sentiment in Clinical Notes for Patients With Mental Health Conditions: Retrospective Cohort Study

2026-09-14T16:30:46-04:00

Background: Clinical notes offer rich, longitudinal insights into patient health trajectories. However, existing clinical sentiment analysis primarily evaluates overall note tone rather than the patient’s distinct perspective. This gap is particularly critical in mental health care, where patient-perspective sentiment closely correlates with severe clinical outcomes, including mortality. Objective: This study aimed to characterize temporal sentiment trajectories and changes in mental health clinical notes. We are specifically interested in the problem from the patient perspective, as this can differ from the overall and the provider-perspective sentiment. Furthermore, we examine its association with postdischarge mortality to check if there is any correlation. Using large language models (LLMs) and lexicon-based methods, we compared and highlighted the strengths and weaknesses of both approaches. Methods: We conducted a retrospective analysis of 16,447 clinical notes from 6,382 patients from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, focusing on the Brief Hospital Course and Discharge Instructions sections for patients with ICD-10 (International Classification of Diseases, Tenth Revision) mental health diagnoses. Sentiment was labeled from 3 perspectives (patient, physician, and general) using 2 LLMs (DeepSeek-7B and Mistral-7B) and compared to lexicon-based tools (ClinSent-lexicon, TextBlob, and VADER [Valence Aware Dictionary and sEntiment Reasoner]). Temporal trends were quantified at the patient level using Kendall τ. Associations between sentiment patterns and mortality were assessed using independent-samples (Welch) t tests. Model performance was evaluated on a manually annotated subset (n=165) using precision, recall, and F₁-score. Results: Temporal sentiment trajectories showed substantial directional change among patients with multiple hospital admissions, with greater fluctuations in Discharge Instructions compared to Brief Hospital Course notes. Patient-perspective sentiment was more balanced, while physician and general perspectives were predominantly neutral. LLMs showed better alignment with patient-centered annotations than lexicon-based methods. Discharge-note sentiment trajectories were significantly more negative among patients who died within 30-90 days of discharge than among survivors across both LLMs. Conclusions: Temporal sentiment analysis revealed section-dependent patterns in clinical narratives that partially reflected patient-perceived experience. LLM-based approaches improved alignment with patient-centered sentiment, although overall performance remained limited. These findings underscore the need for larger, more robust datasets and modeling strategies for clinical sentiment analysis.

What Single-Topic Summaries Miss in Hospital Reviews—Aspect-Level Evaluative Structure Using Generative Pretrained Transformer–Based Sentiment Analysis: Content Analysis

2026-09-14T16:30:18-04:00

Background: Health care service quality is inherently multidimensional; yet, the dominant practice in applied text analysis assigns each patient review to a single topic via Latent Dirichlet Allocation (LDA). This simplification may systematically compress evaluative information when patients discuss multiple service dimensions with varying sentiments within the same review. Objective: This study compared the dominant-topic operationalization commonly used in applied LDA research with Generative Pretrained Transformer (GPT)–based aspect-based sentiment analysis (ABSA) to examine (1) the extent to which patient reviews contain multiple service-quality aspects and how single-topic summaries represent or obscure this structure, (2) the prevalence and patterning of mixed-sentiment reviews, and (3) whether positive and negative reviews differ in aspect comention profiles, before and after adjusting for marginal aspect prevalence. Methods: We analyzed 5467 Google Reviews posted in 2024 from all 24 medical centers in Taiwan. LDA (K=7 topics) and GPT-based ABSA with structured prompts were applied to the same corpus, with the 7 ABSA categories aligned to the LDA topic labels for a controlled but information-asymmetric comparison. Two independent annotators achieved interrater reliability of Cohen κ=0.82; against the consensus gold standard, GPT-4o achieved an accuracy of 0.89, a weighted F₁ of 0.89, and a Cohen κ of 0.78. Mixed-sentiment reviews were identified as those containing both positive and negative aspect evaluations. Rating-stratified network analysis compared aspect comention patterns between positive and negative reviews using Jaccard similarity, with pointwise mutual information as a prevalence-adjusted sensitivity analysis. Results: Aspect-bearing reviews discussed an average of 2.05 distinct aspects (95% bootstrap CI 2.02-2.08), yielding an illustrative 51.2% representational compression estimate under dominant-topic assignment (95% bootstrap CI 50.6%-51.9%). A soft-assignment LDA baseline reduced count-level compression to 1.7%, but semantic alignment with ABSA aspects remained limited (mean set Jaccard=0.33), and topic assignments carry no aspect-level sentiment polarity. Among multiaspect reviews, 11.0% exhibited cross-aspect mixed sentiment, with Technical-Functional Divergence—praising technical quality while criticizing functional quality—appearing in 61.6% of these cases. Clinical dimensions were more frequently comentioned in positive reviews and operational dimensions in negative reviews; however, pointwise mutual information analysis indicated that these differences were substantially confounded with marginal aspect prevalence rather than reflecting differential co-occurrence tendencies. Conclusions: In this corpus, dominant-topic assignment compressed multiaspect patient feedback; soft-assignment LDA recovered topic counts but did not restore semantic alignment or aspect-level sentiment polarity. A nontrivial subset of reviews exhibited cross-aspect mixed sentiment, most commonly praising clinical competence while criticizing functional service dimensions, and positive and negative reviews discussed different constellations of quality dimensions—differences that primarily reflect which aspects patients discuss rather than prevalence-independent associations. Aspect-level analysis that preserves both multidimensional structure and sentiment polarity may help organize patient feedback at a more diagnostically specific level than single-topic summaries.

Depictions of Depression in Generative AI Video Models: Mixed Methods Study of OpenAI’s Sora 2

2026-09-14T16:15:11-04:00

Background: Generative AI video models are increasingly capable of producing complex depictions of mental health experiences, yet little is known about how these systems represent conditions such as depression. Because AI-generated content may reach people during vulnerable periods, understanding what visual narratives these models produce for sensitive concepts carries clinical relevance. Objective: This study aimed to characterize how OpenAI’s Sora 2 generative AI video model depicts depression and examine whether depictions differ between the consumer app and developer API access points, which differ in their product layer mediation. Methods: We generated 100 videos using the single-word prompt “Depression” across 2 access points: the consumer app (n=50, 50%) and developer API (n=50, 50%). Two trained coders independently coded narrative structure, visual environments, objects, figure demographics, and figure states. Interrater reliability was assessed using the Cohen κ, with dimensions showing insufficient agreement excluded from analysis. Computational features (visual aesthetics, audio, semantic content, and temporal dynamics) were extracted and compared between modalities using 2-tailed Welch tests with Benjamini-Hochberg false discovery rate correction. Results: App-generated videos exhibited a pronounced recovery bias: 78% (39/50) featured narrative arcs progressing from depressive states toward resolution compared with 14% (7/50) of API outputs. This divergence was reinforced across channels. App videos brightened over time (mean slope 2.90, SD 2.43 per second vs −0.18, SD 1.24 per second for the API; Cohen =1.59; <.001) and contained 3 times more motion (Cohen =2.07; <.001). Across both modalities, videos converged on a narrow visual vocabulary: predominantly seated figures (94/100, 94% of the videos); downward gaze (93/100, 93% of the videos); and recurring objects including hoodies (n=194), windows (n=148), and rain (n=83). Transcript language in depressive phases emphasized weight and containment (“heavy,” “drowning,” and “room”), whereas recovery phases reversed these patterns: brightness increased by 26.6% (Cohen =0.68; <.001), gaze shifted upward in 67% (30/45) of recovery videos, and terms such as “light” and “breath” emerged. Figures were predominantly young adults (323/367, 88% aged 20-30 years) and nearly always alone (360/367, 98%). Gender varied by access point: app outputs skewed male (34/50, 68%), and API outputs skewed female (33/56, 59%). Conclusions: Sora 2 does not invent new visual grammars for depression but compresses and recombines cultural iconographies, whereas platform-level constraints substantially shape which narratives reach users. Clinicians should be aware that AI-generated mental health video content reflects training data and platform design rather than clinical knowledge and that patients may encounter such content during vulnerable periods.