Journal of Medical Internet Research

Comparative Effectiveness of AI-Assisted Telerehabilitation, Telerehabilitation, In-Person Care, and Usual Care for Chronic Nonspecific Low Back Pain: Bayesian Network Meta-Analysis

2026-07-03T17:30:31-04:00

Background: Guided exercise is central to rehabilitation for chronic nonspecific low back pain. Telerehabilitation enables remote delivery of guided exercise, but its effectiveness vs other rehabilitation modalities remains uncertain. Objective: This review systematically assessed the comparative efficacy of telerehabilitation, in-person rehabilitation (IPR), and usual care (UC) for improving pain, disability, kinesiophobia, and health-related quality of life in patients with chronic nonspecific low back pain. Telerehabilitation combined with artificial intelligence (TLRH-AI) was evaluated as an exploratory intervention because available evidence was limited. Methods: Following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines, we searched randomized controlled trials in PubMed, Cochrane Library, Web of Science, and Embase from inception to April 30, 2026. A Bayesian network meta-analysis was conducted using R (version 4.4.1). Interventions were ranked using surface under the cumulative ranking curve (SUCRA) values. Evidence certainty was assessed using the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) framework. Findings were interpreted considering heterogeneity, risk of bias, inconsistency, and estimated prediction intervals. Results: Among 2491 records, 20 randomized controlled trials involving 1854 participants were included. For pain intensity, IPR showed the greatest benefit at 4 weeks (low-certainty evidence), telerehabilitation at 8 weeks (moderate-certainty evidence), and telerehabilitation ranked highest at 12 weeks (SUCRA 87.2%; moderate-certainty evidence). For the Oswestry Disability Index–based disability, IPR ranked highest at 4 weeks (SUCRA 98.2%; low-certainty evidence) and 12 weeks (SUCRA 86.7%; low-certainty evidence), whereas telerehabilitation ranked highest at 8 weeks (SUCRA 90.4%; high-certainty evidence). For the Roland-Morris Disability Questionnaire–based disability, IPR was among the more effective interventions (SUCRA 67.3%; low-certainty evidence). For kinesiophobia, IPR ranked highest (SUCRA 99%; low-certainty evidence). For health-related quality of life, telerehabilitation significantly improved the physical component summary score (mean difference 6.05, 95% credible interval [CrI] 2.89-9.22; moderate-certainty evidence), whereas IPR showed a nonsignificant trend toward an improved mental component summary score (mean difference 2.79, 95% CrI −1.61 to 7.17; low-certainty evidence). Evidence for TLRH-AI remained limited and descriptive, suggesting possible short-term benefits with low to very low certainty. No significant small-study effects or global inconsistency were detected, although potentially important local inconsistency was observed in the 4-week Oswestry Disability Index comparison between UC and IPR. Conclusions: This review uniquely compared telerehabilitation, IPR, UC, and exploratory TLRH-AI within a Bayesian network meta-analysis. Unlike previous reviews focused mainly on telerehabilitation vs conventional care, it provides a comparative hierarchy across delivery models, follow-up windows, and outcomes while incorporating evidence certainty and heterogeneity. The findings support individualized rehabilitation selection. In practice, telerehabilitation may offer a scalable option for longer-term pain relief and physical function improvement, whereas IPR may remain important for supervised functional recovery and psychological support. TLRH-AI remains exploratory and should not guide clinical decision-making until adequately powered trials are available. Trial Registration: PROSPERO CRD420251146712; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251146712

Effectiveness of WeChat Public Account Intervention Based on the Information-Motivation-Behavioral Skills Model Among College Students With Internet Addiction: Randomized Controlled Trial

2026-07-03T17:30:03-04:00

Background: Internet addiction (IA) among college students causes multiple harms and increases suicide risk. Effective interventions can reduce these and have public health value. Objective: This study evaluates a WeChat public-account intervention based on the information, motivation, behavioral skills (IMB) model to reduce IA in college students and to examine mechanisms driving behavior change. Methods: A total of 226 college students aged 18 to 24 years with IA were recruited from a university in Anhui Province, China, and randomized (stratified by gender) to intervention (n=113) or control (n=113) groups. The intervention comprised a 6-week structured online program delivered via a WeChat public account, organized into 3 modules: information, motivation, and behavioral skills. Controls received no intervention. Preintervention data were subjected to statistical comparison using an independent sample t test, Mann-Whitney U test, and chi-square test. Primary and secondary indicators were measured using the Internet Addiction Test (IAT), Chinese Treatment Self-Regulation Questionnaire, and Internet Control Self-Efficacy Questionnaire. To evaluate the intervention’s effects and identify pre- to postintervention changes, we used generalized linear mixed models and generalized estimating equations for analysis with SPSS software (version 23.0), considering statistical significance at α=.05. Participants were further stratified into mild and moderate to severe IA subgroups based on their baseline scores, and subgroup analyses were conducted to examine both the intervention effects within each stratum and the between-group differences across different addiction severity levels. Results: The intervention group showed a significant reduction in IA (χ²₁=14.154; P<.001). This paper confirmed a substantial reduction in addiction levels after the intervention compared to the control group (adjusted mean change –4.531, 95% CI –7.281 to –1.781; P=.001), with no significant differences in internet use duration. Consistent findings were obtained in the subgroup analyses. A time-group interaction was identified for sleep duration (P<.001), with a notable increase observed in the intervention group (odds ratio 2.186, 95% CI 1.142-4.183; P=.02). No significant differences were observed between the groups in sleep quality and somatic-psychological symptoms. Consistent findings were observed in the subgroup analyses of the above studies. The intervention group showed increased motivation (adjusted mean change 4.283, 95% CI 0.804-7.763; P=.02) and behavioral skills (adjusted mean change 3.407, 95% CI 1.771-5.043; P<.001) after the intervention. Only behavioral skills were improved in the mild IA subgroup, whereas only motivation showed improvement in the moderate to severe subgroup. Conclusions: The WeChat public-account intervention based on the IMB model significantly reduced the level of IA in college students and simultaneously had a positive effect on increasing sleep duration. The decline in IAT scores observed in the moderate to severe subgroup was twice as large as that in the mild subgroup. Trial Registration: ClinicalTrials.gov NCT06704984; https://clinicaltrials.gov/study/NCT06704984

Centering Equity During Health Technology Innovation: Scoping Review of Methods and Research Adjustments to Promote Inclusive Coproduction

2026-07-03T16:45:05-04:00

Background: Digital health has the potential to mitigate health inequity for priority populations who are underserved or marginalized by the health system. However, there is a lack of practical guidance on how to include priority communities in the coproduction of digital health technologies, particularly across the entire lifecycle of innovation, including research, development, and evaluation. Objective: The aim of this scoping review was to systematically identify and assess published methods used during digital health innovation to promote equitable inclusion of priority communities at every stage of the Centre for eHealth Research roadmap for digital health technologies. Methods: This review was based on the Arksey and O’Malley framework for scoping reviews. A 6-stage framework was used to execute the review. To increase the trustworthiness of the findings, an expert advisory group was consulted, and their feedback incorporated into the final manuscript. The Participant, Concept, and Context framework was used to structure the inclusion criteria. Results: The review identified a total of 106 articles, 58 methods, 4 approaches, and 17 research adjustments used to coproduce digital health technologies with priority communities. Common methods across multiple stages included interviews, focus groups, surveys, and workshops; however, the most accessible way to make equity a practical reality during health technology innovation is to appoint a priority population community advisor, or advisory group, from project inception to project closure. Visual and creative methods like photovoice, home tours, and body-mapping were also used, often by priority population researchers themselves. Research adjustments that promote patient safety and comfort, enhanced literacy, peer-support, and recognize sociocultural and demographic considerations have been used to increase the inclusion of priority populations during digital health innovation. Conclusions: Embedding equity is possible using the practical methods and research adjustments identified to promote inclusive coproduction. Professionals working across health care, health informatics, research, digital health, and technology development can use these findings to center digital health equity during technology innovation. This research also recognizes that coproduction must draw on epistemological frameworks, or ways of thinking, which support Indigenous and other priority population knowledge systems. A solely Western lens risks reinforcing structural barriers and overlooking essential knowledge, as demonstrated by this review when the search strategy missed key scholarly works by priority population authors themselves. International Registered Report Identifier (IRRID): RR2-10.2196/53855

Patient Perceptions of Artificial Intelligence–Supported Shared Decision-Making in UK Primary Care for Multiple Long-Term Conditions: Qualitative Study

2026-07-03T16:45:05-04:00

Background: The prevalence of multiple long-term conditions (MLTCs) is increasing globally, leading to complex health care needs and polypharmacy. Shared decision-making (SDM) is important for supporting patient-centered care, yet barriers such as limited consultation time, discontinuity of care, and communication challenges hinder implementation. Artificial intelligence (AI) has the potential to support SDM by providing personalized, data-driven recommendations, particularly for medication management in patients with MLTCs. Objective: This study aimed to explore the perspectives of patients with MLTCs regarding SDM with their general practitioners (GPs) and to explore patients’ views about the use of an AI tool to support SDM, particularly in relation to prescribing decisions. Methods: This qualitative study explored the perspectives of 18 patients with MLTCs on SDM and the use of an AI tool prototype during GP consultations. Semistructured interviews used a simulated patient vignette and a visual AI tool dashboard to facilitate discussion. Participants were recruited through GP practices via the Clinical Practice Research Datalink and community-based organizations across the West Midlands. The data were then analyzed using thematic analysis. Results: Two overarching categories were identified: SDM in GP consultations and the AI tool for SDM. Within SDM, themes included communication and collaboration and system-level barriers, such as limited consultation time, lack of continuity, and fragmented records. Within the AI tool category, themes were related to practical design and implementation, implications for clinical practice and decision-making, and perceived risks and limitations. Participants valued the tool’s potential to summarize health information and support discussions but highlighted the need for clear explanations, accessible design, and clinician guidance. Concerns included time pressures, depersonalization, trust, and transparency, with participants emphasizing that AI should support rather than replace clinical judgment. Conclusions: Overall, patients perceived AI as a promising way to enhance SDM by improving communication and collaboration between patient and clinician. However, patients also had concerns about the accuracy and veracity of AI. The study provides recommendations for AI tools in GP consultations, emphasizing clear, accessible outputs and the use of lay language. AI tools should enhance rather than replace clinical judgment, be transparent about data sources, and be developed with diverse patient input to ensure inclusivity and usability, particularly for those with MLTCs.

Detecting and Preventing Fraudulent Participation in Qualitative Research: Content Analysis of Two Multisite Studies

2026-07-03T16:30:20-04:00

Background: The use of web-based approaches to identify, recruit, enroll, survey, and interview health-related research participants has increased over time, with rapid acceleration since the COVID-19 pandemic. These approaches can make research more accessible to a broader population, but also increase the risk of fraudulent or imposter participants infiltrating research studies. While this threat has been discussed extensively in quantitative survey research, less has been reported in qualitative and mixed methods studies. Objective: This study aims to identify recurring patterns of fraudulent study participation and to offer strategies for identification, remediation, and reporting. Methods: Encounters with fraudulent or imposter individuals during recruitment, enrollment, survey distribution, data collection, and focus group sessions in 2 multisite qualitative and mixed methods research studies are presented. Content from both studies was analyzed to identify common themes and develop strategies for prevention and remediation. Results: Investigators across 2 multisite studies observed several indicators of suspected fraudulent activity, including large response volumes over a short period, highly repetitive email addresses, higher-than-expected proportions of phone numbers with area codes outside the study area, and unusual email/phone responses using atypical language and phrasing. Several imposter or fraudulent individuals disrupted online focus group sessions. To mitigate these issues, both studies implemented remediation strategies, including enhanced screening procedures at baseline, cross-checking of survey responses, and additional identity verification methods prior to participation. Studies took various actions to address these experiences, including notifying the institutional review board, recruitment platforms, and funders. Conclusions: This multisite study identified multiple ways that imposter or fraudulent participants can pose a significant and evolving threat to the integrity of qualitative and mixed methods. These types of fraudulent actors can distort data and undermine research credibility. Lessons learned highlight the importance of real-time recruitment and enrollment analysis and the need for transparent reporting. Addressing this issue will require a comprehensive approach to prevent and address fraudulent study participation that includes collaboration with multiple stakeholders and the broader research community to effectively address this issue.

Impact of Telemedicine-Enhanced Integrated Management of Gestational Diabetes on Pregnancy Outcomes and Glycemic Control: Real-World Study Using TangMama App

2026-07-03T15:30:05-04:00

Background: Gestational diabetes mellitus (GDM) is associated with substantial risks of adverse maternal and neonatal outcomes. Contemporary management approaches for GDM exhibit insufficient implementation, resulting in suboptimal glycemic control and preventable perinatal complications. The rapid evolution of mobile health technologies offers potential to enhance GDM care, yet evidence from large real-world studies remains limited. Objective: This study aimed to evaluate the impact of a telemedicine-enhanced integrated management system on pregnancy outcomes and glycemic control in women with GDM and to explore the dose-response relationship between telemedicine engagement intensity and clinical outcomes. Methods: In this real-world, prospective cohort study conducted at a provincial-level medical center in China, women with GDM were categorized into a standard care group and a telemedicine-enhanced group receiving the TangMama smartphone app in addition to standard care. We compared pregnancy outcomes and glycemic parameters between the 2 groups in an inverse probability of treatment weighting population based on propensity scores. Mediation analyses and dose-response analyses were additionally conducted to explore potential mechanisms and engagement effects. Results: A total of 4621 women with GDM were included, with 1711 in the telemedicine-enhanced group and 2910 in the standard care group. Upon inverse probability of treatment weighting analysis, the telemedicine-enhanced group demonstrated significantly lower gestational weight gain (adjusted mean difference −1.49 kg, 95% CI −1.81 to −1.17), reduced rates of excessive gestational weight gain (adjusted odds ratio [aOR] 0.61, 95% CI 0.54-0.69), cesarean section (aOR 0.80, 95% CI 0.71-0.91), hypertensive disorders in pregnancy (aOR 0.76, 95% CI 0.64-0.90), and pre-eclampsia (aOR 0.64, 95% CI 0.49-0.83). Glycemic control in the third trimester was significantly improved, with lower glycated hemoglobin A (HbA) levels (adjusted mean difference −0.05%, 95% CI −0.08 to −0.03) and higher HbA on-target rates. For neonatal outcomes, telemedicine-enhanced management was associated with lower rates of preterm birth (aOR 0.47, 95% CI 0.38-0.59), large-for-gestational age (aOR 0.81, 95% CI 0.69-0.96), neonatal unit admission (aOR 0.80, 95% CI 0.71-0.91), neonatal hypoglycemia (aOR 0.64, 95% CI 0.45-0.93), and neonatal hyperbilirubinemia (aOR 0.69, 95% CI 0.58-0.82). Mediation analyses identified gestational weight gain and third-trimester fasting plasma glucose as significant mediators. Higher telemedicine engagement was associated with improved glycemic control and reduced adverse outcomes in a dose-response manner. Conclusions: Telemedicine-enhanced integrated management is associated with improved maternal glycemic control and substantial reductions of adverse pregnancy outcomes among women with GDM. The observed dose-response relationship between engagement intensity and outcomes underscores the importance of promoting active patient participation. These findings support the broader integration of telemedicine into routine GDM care pathways to optimize maternal and neonatal health.

Integrating Participatory Social Innovation Into Requirements Engineering for AI Health Care Solutions: Case Study

2026-07-03T14:15:16-04:00

Background: The successful design and implementation of artificial intelligence (AI)–driven solutions in health care requires early and continuous multidisciplinary and multiprofessional collaboration. However, diverse disciplinary educational backgrounds, varying languages, and cultural or geographic differences can lead to misunderstandings. To bridge this gap, a structured approach to AI requirements specification can facilitate a shared terminology and a deep mutual understanding among stakeholders, serving both as a guide for technological development and as a means of defining clear pathways for clinical implementation. While technical requirements are well-established in traditional technology development domains, this structured approach remains relatively underused within clinical and social science contexts. Consequently, valuable insights derived from participatory and stakeholder-driven approaches are often overlooked, limiting the relevance and trustworthiness of AI systems in health care settings. Objective: This study presents a methodology for requirements gathering, specification, mapping, and verification, specifically engineered for the complex, multistakeholder environment of clinically applied AI. The methodology was implemented within the specific case of an international multidisciplinary project evaluating an AI-based prediction tool for neoadjuvant chemotherapy treatment response for breast cancer and forming a part of the developed AI validation framework. Methods: The process for AI requirements gathering, specification, and monitoring included 3 iterative rounds of discussion, engaging nearly 150 social, clinical, technical, ethical, and regulatory experts and patients across Europe, South America, North Africa, and Eurasia. It combines established requirements engineering methods (including the MoSCoW [Must have, Should have, Could have, Will not have] framework) with social innovation techniques to ensure inclusivity and contextual relevance. Results: A key finding is the successful development of a structured framework integrating participatory social innovation with formal requirements engineering in an international AI health care setting, through a traceable multisource workflow including clinical, ethical, and regulatory aspects. It is supplemented with an extensive list of 184 actionable consensus-based requirements, categorized by stakeholder group, providing valuable insights for AI researchers in the oncology field with the potential to be transferable to other digital health domains. The requirements align with the fairness, universality, traceability, usability, robustness, and explainability in the AI (FUTURE-AI) framework, ensuring the tool is trustworthy and comprehensive from a multistakeholder perspective, and ensuring comprehensive consideration of all elements of FUTURE-AI. Conclusions: The proposed methodology represents a significant advancement for requirements engineering in digital health by extending traditional technical processes to systematically incorporate nontechnical requirements from diverse global stakeholders. This unified approach is essential for ensuring AI solutions are not only technically robust but also clinically relevant, legally compliant, and socially acceptable.

Are Traditional Registries Becoming Obsolete in the Modern Digital Health Ecosystem?

2026-07-03T13:45:21-04:00

Registries have long been a cornerstone of medical research and public health, providing systematically collected data on diseases, treatments, and health outcomes. However, in the era of digital health, we argue that the traditional model of stand-alone registries needs reconsideration, given the context of increasingly digitized and interoperable health data ecosystems. Unless registries evolve to embrace embedded, standards-based data services, operating across interoperable infrastructure, they will become obsolete while digitalization is reshaping how data can be collected, shared, and used. In this viewpoint, we recount how the present health data ecosystem came to be and what role registries have come to play therein. Following that, we show how recent regulatory initiatives such as the Trusted Exchange Framework and Common Agreement in the United States or the European Health Data Space Regulation signal a shift toward cross-network health information exchange, promoting patient-centric data integration within electronic health record systems. We further illustrate how electronic health records are consequently set to evolve into information hubs, acting as the primary gateway for individuals through which they may access and control their personal health data spread throughout increasingly connected health data ecosystems. This, in turn, might stimulate the creation of digital twins and continuous learning health systems in practice. Following this line of thought, we discuss the opportunities and challenges of interconnected health data ecosystems. Ultimately, we propose that next-generation registries need to be designed as dynamic, service-oriented software stacks for research, leveraging the common data infrastructures that are currently being established around the world. Given the points raised in this viewpoint, we invite health care professionals and researchers alike to equally rethink the role that registries should play within the globally emerging interconnected health data ecosystems and contribute their findings. References included in this viewpoint were identified through searches of PubMed and Google Scholar with various search terms and combinations thereof pertinent to the topics touched on, for example, “patient registry,” “clinical registry,” “digital twin,” “healthcare,” “clinical research,” “virtual twin,” “TEFCA,” or “EHDS.” Only papers in English were reviewed. The final reference list was generated on the basis of originality and relevance to the broad scope of topics covered in this viewpoint, aiming to present a balanced overview of topic-related findings and arguments.

Adoption of Artificial Intelligence–Based Precision Mental Health Technologies Among Psychology Trainees: Mixed Methods Cross-Sectional Survey Study

2026-07-03T13:45:16-04:00

Background: Despite the significant benefits of artificial intelligence (AI) in mental health, real-world implementation remains limited, making it essential to understand the factors that influence adoption. Objective: This study examined the acceptability and intention to use artificial intelligence–based precision mental health technologies (AI-PMHTs) and proposed an empirical, theory-guided model that integrates traditional technology acceptance predictors (eg, perceived usefulness, risk, and ease of use) with emerging psychological factors (eg, AI anxiety, personality, and conspiratorial thinking) that may inform future implementation research, strategic planning, and training program design. Methods: An online survey was distributed to a sample of 357 psychologists in training, including both undergraduate and master’s students. A mixed methods approach was used, combining quantitative measures (via psychometrically validated questionnaires) and qualitative data (through open-ended questions). Descriptive statistics and tests were conducted to characterize the sample, and responses to the open-ended questions on facilitators and barriers were thematically analyzed. Partial least squares structural equation modeling was used to build the empirical model. Results: Participants showed moderate-to-high acceptance and intention to use AI-PMHTs, yet anxiety and perceived risk varied (with higher levels among women), and more frequent use was linked to more favorable acceptance profiles without reducing fear. Thematic analysis revealed that participants viewed AI tools as efficiency-enhancing but raised concerns about reliability, usability, overdependence, and access constraints. Partial least squares structural equation modeling supported a hierarchical adoption pathway in which dispositional and demographic factors shape AI-related fear and perceived risk, which then influence cognitive evaluations and attitudes, ultimately being associated with acceptance and intention to use AI-PMHTs. Predisposing variables (particularly resistance to change and conspiratorial thinking) were the strongest predictors of AI-related anxiety, with gender and extraversion showing smaller but meaningful effects. Fear acted as a key affective mediator, increasing perceived risk and indirectly weakening positive attitudes and perceived usefulness. Acceptance was the most influential downstream construct, directly predicting satisfaction, perceived usefulness, prior experience, and future intention to use, consistent with a reinforcing feedback loop in which early acceptance supports sustained engagement. Conclusions: Findings suggest a layered framework that may inform future implementation research and training program design, addressing (1) predisposing dispositional and emotional profiles; (2) precipitating fear and perceived risk via transparent regulation, explainable design, and policies that strengthen professional agency; and (3) maintenance through high-quality early experiences, usability, and sustained institutional support. This theory-guided model clarifies how psychological, contextual, and experiential factors jointly shape adoption and sustained use of AI-PMHTs among psychologists in training, informing targeted educational and implementation strategies for this population.

Performance of DeepSeek V3.2 and ChatGPT 5.1 in Musculoskeletal Triage and Differential Diagnosis of Outpatients With Low Back Pain: Multidimensional Comparative Study

2026-07-03T13:45:16-04:00

Background: Outpatients presenting with low back pain (LBP) often require efficient preconsultation triage and early differential diagnostic support. Large language models may assist these text-based tasks, but their performance under different clinical information conditions remains unclear. Objective: This study aimed to compare the performance of ChatGPT (5.1; OpenAI) and DeepSeek (V3.2; DeepSeek AI) in musculoskeletal disorders (MSDs) triage and the differential diagnosis of outpatients with LBP using real-world outpatient records under 2 simulated information conditions. Methods: This retrospective comparative study was conducted at a tertiary academic teaching hospital in Beijing. A total of 160 cases were included using a balanced design across 8 diagnostic categories (20 per category); 6 MSDs and 2 non-MSDs. Evaluation was performed in 2 phases: Phase 1 (chief complaint) and Phase 2 (structured questionnaire with 7 domains or 33 items), both executed in a zero-shot setting using standardized prompts. Outcomes included (1) triage accuracy, (2) preliminary diagnosis accuracy, and (3) differential diagnosis agreement. In Phase 2, 3 senior orthopedic evaluators additionally rated model rationales across 5 domains using a 5-point Likert scale. Results: For triage accuracy across all 160 cases, DeepSeek V3.2 improved from 84.4% to 90.6% (risk difference [RD] 6.2%, 95% CI –0.7% to 13.3%), and ChatGPT 5.1 improved from 75.6% to 93.1% (RD 17.5%, 95% CI 10.2%-24.9%). For preliminary diagnosis accuracy across the 120 musculoskeletal cases, DeepSeek V3.2 improved from 48.3% to 76.7% (RD 28.3%, 95% CI 16.8%-38.8%), whereas ChatGPT 5.1 improved from 35.0% to 87.5% (RD 52.5%, 95% CI 42.8%-60.6%). The mean number of correct differential diagnoses increased from 1.27 (SD 0.71) to 2.02 (SD 0.74) for DeepSeek V3.2 and from 1.34 (SD 0.70) to 2.03 (SD 0.77) for ChatGPT 5.1. In Phase 2, rationale ratings were generally good for both models, with ChatGPT 5.1 scoring higher in understanding and reasoning. Recognition of multiple myeloma (MM) remained limited, improving only from 45% to 55% (DeepSeek V3.2) and 55% to 60% (ChatGPT 5.1). Structured input reduced safety-risk errors in both models, but residual errors remained, especially for MM and metastatic spinal tumor. Conclusions: Both ChatGPT 5.1 and DeepSeek V3.2 demonstrated potential in text-based triage and differential diagnosis of MSDs for LBP, with structured clinical information generally improving performance, particularly for preliminary diagnosis accuracy and differential diagnosis agreement. However, their suboptimal sensitivity for red-flag conditions such as MM highlights significant safety concerns, indicating that they should not be used as stand-alone triage tools without clinician oversight. ChatGPT 5.1 showed stronger reasoning with structured inputs based on rationale ratings, whereas DeepSeek V3.2 showed better performance under chief-complaint-only input, with significantly higher Phase 1 preliminary diagnostic accuracy and numerically higher Phase 1 triage accuracy. These findings underscore the need for further model refinement, rigorous prospective validation, and integration with clinician oversight before clinical implementation.