In response to the dual demands of enhancing safety and greening transportation infrastructure under China's "Dual Carbon" goals, this study overcomes the limitations of traditional guardrail materials and processes by proposing and validating a synergistic pathway for optimizing both the safety and carbon emission reduction of alloy steel corrugated beam guardrails. A systematic "material design - process optimization - performance verification - environmental assessment" framework was established. High-performance 700L-grade alloy steel was produced through low-carbon alloy design combined with ESP short-process rolling technology. Its safety performance was quantitatively evaluated via SB-level full-scale vehicle crash tests, and its life-cycle carbon footprint was quantified using an ISO 14067-compliant model implemented in eFootprint software with the CLCD database. The results demonstrate that the alloy steel achieves a synergistic optimization of strength and plasticity, with a tensile strength of 766-781 MPa and a product of strength and elongation (PSE) exceeding 0.175 GPa·%. In the full-scale vehicle crash tests, all occupant risk indicators were superior to the safety limits. For instance, the key risk parameters for the small passenger car, such as the longitudinal velocity (Vx = 4.3 m/s) and the lateral acceleration (ay = 126.0 m/s²), demonstrated excellent performance. The maximum dynamic outward inclination equivalent values for the medium and large trucks were 1.75 m and 2.45 m, respectively. These results confirm that the safety performance of the guardrail fully meets the SB-level standard, even with a lightweight design featuring a 25% reduction in beam thickness and a 50% reduction in post thickness. Life-cycle analysis revealed that the carbon footprint per kilometer of guardrail was reduced to 80.86 t CO2e, representing a 73.8% reduction compared to the conventional solution. Sensitivity analysis identified iron input and electricity consumption as the core influencing parameters. Furthermore, a cost-benefit analysis indicated superior life-cycle cost advantages. This study elucidates the mechanism for achieving synergistic gains in safety and emission reduction through material and process innovation, providing a systematic solution and data support for the green, low-carbon, and safe transformation of highway infrastructure.
Patient-facing large language model (LLM) systems are increasingly proposed as scalable tools for chronic disease self-management support. In this setting, safety depends not only on factual accuracy but also on whether outputs are interpretable, trusted, and used safely over time. To explain how safety guardrails shape outcomes in patient-facing LLM-supported self-management across different task-risk, user, and interaction contexts. We conducted a realist review of LLM-based or LLM-enabled generative conversational systems used for self-management of long-term physical health conditions. Searches of PubMed, Web of Science Core Collection, IEEE Xplore, ACM Digital Library, and arXiv covered all indexed years to 1 April 2026 and used terms for chronic disease or self-management, patient-facing conversational systems, and LLMs or generative AI, identifying 1,154 records before deduplication. The core evidence base comprised 21 studies and 38 context-mechanism-outcome configurations (CMOCs). The patient-facing self-management task, rather than disease label, was the unit of synthesis. At the study level, the dominant patient-facing task was coded as low risk in 6 studies, moderate risk in 11, and high risk in 4; 17 studies evaluated mainly single-turn interactions and 4 included multi-turn, sequential, or simulated-consultation elements. Most evidence concerned simulated or expert-judged patient-facing tasks rather than sustained real-world deployment. Three patterns recurred. Provenance-related safeguards improved transparency and checkability more consistently than they ensured safe downstream action. Communication-oriented safeguards improved readability or perceived comprehensibility while leaving recurring gaps in completeness or actionability. Boundary-control strategies, including source-bounded retrieval, clinician deferral, and escalation support, became more important as task actionability and interaction complexity increased. In patient-facing chronic disease self-management, safety cannot be judged adequately by answer plausibility alone. This review develops a refined programme theory and a risk-linked, theory-generating heuristic framework, but many proposed mechanisms remain indirect and require real-world, longitudinal, multi-turn testing before deployment.
Early childhood caries (ECC) is a complex, multifactorial disease shaped by biofilm ecology, host susceptibility, diet and behaviors, and structural determinants of health. Silver diamine fluoride (SDF) is an effective non-restorative option for arresting cavitated lesions in many settings and can support access when definitive care is delayed. However, translating short-horizon "arrest" outcomes into broad policy claims-that SDF-first, delegated pathways can substitute for dentist-led diagnosis and comprehensive rehabilitation-risks institutionalizing a two-tier standard of care for children facing the greatest access barriers. This perspective critically appraises evidence-to-implementation pathways for SDF and delegated ECC management, using risk-of-bias and reporting guidance as interpretive tools and drawing on pragmatic regimen trials, microbiome substudies, oral health-related quality of life (OHRQoL) analyses, and implementation work including the Canadian Caries Risk Assessment Tool (CCRAT) in primary care. We explicitly distinguish what studies demonstrate (e.g., feasibility and short-term arrest differences by reapplication interval) from what they do not establish (e.g., long-term tooth survival, pulpal outcomes, definitive treatment completion, and equity impacts). We propose practical guardrails that position SDF as interim management within a continuum of care: dentist-led diagnosis and escalation when pulpal risk is suspected; time-bound referral pathways with completion tracking; protocolized follow-up aligned with lesion/risk status; outcome sets that extend beyond "arrest" to include pain, function, OHRQoL, tooth survival, and equity stratification; and lesion-site sampling plus preregistered analyses when mechanistic claims are advanced.
Artificial Intelligence (AI) documentation systems are being rapidly adopted throughout the healthcare industry as solutions to help alleviate the clinician documentation burden. The use of Ambient AI documentation systems enables passive capture of clinical conversations and the creation of medical documentation. Unlike routine clinical encounters, trauma resuscitation is characterized by fragmented speech, overlapping conversations, temporal complexities, and immediacy of decision-making. All of these aspects could present challenges to the many AI models based on commonly held assumptions about communication. Without trauma-specific safeguards, AI documentation systems may introduce automation bias, speaker identity errors, time relationship errors, and convert uncertainty into definitive statements. This article provides an examination of the risks associated with the use of AI documentation systems in trauma care and proposes several safety guardrails, including mandatory human review of high-risk elements, preserving uncertainty in documentation, structuring documentation of safety-critical data, providing multidisciplinary oversight of implementation, conducting trauma-specific validation testing, and ongoing auditing of performance.
External evidence from prior trials, registries, and fit-for-purpose real-world data can improve drug development efficiency. Hybrid-controlled designs are particularly appealing for reducing concurrent control enrollment while simultaneously providing internal validity with a randomized control arm. Yet regulatory adoption is limited due to major concerns around bias due to possible differences in characteristics and outcomes between the external data and the trial. To realize the benefits of the hybrid approach without compromising credibility, methodological guardrails are crucial for mitigating bias and enabling valid inference. We assessed eight statistical methods which proactively address differences between external data and trial data. We apply these methods to both a large clinical trial as a case study, as well as within a comprehensive simulation study with continuous outcomes that varied the amount of measured versus unmeasured confounding, the severity of the between-data-source heterogeneity, and the number of external data sources. Results show that two-step strategy, propensity score-based balancing followed by Bayesian dynamic borrowing, consistently delivered the most favorable trade-off between precision gain and bias control. This approach when used with fit-for-purpose external data can provide a robust implementation of the hybrid trial design beyond the narrow set of conditions where there is currently precedent.
Currently, artificial intelligence (AI) is clinically relevant to mood and anxiety care, but the evidence base is uneven across use cases. This narrative review synthesizes recent literature most relevant to clinicians and investigators. Five themes dominate the current field: patient-facing adjunctive tools, failure modes and safety risks, clinician-facing decision support, passive sensing and measurement infrastructure, and governance. Recent randomized evidence supports a narrow efficacy claim for structured chatbot interventions, with small improvements in depressive and anxiety symptoms and more consistent effects on engagement than on symptom superiority. These studies do not support autonomous psychotherapy, and they do not establish a therapeutic advantage for open-ended large language model systems over more constrained designs. Safety studies, by contrast, identify active concerns: harmful endorsement, weak youth risk assessment, inconsistent crisis handling, and anxiety/OCD reassurance loops. The strongest current clinical signal lies in supervised clinician-facing decision support, where recent trials of AI-assisted antidepressant selection improved treatment persistence and some downstream symptom outcomes. Passive sensing detects behaviorally meaningful signals, but evidence that alert-driven deployment improves care remains insufficient for routine practice. Across stakeholder and policy sources, the most defensible deployment model is human-in-the-loop, stepped, and bounded by explicit handoff rules.
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
GenAI is increasingly used to draft replies to patient portal messages to reduce clinician workload. Evidence shows modest utilization and measurable workflow effects, alongside safety risks when clinicians miss errors or accept drafts with minimal editing. As deployments scale with deeper EHR integration and expanding transparency mandates, this Comment proposes governance that preserves efficiency without normalizing unsafe delegation: scoped use, risk tiering, accountable human authorship, auditability, and patient-facing transparency.
暂无摘要（点击查看详情）
Large language models often mishandle psychiatric emergencies, offering harmful or inappropriate advice. This study evaluated the Verily Mental Health Guardrail (VMHG) on two clinician-labeled datasets: the Verily Mental Health Crisis Dataset v1.0, containing 1800 simulated messages and the NVIDIA Aegis AI Content Safety Dataset subsetted to 794 mental health-related messages. Performance was benchmarked against OpenAI Omni Moderation Latest and NVIDIA NeMo Guardrails. The VMHG demonstrated high sensitivity (0.990) and specificity (0.992) on the Verily dataset, with an F1-score of 0.939 and high category-level sensitivity (0.917-0.992) and specificity (≥0.978). On the NVIDIA dataset, it maintained strong sensitivity (0.982) and accuracy (0.921) with reduced specificity (0.859). Compared with NVIDIA and OpenAI guardrails, the VMHG achieved significantly higher sensitivity (all p < 0.001) and comparable specificity (NVIDIA p < 0.001, OpenAI p = 0.094). Overall, the VMHG demonstrated robust, generalizable, and clinically oriented safety performance that prioritizes sensitivity to minimize missed mental health crises.
Public discourse on "prompt engineering" has grown rapidly, but large-scale evidence on how people discuss it is limited. The nascent coinage describes a user journey of optimizing text-based parameters within large language models and generative AI platforms in a trial-and-error process, adding modifiers or key phrases to yield a satisfactory output. While several studies have systematically and taxonomically mapped these techniques, no big data studies have specifically delved into the public reception to the concept. The study significantly contributes to the literature by documenting emerging new trends. It is prudent to keep abreast of discursive content as these techniques enter into the public lexicon, by capturing public discourse, sentiments and dominant themes. This paper maps prominent discussion peaks and quarter-by-quarter themes and sentiment in posts mentioning "prompt engine*" on X and Reddit (1 May 2023-30 April 2024). We analyzed a large social media corpus (n = 298,774) of publicly-available English-language posts containing "prompt engine*" (engine/engineer/engineering/engineered). Our first research direction detected discussion spikes using peak-prominence modeling (pre-specified prominence threshold at 95th-percentile significance), then qualitatively coded the top-engagement posts on peak dates. Our second research direction computed quarterly sentiment, frequent bigrams, and emoji usage. Six peaks centered on: (i) skill validity, (ii) future ramifications, (iii) title longevity, (iv) artist-attribution disputes, (v) model bias/guardrails, and (vi) analogies to child-rearing. Negative sentiment increased from 14.1% (Q1) to 37.3% (Q4), while positive sentiment fell from 36.4% to 22.8%. Tutorials dominated early content; later quarters featured humor, satire and legal liability discussions. Public narratives around prompt engineering increasingly question legitimacy, longevity, and ethics. Clearer terminology, transparency about guardrails, and guidance on responsible use may help align expectations and practice.
Generative Artificial Intelligence (GenAI) has rapidly permeated education, with growing implications for disability-inclusive practice. Objective: This review maps GenAI uses for students with disabilities since public LLM adoption, identifies research clusters, and surfaces gaps. Methods: Following SPAR-4-SLR, we searched Scopus and Google Scholar (publications Jan 1, 2022-Feb 6, 2025; English; journals/conferences). After screening, 88 records were retained for the qualitative SLR; a relevance subset (n = 49, score = 3) underwent bibliometric and text-mining analysis using TF-IDF, K-Means (k = 5), and PCA based visualisation; keyword co-occurrence networks were built in VOSviewer. Results: Five clusters emerged: (1) adaptive tools for autism and language learning; (2) inclusive/early-childhood AI integration; (3) game-based/adaptive learning; (4) broad ChatGPT applications across K-12/higher/special education; and (5) frameworks/conceptual models. ASD and dyslexia dominate; visual/hearing/motor impairments are underrepresented. Conclusions: GenAI shows promise for personalisation, AAC, and teacher support, but evidence is early-stage and uneven across disabilities. Recommendations include standardised reporting (datasets, prompts, guardrails), longitudinal evaluation, and policy frameworks aligning Universal Design for Learning (UDL) with AI ethics and privacy. A replication package (data/code) and a taxonomy for GenAI-inclusive learning are proposed. Expand GenAI interventions beyond ASD and learning disabilities to address major gaps in rehabilitation support for visual, hearing, motor, and other complex disabilities as well: The current evidence base is disproportionately narrow, signalling an urgent need for inclusive GenAI design that meets the rehabilitation needs of all disability groups.Position GenAI as a catalyst for personalised, multimodal rehabilitation by integrating text, speech, image, and VR/AR outputs to strengthen Augmentative and Alternative Communication (AAC), social-communication training, motor practice, and real-world simulation: Such multimodal tools can enhance functional engagement and reduce barriers in both clinical and educational rehabilitation settings.Embed accessibility-first and ethics-centred frameworks such as Universal Design Language (UDL),Web Content Accessibility Guidelines (WCAG) including privacy-preserving architectures and bias mitigation into GenAI development to ensure safety, equitable access, and meaningful participation of neurodivergent and disabled learners: These guardrails are essential for responsible deployment in rehabilitation contexts.Prioritise longitudinal, co-designed, and multidisciplinary research to validate real-world rehabilitation outcomes, ensure cultural and linguistic relevance, and support scalable adoption especially in low-resource or underserved areas: Sustained evidence generation will enable GenAI tools to evolve from promising prototypes into reliable components of rehabilitation practice.
Preventive campaigns for older adults must decide how to allocate limited resources across media channels. However, these channel allocation and budget decisions rarely use explicit criteria for distributional equity or structured strategic planning tools. Consequently, health systems may optimize average uptake while leaving large gaps across socioeconomic groups and media use profiles. This study aimed to develop and apply a data-driven agent-based model as a strategic planning tool for preventive campaigns targeting older adults, comparing channel allocation, personalization, and loss framing options under explicit budget and equity guardrails. We built an agent-based model calibrated to national survey data from South Korea on influenza vaccination and routine health screening among older adults (vaccination, N=2405; screening, N=2400). Fifteen prespecified campaign scenarios varied channel allocation across television, digital, and print media; budget intensity; 2 equity-focused personalization strategies; and graded loss framing. Primary outcomes were final adoption and time to adoption. Equity outcomes included the minimum class-level adoption and 90-10 gap across latent classes. Each scenario was simulated over 12 monthly steps with 100 Monte Carlo replications. We conducted sensitivity analyses varying link functions and key social reinforcement parameters. Personalization improved uptake and equity relative to the integrated baseline. In the vaccination model (N=2405), adoption increased from 91.2% (n=2193) to 93.3% (n=2244) and 94.6% (n=2275). Minimum class-level adoption increased from 86.8% to 90.3% and 90.9%. The 90-10 gap narrowed from 5.7 to 4.5 and 4.7 percentage points. In the screening model (N=2400), adoption increased from 83.8% (n=2011) to 88.2% (n=2117) and 89.5% (n=2148). Minimum class-level adoption increased from 77.6% to 83.2% and 85.3%. The 90-10 gap narrowed from 9.2 to 7.4 and 6.2 percentage points. Television-only strategies achieved high adoption but had less favorable equity profiles than personalization. High-budget strategies achieved high adoption but required higher total exposure. Stronger loss framing produced small, monotonic gains in adoption and shortened the time to adoption without worsening equity in the tested range. Scenario rankings were stable in sensitivity analyses. This agent-based modeling study illustrates how ex ante planning can improve preventive campaign design by comparing channel allocation and personalization options under explicit equity and budget criteria. For campaigns targeting older adults, equity-focused reweighting and class-tailored television-digital portfolios improved or preserved mean adoption while strengthening distributional equity under fixed budgets. In contrast, undifferentiated channel diversification without personalization offered a less favorable efficiency-equity trade-off. These findings support integrating explicit equity guardrails into early-stage channel allocation and prioritizing targeted personalization over simple channel diversification. Future work should validate these patterns in other populations and health systems and link simulated diffusion trajectories with observed exposure and engagement in real-world campaigns. It should also extend guardrail-based planning tools to organizational settings and multiyear decision contexts.
Artificial intelligence (AI), including machine learning, natural language processing, and large language models, may support implementation practice and research in tasks such as evidence synthesis, determinant assessment, strategy selection, monitoring, adaptation, and theory development. However, these applications of AI do not form a single, uniform category. They span a continuum from practice-facing applications that support local implementation work to research- and methods-facing applications that support evidence generation and synthesis. The guidance on how to classify, evaluate, and report these uses of AI remains limited. The AI Methods for Implementation Science (AIM-IS) program aims to develop, validate, and maintain a suite of products to guide the responsible use of AI across implementation practice, implementation research, and bridging use cases. AIM-IS is a multi-phase, multi-method methodological development program. The unit of analysis is the AI-for-implementation use case: a specific AI capability supporting a defined implementation practice or research task within a workflow, decision point, and governance context. Phase 1 is a living scoping review mapping published AI use cases in implementation science, including how they are evaluated and what risks they raise. Phase 2 is a qualitative interview study with implementation researchers, practitioners, AI experts, community members, and data infrastructure and governance experts to refine use cases and identify feasibility constraints, outcome priorities, and reporting needs. Phase 3 will integrate findings from Phases 1 and 2 to develop the draft AIM-IS products, including a framework, a taxonomy of use cases, guardrails for responsible use, a practical guide, outcome domains, and reporting items. Phase 4 will use an eDelphi process and consensus meeting to refine and finalize these products. Phase 5 will conduct usability testing to improve clarity and ease of use, resulting in the finalized AIM-IS products. AIM-IS is informed by implementation science, sociotechnical systems, equity, and responsible AI frameworks, and includes a living-update approach to support ongoing refinement. The AIM-IS program will deliver a suite of products, including a framework, toolkit and reporting standard, to support the specification, governance, evaluation, and reporting of AI in implementation science. Together, these products aim to strengthen transparency, comparability, accountability, and attention to equity in how AI is used by implementation practitioners and researchers over time. Open Science Framework, March 15, 2026: https://doi.org/10.17605/OSF.IO/BX35K.
Artificial intelligence (AI) is rapidly entering clinical practice, yet the governance models needed to ensure its safe, ethical, and equitable use have not kept pace-particularly in community and safety-net settings. Existing frameworks, designed for large academic systems, are often impractical for frontline physicians, creating a dangerous gap between AI adoption and oversight. This article reframes AI governance as a clinician-centered enabler rather than a compliance burden. We propose a pragmatic model built on clear accountability, defined use guardrails, basic safety and bias checks, transparency, and lightweight workflows-supported by a scalable hub-and-spoke approach leveraging trusted professional organizations. Without intentional governance, AI risks amplifying disparities and eroding trust. Done well, it becomes a force multiplier-extending high-quality, equitable care into the communities that need it most. For community physicians, AI governance is not optional; it is essential to protecting patients, preserving clinical judgment, and ensuring that innovation advances equity rather than harm.
Human gut microbiome research has generated many disease associations, yet few translate into clinical applications. A central obstacle is not a lack of data, but the limited integration of causal reasoning, as most studies report correlations without establishing directionality, confounding control, or mechanistic evidence. We propose a unified causal inference framework that integrates directed acyclic graphs, Mendelian randomization, double machine learning, mediation analysis, and tests of causal reversibility into a single decision-aware workflow. Unlike prior applications of these tools in isolation, our framework explicitly separates assumption mapping, causal identification, effect estimation, and mechanistic interpretation, introducing "assumption guardrails" that constrain interpretation at each stage and prevent overinterpretation of observational findings. Using a colorectal cancer case study with public metagenomic data, we demonstrate how the framework operates under real-world constraints, transforming observational associations into testable, mechanism-based hypotheses. The contribution is architectural in that it organizes existing tools into a disciplined, integrated pipeline that clarifies the strength of evidence at each stage. This operational blueprint provides a reproducible path from correlation to causation in microbiome research and toward precision interventions.
To evaluate contemporary deep-reasoning large language models (LLMs) for early assessment of ophthalmic emergencies using sequential, workflow-mimicking information levels. Cross-sectional, vignette-based, head-to-head comparative evaluation. Thirty-four deidentified emergency ophthalmology teaching cases curated from a publicly accessible repository. Each case was reconstructed into 3 sequential information levels (level 1 [L1]: history; level 2: basic examination; level 3 [L3]: specialist examination). Six LLMs (Doubao, DeepSeek, Kimi-2, ChatGPT-5, Gemini-3, and Grok-4), operating in deep-reasoning mode, generated outputs that were independently scored by 2 ophthalmologists. Diagnoses were graded as fully correct, partially correct, or incorrect; triage category (typical vs. atypical emergency) was rated as correct or incorrect. Ancillary test recommendations were mapped to a prespecified 10-category taxonomy and classified as undertesting, exact match, or overtesting. A four-level composite outcome integrated diagnostic correctness, triage accuracy, and testing. Diagnostic correctness (fully correct, partially correct, incorrect), triage-category accuracy, ancillary test recommendation patterns, and composite outcome (ideal, safe but overtesting, potentially dangerous, intermediate). Across 612 model-case-level outputs, 46.9% of diagnoses were fully correct, 24.5% partially correct, and 28.6% incorrect. Fully correct diagnoses increased from 43.1% at L1 to 53.9% at L3 (P = 0.048). Overall triage category accuracy was 85.3% (range, 76.5%-94.1% across models; P = 0.003) and did not differ across information levels (P = 0.89). Ancillary test recommendations most commonly reflected undertesting (51.0%), followed by overtesting (27.5%) and exact matches (21.6%) (P < 0.001 across models). In generalized estimating equation pairwise comparisons, ChatGPT-5 showed higher odds of a fully correct diagnosis than DeepSeek (odds ratio [OR], 3.54; 95% confidence interval [CI], 1.49-8.43) and Gemini-3 (OR, 2.24; 95% CI, 1.31-3.83), and lower odds of potentially dangerous composite outcomes than DeepSeek (OR, 0.28; 95% CI, 0.10-0.74) and Gemini-3 (OR, 0.31; 95% CI, 0.11-0.89). Deep-reasoning LLMs demonstrated high triage category accuracy and moderate diagnostic performance for ophthalmic emergencies, with diagnostic correctness improving at higher information levels. However, ancillary testing patterns varied substantially, and ideal composite safety profiles were uncommon, supporting cautious, supervised deployment with explicit guardrails governing workup recommendations. The author has no/the authors have no proprietary or commercial interest in any materials discussed in this article.
This review aims to summarize recent literature on artificial intelligence (AI) tools for adolescent mental health, including the types of tools available, their clinical applications, effectiveness, and safety, as well as relevant ethical considerations. For clinicians, AI can facilitate clinical documentation, enhance therapy, and support the diagnosis process. Adolescents show interest in using AI for their mental healthcare and can benefit from AI-guided therapy apps and chatbots. Most studies that show effectiveness focus on depression treatment. Many tools are only in the prototyping stage, not tested on clinical samples, or lack safety measures, highlighting the need for further safety evaluation before specific app recommendations can be made. AI is increasingly being implemented in pediatric health systems and adolescents' daily lives. Adolescent medicine practitioners should recognize the growing potential for certain AI applications to enhance access and support adolescents, review and utilize those applications that have empirical support of efficacy, and that provide guardrails for safe and ethical use.