Background Clinical toxicology research depends on reusable observational data, but toxicology information is often fragmented across systems and difficult to extract reliably. Learning health systems (LHS) use routinely collected data to generate knowledge and improve care. This study examined the conditions needed to support an LHS for clinical toxicology through reliable, clinically meaningful and reusable data. Methods We conducted a cross-sectional mixed-methods survey of professionals with expertise in clinical toxicology, poisons information services, emergency care, digital health, informatics, data governance, data engineering or health service data use. Survey items assessed six areas: data quality and standardisation, data access and governance, technical infrastructure, clinical workflow fit, organisational readiness, and future use and scalability. Open-text responses were analysed using descriptive content analysis and integrated with quantitative findings using a joint display. Results Thirty-four respondents initiated the survey, and 21 provided analysable Likert data. Section summaries showed strongest agreement for future use and scalability and technical infrastructure, and lowest agreement for data quality and standardisation. Strong support was observed for timely clinical access (19/19), co-designed governance, workflow and technical design (17/17), sustainability through standards and stewardship (16/16), automation with human validation (16/17), and a concise essential dataset (19/21). Agreement was lower for integrating clinical terminologies into extraction and reporting (5/21) and prioritising completeness and accuracy over real-time entry speed (10/21). Integrated findings identified clinically meaningful source data, automation with clinical validation, timely access, proportionate governance, reliable and adaptable infrastructure, workflow fit and sustained stewardship as conditions supporting reliable data extraction and use. Conclusion Establishing an LHS for clinical toxicology depends on source data quality, validation, governance, workflow fit and organisational capacity to sustain data use. These findings can guide data environment design for research, surveillance and quality improvement. Targeted, non-probability recruitment and the Australia-New Zealand sample may limit representation and transferability.
Introduction Digital twin (DT) technology is increasingly used to support data-driven hospital operations, but evidence remains fragmented across workflows, technological infrastructures, levels of decision support, implementation stages, and methodological quality. Objective This study aimed to synthesise DT applications in hospital operations management and to develop a structured framework linking data infrastructure, workflow modelling, decision support, resource allocation, implementation, and operational resilience. Methods A structured review and quantitative content analysis was conducted across PubMed, Scopus, and Web of Science on 23–24 March 2026. Of 521 records identified, 456 remained after duplicate removal and 20 studies met the eligibility criteria following title/abstract and full-text screening. Studies were independently classified by two reviewers according to application domain, technological infrastructure, decision-support maturity, operational outcomes, and implementation stage, and were formally appraised for methodological and reporting quality using design-appropriate QuADS and JBI tools. Results Twenty studies were included. Among the 13 studies demonstrating operational decision-support capability, 2 (15.4 %) were descriptive, 3 (23.1 %) predictive, 5 (38.5 %) prescriptive, and 3 (23.1 %) adaptive. Implementation remained limited: 6 studies (30 %) were contextual/conceptual, 8 (40 %) simulation or proof-of-concept, 4 (20 %) hospital-context validation or pilot studies, and only 2 (10 %) demonstrated operational implementation. Among the 17 studies assessed with QuADS, the final consensus median score was 36/39 (IQR 31–38; range 21–38). Conclusions DTs are emerging as hospital information-system infrastructures for analytics-enabled, workflow-integrated, and human-supervised decision support. However, technical decision-support sophistication does not necessarily indicate real-world implementation or stronger evidence. Wider adoption requires interoperable data systems, workflow integration, governance, organisational readiness, and rigorous evaluation in operational hospital settings.
Aim To map artificial intelligence (AI) systems developed for triage in primary care and describe their characteristics, performance and clinical applicability. Methods A scoping review was conducted in accordance with PRISMA-ScR. PubMed, MEDLINE (via Ovid), Embase, Web of Science, CINAHL (via EBSCOhost), IEEE Xplore and the ACM Digital Library were searched until March 14th 2026. Two reviewers independently screened studies and extracted data on system characteristics, triage function, performance, validation, safety, equity and implementation. Each system was charted by triage function and by whether it was patient-facing, clinician-facing or fully automated. Results Seventeen studies were included, published between 2013 and 2026 across 12 countries (14/17 high-income). Most were clinician-facing decision support (14/17), three had a patient-facing component and none were autonomous. Diagnostic or screening triage was the most common function (n = 7), followed by prioritisation (n = 4) and referral routing (n = 4). Only two performed urgency classification or self-triage. As a clinician aid, AI improved performance (fundus accuracy 82.8 % to 91.1 %; dermoscopy management sensitivity 72.9 % to 94.2 %), but standalone models sometimes fell below clinicians. Reported accuracy ranged from 41.2 % to 98.28 % and AUROC from 0.65 to 0.90, on heterogeneous outcomes. Validation was mostly internal (11/17), external in four. None reported a prospective randomised comparison or measured patient outcomes. Conclusion AI triage in primary care is concentrated on clinician-facing diagnostic and referral tasks and is most valuable for augmenting clinician workflows, particularly non-specialists. The evidence base remains immature, constrained by internal validation, unreliable reference standards, sparse equity assessment and absent prospective outcome data. Future work should prioritise prospective, outcome-anchored evaluation, with attention to safety and equity.
Background Digital tools for pharmacovigilance (PV) have been widely promoted to strengthen adverse drug reaction (ADR) reporting in low- and middle-income countries (LMICs). Despite growing evidence on the implementation of digital pharmacovigilance tools, limited evidence has examined the factors influencing their institutionalization in LMICs. This scoping review examines the institutionalization of digital PV tools in LMICs and identifies factors shaping their long-term integration. Methods A scoping review was conducted following PRISMA-ScR guidelines. 2 reviewers independently screened studies identified through MEDLINE, Embase, Web of Science, Google Scholar, and selected grey literature. Eligible studies examined digital PV tools implemented in LMICs and reported implementation-related findings. Data were synthesized using an inductive-deductive thematic approach informed by the Non-adoption, Abandonment, Scale-up, Spread, and Sustainability (NASSS) framework. Results 21 studies from 14 LMICs were included of which the majority are in sub-Saharan Africa, with fewer studies from South/Southeast Asia, the Middle East, and none from Latin America. Evidence was heavily concentrated on technological features, user adoption, and short-term organizational readiness. Key facilitators included simplified reporting interfaces, offline functionalities, training, and alignment with WHO PV standards. However, persistent barriers were identified at organizational and system levels, including weak or absence of governance, limited feedback mechanisms, fear of blame, fragmented workflows, and heavy reliance on external donor funding. Evidence related to long-term embedding, such as routinization, domestic and sustainable financing, and institutional ownership, was sparse. 16 of 21 included studies (76%) examined tools at the implementation stage (pilot/trial/early routine use), 4 (19%) at the pre-implementation stage, and only 1 (5%) at post-implementation/sustained use. Conclusions Digital PV in LMICs is frequently treated as a technical intervention rather than a complex socio-technical and institutional change. Strengthening the long-term integration of digital PV may require greater attention to legal mandates, sustainable financing, organizational and professional capacities, and political commitment.
Background Large language models (LLMs) and multimodal foundation models are increasingly proposed for diagnostic decision support, symptom triage, report interpretation, and electronic health record (EHR)-grounded workflows. Because medical informatics is concerned with how information and communication technologies behave inside clinical practice, workflow evaluation and deployment readiness are central rather than optional. Objective To synthesise diagnostic LLM evidence through a medical-informatics lens and to propose a stage-gated framework, with an accompanying self-assessment instrument, for trustworthy clinical use. Methods We conducted a critical narrative review supported by targeted evidence mapping of 64 cited records published or updated primarily between 2019 and August 2026, with selected foundational sources retained when directly relevant and emphasis on LLM evidence from 2022 onward. No pooled estimates were calculated. Results Twenty empirical diagnostic or patient-facing evaluations were mapped to an eight-level evidence ladder: two at Level 1 (benchmark or examination), one spanning Levels 1–2, ten at Level 2 (case vignettes or synthetic reasoning), three at Level 3 (retrospective clinical records), and four at Level 6 (clinician-in-the-loop workflow), one of which was a prospective single-centre deployment that also reported operational outcomes. Published examples of bias measurement, improvement strategies, and implementation-oriented outcomes exist, but external validation, prospective silent-mode evaluation, controlled clinical-impact trials, and post-deployment surveillance remain rare. Recurring risks include hallucination, dangerous omission, premature closure, unsafe reassurance, poor calibration, automation bias, privacy leakage, prompt injection, and model drift. ConclusionsDiagnostic LLMs should be evaluated as clinical information-system components. We propose a five-gate deployment-readiness framework requiring intended-use specification, evidence matched to clinical risk, a tested safety case, workflow and interoperability readiness, and post-market surveillance, operationalised as a 29-item, use-case-stratified self-assessment instrument.
