Role of artificial intelligence in virtual emergency care: a protocol for a systematic review

Abstract

Introduction

Artificial intelligence (AI) has the potential to revolutionise healthcare delivery, particularly in the domain of emergency medicine. With the rise of telemedicine and virtual care, AI-powered tools could assist in triage, diagnosis and treatment recommendations for patients seeking emergency care remotely. This systematic review aims to synthesise the current state of research on AI applications in virtual emergency care, identify key challenges and opportunities and provide recommendations for future research and implementation.

Methods and analysis

We will conduct a comprehensive search of multiple electronic databases (PubMed, Web of Science, Embase, CINAHL, MEDLINE, The Cochrane Library, Scopus) from each database’s inception to March 2025. The search will include terms related to AI, machine learning, deep learning, virtual care, telemedicine and emergency medicine. We will include original research articles, conference proceedings and preprints that describe the development, validation or implementation of AI models for virtual emergency care. Two reviewers will independently screen titles and abstracts, review full texts, extract data and assess risk of bias using the PROBAST (Prediction model Risk Of Bias ASsessment Tool) tool for prediction model studies, Cochrane Risk-of-Bias tool for randomised trials for randomised trials and Risk Of Bias In Non-randomised Studies of Interventions for non-randomised studies. Data synthesis will involve a narrative review of included studies, summarising key findings, methodological approaches and implications for practice and research. The results will be reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines.

Ethics and dissemination

No ethical approval is required for this systematic review as it will use only published data. The findings will be disseminated through publication in a peer-reviewed journal, presentations at relevant conferences and engagement with clinicians, health system leaders, policymakers and researchers. This review will provide a timely and comprehensive overview of the applications of AI in virtual emergency care to inform evidence-based guidelines, policies and practices for leveraging these technologies to enhance access, quality and efficiency of emergency care delivery.

PROSPERO registration number

CRD42025648202.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This systematic review will employ a comprehensive search strategy across seven major databases from their inception to March 2025, capturing a wide range of literature on artificial intelligence (AI) in virtual emergency care.

• The protocol uses established methodological frameworks including Preferred Reporting Items for Systematic Review and Meta-Analysis Protocols reporting guidelines, Population, Intervention, Comparison, Outcomes and Study Design framework for eligibility criteria, and validated tools (PROBAST (Prediction model Risk Of Bias ASsessment Tool), Cochrane Risk-of-Bias tool for randomised trials, Risk Of Bias In Non-randomised Studies of Interventions) for assessing risk of bias.

• The review incorporates a socio-technical perspective through the Human-AI Collaboration Framework, acknowledging that AI implementation involves complex human, organisational and societal factors beyond technical performance.

• A multidisciplinary team of experts in emergency medicine, virtual care, AI and systematic review methodology will ensure scientific rigour, clinical relevance and practical utility of findings.

• Limitations include potential challenges in synthesising heterogeneous AI models and outcomes, the rapid evolution of AI technologies potentially outpacing published literature and the exclusion of non-English publications which may introduce language bias.

Introduction

Emergency departments (EDs) worldwide face increasing patient volumes, acuity and complexity, leading to overcrowding, prolonged wait times and suboptimal outcomes.¹ In the USA alone, in 2022, there were an estimated 155 million ED visits in the USA, with a total ED visit rate of 47 visits per 100 people.² The COVID-19 pandemic has further strained emergency care systems, with surges in patient demand, limited resources and infection control challenges.³ In response, many healthcare organisations have rapidly expanded virtual care offerings, including telemedicine consultations, remote monitoring and digital triage tools.⁴

However, the sudden pivot to virtual care has also exposed gaps in the quality, safety and equity of these services.⁵ Emergency medicine is a high-stakes, time-sensitive and information-intensive specialty that requires rapid clinical decision-making under conditions of uncertainty.⁶ Traditional virtual care models, such as synchronous video visits or asynchronous symptom checkers, may not fully capture the complexity of emergency patient presentations or provide adequate support for clinical reasoning and risk stratification.⁷ Moreover, virtual care can exacerbate disparities in digital access, health literacy and trust among vulnerable populations, leading to delays in care-seeking and worse outcomes.⁸

Artificial intelligence (AI) has emerged as a promising solution to augment human decision-making and improve efficiency in healthcare delivery.⁹ AI encompasses a range of computational methods, including machine learning, deep learning, natural language processing and computer vision, that enable systems to learn from data and perform tasks that typically require human intelligence.¹⁰ In the context of emergency medicine, AI has been applied to various use cases such as triage, diagnosis, risk stratification, resource allocation and treatment recommendations.^{11 12} For example, machine learning models have been developed to predict hospital admission, critical illness and mortality among ED patients using structured and unstructured data from electronic health records (EHRs).¹³,¹⁵ Deep learning algorithms have been used to analyse chest X-rays and CT scans for the detection of life-threatening conditions such as pneumothorax, pulmonary embolism and intracranial haemorrhage.¹⁶,¹⁸ Natural language processing tools have been applied to automatically extract relevant clinical information from ED physician notes and generate standardised documentation.¹⁹,²¹

The integration of AI into virtual emergency care platforms has the potential to transform how patients access and receive timely, high-quality care. AI-powered chatbots and symptom checkers can guide patients through self-assessment and provide personalised recommendations for seeking care based on their reported symptoms, risk factors and preferences.²²,²⁴ Machine learning models can analyse patient-reported data, vital signs and visual information from smartphones or wearable devices to predict disease severity and prioritise cases for provider review.²⁵,²⁷ Natural language processing can extract relevant information from EHRs and generate automated documentation to facilitate care coordination and handoffs between virtual and in-person settings.²⁸,³⁰ Computer vision algorithms can analyse facial expressions, body language and environmental cues during video consultations to detect signs of distress, discomfort or domestic violence.³¹,³³ Collectively, these AI applications could help triage patients more accurately, diagnose conditions more quickly and personalise treatment plans more effectively, leading to better outcomes and lower costs.

Despite the growing interest and investment in AI for virtual emergency care, significant challenges remain in developing, validating and implementing these technologies in real-world settings. First, AI models require large, diverse and representative datasets for training and testing, which can be difficult to obtain given the heterogeneity of patient populations and clinical contexts.³⁴ Many existing AI models have been developed using data from a single institution or a limited set of patients, raising concerns about their generalisability and potential for bias.³⁵ Second, ensuring the safety, reliability and accountability of AI systems is critical to prevent unintended consequences and protect patient interests.³⁶ AI models can be prone to errors, drift and adversarial attacks, especially when faced with novel or extreme scenarios common in emergency care.³⁷ Third, integrating AI seamlessly into clinical workflows and decision support tools requires careful design, user testing and change management strategies to ensure adoption and sustainability.³⁸ Provider trust, liability concerns and alert fatigue are major barriers to the implementation of AI in emergency care.³⁹ Finally, addressing ethical and legal considerations around data privacy, informed consent, algorithmic fairness and human oversight is essential to foster public trust and social responsibility in the use of AI for virtual emergency care.⁴⁰

To address these challenges and realise the full potential of AI in virtual emergency care, there is a need for a comprehensive and critical synthesis of the current state of research in this rapidly evolving field. Previous reviews have focused on specific AI techniques,⁴¹,⁴³ clinical applications¹²⁴⁴,⁴⁶ or care settings,⁴⁷,⁴⁹ but have not provided an integrated view of the use of AI across the virtual emergency care continuum. Existing research highlights AI’s role in triage,^{46 50} workflow optimisation^{51 52} and emergency department efficiency,^{53 54} yet there remains a gap in understanding its broader implications across virtual emergency care. This systematic review aims to fill this gap by conducting a broad and inclusive search of the literature, appraising the quality and scope of evidence using established frameworks and generating actionable insights for researchers, clinicians and policymakers.

Theoretical framework

To guide the conceptualisation and interpretation of studies in this review, we will draw on socio-technical frameworks that view AI not as a standalone technology, but as part of a complex adaptive system involving human, organisational and societal factors.^{55 56} Specifically, we will use the Human-AI Collaboration Framework proposed by Frey and colleagues, which outlines five key dimensions for successful integration of AI into clinical workflows: task allocation, interface design, training and feedback, team dynamics and governance.⁵⁷ This framework emphasises the importance of designing AI systems that complement rather than replace human expertise, foster meaningful interactions and shared decision-making between clinicians and patients, provide explanations and opportunities for feedback and improvement, cultivate trust and psychological safety within teams and establish clear policies and accountability mechanisms. We will apply this framework to analyse the characteristics of AI models, the contexts of their development and use and the factors influencing their implementation and impact in virtual emergency care settings.

Objectives

The primary objective of this systematic review is to synthesise the current state of research on AI applications in virtual emergency care. Specifically, we aim to:

1. Identify and characterise the types of AI models and use cases that have been developed and evaluated for virtual emergency care settings.

2. Assess the methodological quality and risk of bias of included studies, focusing on aspects such as dataset characteristics, model performance metrics, validation approaches and implementation outcomes.

3. Summarise the key findings, trends and gaps in the evidence base, highlighting promising applications, challenges and future directions for research and practice.

4. Provide recommendations for stakeholders (clinicians, health system leaders, policymakers and researchers) on leveraging AI to enhance the access, quality and efficiency of emergency care delivery through virtual modalities.

Ultimately, this review aims to provide a comprehensive and critical assessment of the state-of-the-art and future directions of AI in virtual emergency care, and to inform evidence-based guidelines, policies and practices for the responsible and effective use of these technologies in the service of patient and population health.

Methods

Protocol registration

This protocol has been registered in the International Prospective Register of Systematic Reviews (PROSPERO) database (registration number: CRD42025648202) and will be reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) Protocols 2015 statement.⁵⁸

Review question and framework

The overarching review question is: What is the current state of research on AI applications in virtual emergency care, and what are the key challenges, opportunities and implications for future research and implementation?

To operationalise this question and guide the review process, we will use the Population, Intervention, Comparison, Outcomes and Study Design framework:⁵⁹

• Population: adult patients (aged 18 years or older) seeking or receiving emergency care through virtual modalities, including telemedicine, remote monitoring or digital triage tools, for any acute condition or symptom. We will exclude studies focusing solely on paediatric or neonatal populations, or those conducted in non-emergency care settings such as primary care, specialty clinics or inpatient wards. While studies focusing on paediatric or neonatal populations represent an important area of AI application in virtual emergency care, we have excluded them from this review to maintain focus and homogeneity in our analysis. Paediatric emergency presentations differ significantly from adult presentations in terms of symptom patterns, communication approaches and clinical decision-making processes. Including both populations would introduce substantial heterogeneity that could compromise the clarity and applicability of our findings. Future systematic reviews specifically focused on AI applications in paediatric virtual emergency care would be valuable to address this important population.

• Intervention: any AI model or application designed to support clinical decision-making, triage, diagnosis, risk stratification, resource allocation, treatment recommendations or other relevant tasks in virtual emergency care settings. This includes machine learning, deep learning, natural language processing, computer vision and hybrid approaches that combine multiple AI techniques or modalities. We will include both standalone AI tools and those integrated into existing clinical decision support systems, EHRs or virtual care platforms. Given the broad and evolving nature of AI applications, we will adopt an inclusive approach to defining AI interventions. We will include studies involving any computational method that demonstrates learning from data or automated decision-making capabilities, including but not limited to: machine learning algorithms (supervised, unsupervised, reinforcement learning), deep learning models (neural networks, convolutional neural networks, recurrent neural networks), natural language processing tools, computer vision systems, expert systems and hybrid approaches combining multiple AI techniques. We will categorise identified AI applications into major paradigms during data extraction to facilitate systematic analysis: (1) predictive modelling and risk stratification, (2) diagnostic support and image analysis, (3) natural language processing and documentation, (4) conversational AI and symptom assessment, (5) workflow optimisation and resource allocation and (6) monitoring and surveillance systems. This taxonomy will ensure comprehensive coverage while enabling structured analysis of different AI application types.

• Comparison: standard care without AI support, non-AI clinical decision support tools or alternative AI models or applications, depending on the study design and objectives. Studies without a direct comparator will also be included if they report relevant outcomes.

• Outcomes: the primary outcomes of interest are the performance and validity of AI models, including metrics such as discrimination (eg, area under the receiver operating characteristic curve (AUROC), sensitivity, specificity), calibration (eg, calibration plot, Hosmer-Lemeshow test), clinical utility (eg, net benefit, decision curve analysis) and equity (eg, performance across subgroups, fairness metrics). Secondary outcomes include implementation outcomes (eg, user satisfaction, usability, adoption, sustainability), patient outcomes (eg, mortality, morbidity, quality of life, experience of care), health system outcomes (eg, resource utilisation, costs, efficiency) and societal outcomes (eg, access to care, health disparities, social acceptability).

• Study design: we will include a broad range of study designs, including experimental (eg, randomised controlled trials), quasi-experimental (eg, interrupted time series, before–after studies) and observational (eg, cohort, case–control, cross-sectional) studies, as well as qualitative (eg, interviews, focus groups, ethnography) and mixed-methods studies.

Search strategy

We will conduct a comprehensive search of multiple electronic databases, including PubMed, Web of Science, Embase, CINAHL, MEDLINE, The Cochrane Library and Scopus, from each database’s inception to March 2025. The search strategy will be developed in consultation with a medical librarian and will include a combination of controlled vocabulary terms (eg, Medical Subject Headings) and keywords related to the following concepts:

1. Artificial intelligence: “artificial intelligence”, “machine learning”, “deep learning”, “neural network*“, “natural language processing”, “computer vision”, “automated decision support”, “clinical prediction model*”

2. Virtual care: “telemedicine”, “telehealth”, “virtual care”, “remote consultation*“, “digital health”, “mhealth”, “ehealth”, “mobile health”, “wearable*“, “biosensor*“, “smart device*”

3. Emergency medicine: “emergency medicine”, “emergency department*“, “emergency room*“, “emergency service*“, “acute care”, “urgent care”, “triage”, “diagnosis”, “risk assessment”, “resuscitation”, “critical care”

The specific search strings will be adapted for each database and will be based on the following template: ((“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network*” OR “natural language processing” OR “computer vision” OR “automated decision support” OR “clinical prediction model*“) AND (“telemedicine” OR “telehealth” OR “virtual care” OR “remote consultation*” OR “digital health” OR “mhealth” OR “ehealth” OR “mobile health” OR “wearable*” OR “biosensor*” OR “smart device*“) AND (“emergency medicine” OR “emergency department*” OR “emergency room*” OR “emergency service*” OR “acute care” OR “urgent care” OR “triage” OR “diagnosis” OR “risk assessment” OR “resuscitation” OR “critical care”))

The full search strategies for each database will be reported in the online supplemental materials. We will also hand-search the reference lists of included studies and relevant reviews, as well as the proceedings of key conferences (eg, American College of Emergency Physicians Scientific Assembly, Society for Academic Emergency Medicine Annual Meeting, IEEE (Institute of Electrical and Electronics Engineers) International Conference on Healthcare Informatics) and the websites of leading AI and healthcare organisations (eg, Google Health, Microsoft Healthcare, IBM Watson Health, HIMSS (Healthcare Information and Management Systems Society), AMIA (American Medical Informatics Association)) to identify additional eligible studies.

We will also include conference proceedings, preprints and grey literature sources that meet the eligibility criteria, to capture the latest developments in this fast-moving field.

Eligibility criteria

We will include studies that align with specific criteria to ensure relevance and rigour in our review. These criteria cover aspects such as the population, intervention, outcomes, study design, language and publication type. Similarly, we have outlined exclusion criteria to filter out studies that do not contribute to our research focus. The table 1 below summarises the inclusion and exclusion criteria.

Table 1. Inclusion and exclusion criteria.

Criteria	Inclusion	Exclusion
Population	Adults (≥18 years) seeking or receiving emergency care through virtual modalities	Paediatric/neonatal populations or non-emergency care settings
Intervention	AI models supporting clinical decision-making, triage, diagnosis, risk stratification, etc	Non-AI decision support tools or AI models not applied to virtual emergency care
Outcomes	At least one performance, validity, implementation, patient, health system or societal outcome	No relevant outcome reported
Study design	Experimental, quasi-experimental, observational, qualitative or mixed-methods studies	Reviews, commentaries, editorials or letters without original data
Publication date	From database inception to March 2025	N/A
Publication type	Original research articles, conference proceedings, preprints or grey literature sources	N/A

AI, artificial intelligence.

We will exclude systematic reviews and narrative reviews from our primary analysis to focus on original research findings. However, we will hand-search the reference lists of relevant systematic reviews identified during our search to ensure comprehensive capture of primary studies that meet our inclusion criteria. This approach allows us to benefit from the bibliographic work of previous reviews while maintaining our focus on original research contributions.

Study selection

Two reviewers will independently screen the titles and abstracts of all retrieved records for eligibility using Covidence software. Records that meet the inclusion criteria or require further assessment will be obtained in full text. The reviewers will then independently assess the full-text articles against the eligibility criteria. Any discrepancies will be resolved through discussion or consultation with a third reviewer. The reasons for exclusion at the full-text stage will be recorded and reported. The study selection process will be documented using a PRISMA flow diagram (figure 1).

Figure 1. Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow diagram.

Data extraction

For each included study, two reviewers will independently extract relevant data using a standardised form in Covidence. The data extraction form will be piloted on a sample of studies and refined as needed. The extracted data will include study characteristics such as authors, year, country, funding source, study design, setting, duration and sample size. Population characteristics will cover aspects like age, gender, race/ethnicity, socioeconomic status, comorbidities, acute conditions and severity. Intervention characteristics will focus on the AI model type (eg, machine learning, deep learning, natural language processing), data sources and inputs (eg, electronic health records, imaging, wearables), model architecture and hyperparameters, output type (eg, probability, risk score, recommendation), intended use and users and integration with existing systems. If applicable, comparator characteristics will be recorded, including the type of comparator (eg, standard care, non-AI tool, alternative AI model) and key features distinguishing it from the intervention.

Outcome measures will be assessed across multiple dimensions. Performance and validity will be evaluated using metrics such as discrimination (eg, AUROC, sensitivity, specificity), calibration (eg, calibration plot, Hosmer-Lemeshow test), clinical utility (eg, net benefit, decision curve analysis) and equity (eg, performance across subgroups, fairness metrics). Implementation aspects will be examined in terms of user satisfaction, usability, adoption, sustainability and barriers and facilitators. Patient-centred outcomes will include mortality, morbidity, quality of life, experience of care, decisional conflict and trust. Health system-related outcomes will focus on resource utilisation, costs, efficiency, patient flow and safety events, while societal outcomes will consider access to care, health disparities, social acceptability and ethical and legal issues. Additionally, key findings such as main results, statistical analyses, subgroup analyses and sensitivity analyses will be documented. The conclusions section will provide an interpretation of the results, generalisability, implications, limitations and future directions. (Details in appendix 1 in online supplemental file 1).

Disagreements in data extraction will be resolved through discussion or adjudication by a third reviewer. Missing or unclear data will be requested from study authors where feasible.

Risk of bias assessment

We will assess the methodological quality and risk of bias of each included study using tools appropriate for the study design. For prediction model studies, we will use the most current version of prediction model risk assessment tools. Where available, we will apply PROBAST+AI (Prediction model Risk Of Bias ASsessment Tool for Artificial Intelligence" + "Artificial Intelligence"),⁶⁰ an extension of PROBAST⁶¹ specifically designed for AI and machine learning prediction models, which addresses unique considerations such as data preprocessing, model interpretability and algorithmic bias. For studies where PROBAST+AI is not applicable, we will use the standard PROBAST tool. Additionally, we will incorporate reporting quality assessment using TRIPOD+AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis+Artificial Intelligence) guidelines where applicable, to evaluate the completeness and quality of AI model reporting.⁶² For randomised controlled trials, we will use the Cochrane Risk-of-Bias tool for randomised trials,⁶³ which assesses the risk of bias arising from the randomisation process, deviations from intended interventions, missing outcome data, measurement of the outcome and selection of the reported result. For non-randomised studies, we will use the Risk Of Bias In Non-randomised Studies of Interventions tool,⁶⁴ which covers seven domains: confounding, selection of participants into the study, classification of interventions, deviations from intended interventions, missing data, measurement of outcomes and selection of the reported result. For qualitative studies, we will use the Critical Appraisal Skills Programme Qualitative Checklist,⁶⁵ which includes 10 questions covering issues such as appropriateness of the research design, recruitment strategy, data collection, researcher-participant relationship, ethical issues, data analysis and reporting of findings.

Two reviewers will independently assess the risk of bias for each study, with disagreements resolved through discussion or consultation with a third reviewer. The results of the risk of bias assessment will be presented graphically and narratively, and will inform the interpretation and confidence in the review findings.

Data synthesis

We will perform a narrative synthesis of the included studies due to the anticipated heterogeneity of AI models, clinical applications, outcomes and study designs. The synthesis will focus on identifying patterns, similarities and differences across the evidence base, structured around the review objectives and guided by the Human-AI Collaboration Framework.⁵⁷ We will use tables, figures and data visualisations to summarise the key characteristics, findings and quality of the studies, highlighting strengths, limitations and gaps in current research.

The synthesis will cover the following dimensions: (1) characteristics and types of AI models and virtual emergency care applications, discussing advantages and disadvantages of different AI approaches; (2) methodological quality and risk of bias, identifying limitations and making recommendations for future studies; (3) performance and validity of AI models, exploring factors associated with success and challenges in evaluation; (4) implementation approaches and challenges, assessing usability, acceptability, adoption and regulatory, legal and ethical issues; (5) impact of AI on patient, health system and societal outcomes, considering unintended consequences and trade-offs; and (6) gaps, challenges and future directions for AI research and implementation, proposing a prioritised research agenda and envisioning the future landscape.

Throughout the synthesis, we will maintain a critical and reflexive stance, acknowledging the complexity and context-dependence of AI and virtual emergency care, and the limitations and biases in the review process. We will follow best practices for transparent and reproducible reporting, as per the⁶⁶ PRISMA 2020 statement,⁶⁶ and will make data and code publicly available where possible.

Where sufficient homogeneity exists in terms of study populations, AI interventions, comparators and outcomes, we will consider conducting quantitative meta-analysis using random-effects models. We will assess clinical and methodological heterogeneity using the I² statistic and χ² test. If a meta-analysis is feasible, we will conduct subgroup analyses based on AI model type, clinical application, study design and risk of bias level. However, we anticipate that the diversity of AI applications and outcome measures may limit opportunities for meaningful quantitative synthesis, and our primary approach will remain narrative synthesis guided by the Human-AI Collaboration Framework.

Ethics approval and dissemination

Ethical approval is not required for this systematic review as it will only involve the analysis of previously published data. However, for any research integrity or methodological inquiries regarding this systematic review, please contact the Research Office at the National University Health System Singapore at research@nuhs.edu.sg. The findings will be disseminated through publication in a peer-reviewed journal, presentations at relevant conferences and engagement with clinicians, health system leaders, policymakers and researchers. This review will provide a timely and comprehensive overview of the applications of AI in virtual emergency care to inform evidence-based guidelines, policies and practices for leveraging these technologies to enhance access, quality and efficiency of emergency care delivery.

Patient and public involvement

None.

Discussion

This protocol presents a systematic and rigorous methodology for conducting a comprehensive review of the research evidence on AI applications in virtual emergency care. The COVID-19 pandemic has accelerated the adoption of virtual care models and highlighted the potential for AI to support clinical decision-making and improve patient outcomes, but also exposed the challenges and limitations of current approaches.^{67 68} As emergency departments continue to face increasing demands and resource constraints, there is an urgent need to critically appraise and synthesise the existing knowledge base to inform future research, policy and practice in this rapidly evolving field.⁶⁹

The strengths of this protocol include its use of a comprehensive search strategy across multiple electronic databases, clear and explicit eligibility criteria, standardised data extraction and quality assessment procedures using validated tools and a structured and transparent approach to data synthesis and reporting following best practices.⁶⁶ The involvement of a multidisciplinary team of experts in emergency medicine, virtual care, AI and systematic review methodology will ensure the scientific rigour, clinical relevance and practical utility of the review findings.

However, the protocol also has several limitations that should be acknowledged. First, the focus on peer-reviewed literature may exclude relevant studies in non-biomedical domains. Second, the anticipated heterogeneity and variability of AI models, virtual emergency care applications and outcome measures may pose challenges for data synthesis and interpretation, and may limit the generalisability and comparability of the findings across contexts. Third, the rapid pace of technological innovation and the evolving nature of the field may require regular updates and revisions to the protocol and the review to maintain their currency and relevance.

Despite these limitations, this protocol represents a timely and important contribution to the methodological literature on conducting systematic reviews of AI applications in healthcare. By providing a clear and detailed roadmap for the identification, appraisal and synthesis of the available evidence, the protocol can serve as a useful template and guide for researchers, clinicians and policymakers seeking to understand and evaluate the scope, quality and impact of AI in virtual emergency care and related domains. The protocol also highlights the importance of taking a socio-technical perspective that considers not only the technical performance and clinical effectiveness of AI models, but also their usability, acceptability and compatibility with existing care processes and workflows, as well as their broader ethical, legal and societal implications.

Furthermore, the protocol emphasises the need for a critical and reflexive approach to evidence synthesis that acknowledges the complex and context-dependent nature of AI and virtual emergency care, as well as the inherent limitations and biases of the review process itself. By explicitly reporting the assumptions, judgements and decisions made at each stage of the review, and by engaging with diverse stakeholders and perspectives throughout the process, the protocol aims to enhance the transparency, accountability and credibility of the review findings, and to foster a more inclusive and participatory approach to knowledge generation and translation in this field.

This protocol presents a rigorous and comprehensive methodology for conducting a systematic review of AI applications in virtual emergency care that addresses a pressing need for evidence-based guidance in this rapidly evolving field. By critically appraising and synthesising the available research evidence, identifying key gaps and challenges and proposing a roadmap for future research and implementation, the review will inform and advance the responsible and impactful development and use of AI in emergency care. The protocol also contributes to the broader methodological literature on conducting systematic reviews of AI in healthcare, and highlights the importance of taking a socio-technical, critical and reflexive approach to evidence synthesis in this complex and dynamic domain. As such, the protocol represents an important foundation for future research, policy and practice aimed at realising the full potential of AI to improve the quality, safety and equity of emergency care delivery in the digital age.