This theme, development of AI, explores how AI and its uses in prehospital care have developed over time and considers the directions they may continue to grow in. Some articles noted how AI technology has progressed to become more advanced and manage larger quantities of data (35, 39, 43, 46). Most authors discussed how the results of their studies can drive future AI applications and research (36, 38, 41–43, 45–48), with some studies forecasting the potential of combining AI tools with statistical or other AI models (37, 38, 40, 43, 44). Key quotes from these studies are shown in Table 3.

Experts also noted the progression towards more advanced AI technology (ID 1, ID 5, ID 7), although some acknowledged there is still a need for further refinement (ID 1, ID 7, ID 9). The majority of experts believed that AI has the potential to bring significant benefits to the PECS of LMICs (ID 1, ID 3, ID 4, ID 5, ID 6, ID 7, ID 8), and almost all experts interviewed discussed the various roles AI may play in this (ID 1, ID 2, ID 3, ID 5, ID 6, ID 7, ID 8, ID 9). Table 4 highlights some relevant quotes from the interviews. Consultations with experts also revealed a divide in opinion on the future progression of AI models in the context of LMICs. Some experts believed that explainability will be a key consideration in the implementation of AI in this context, including simpler models and the development of explainable deep learning AI models (XAI; ID 3, ID 4, ID 5, ID 8), while another expert believed that generative AI applications will transform future AI use in prehospital care (ID 1).

Overall, AI development is on a continued upward trajectory as the technology becomes increasingly complex and sophisticated. While there are some differences in opinion amongst experts around key considerations on the future direction of AI in prehospital care, subsequent research can leverage off previous studies to promote further innovation.

Some articles discussed how implementing AI solutions in EMS can reduce response times, save money and minimise resource use (37, 43, 45, 49) allowing for more equitable resource allocation (43, 44). As shown in Table 5, authors referred to the current lack of resources in LMICs as an important barrier to adequate prehospital emergency care (44, 45, 48–50). However, few studies considered the ability of countries to access the computational resources that would need to be involved to implement AI in some settings (38, 40, 45, 47). Moreover, Yang et al. (43) proposed using a statistical model in favour of AI to combat these computational demands.

Table 6 highlights some relevant quotes related to this theme from the expert interviews. Just under half of experts interviewed acknowledged the advantages of using AI to tackle resource limitations (ID 1, ID 2, ID 3, ID 5). Despite several experts noting that current IT infrastructure and expertise may be a limitation for using AI in some settings (ID 1, ID 2, ID 5, ID 8), these potential physical infrastructure challenges could potentially be overcome by leveraging off cloud architecture and mobile health solutions (ID 2, ID 5). Experts also considered the infrastructure required for adequate data security to protect detailed EHR as a potential limitation in the implementation of AI in LMICs (ID 3, ID 5, ID 7). In addition, experts noted the importance of funding and support for a project (ID 3, ID 5, ID 6, ID 8), with one expert admitting that the economy comes first (ID 3).

AI is generally considered to be able to optimise resource allocation in prehospital care. This is particularly important in LMICs that often lack appropriate levels of resources. However, other factors such as funding and infrastructure must be evaluated before attempting to implement any AI tools.

Several studies noted that model precision could be improved with access to additional high-quality data (36, 42–44, 49, 50). With the majority of studies using retrospective datasets, some papers acknowledged the potential for bias in study design and participant selection (35, 39, 41, 42, 47). Minimising inherent biases in data collection methods and ensuring models are only trained on high quality data that represents the population it is serving, were seen by some as two key processes to limit the effect of data bias (39, 43, 47, 50). Relevant quotes from the literature are included in Table 7.

Expert opinions were largely congruent with study authors on the importance of, and how to limit data bias (ID 2, ID 3, ID 5, ID 7, ID 8, ID 9). Access to a high quality and quantity of representative data was considered amongst experts to be one of the most significant limitations to the implementation of AI in PECS in LMICs (ID 2, ID 3, ID 5, ID 7, ID 9). Some experts commented on the need for LMICs to digitise health records before they can embrace AI solutions (ID 5, ID 7, ID 8). Most experts warned of the risk of data bias and the consequences this could have on specific populations (ID 2, ID 3, ID 5, ID 6, ID 7). Examples of applicable quotes can be found in Table 8.

AI model accuracy can be improved by increased focus on data collection procedures. This includes ensuring you collect a large quantity of quality data, and that it is debiased, so that it better represents the population. LMICs aiming to use AI in their prehospital system, can facilitate algorithm access to larger quantities of data by digitising their existing health records.

Some studies advise that increased collaboration between researchers, developers and healthcare workers will increase the chance of successful implementation (43, 46, 47). However, factors such as human error in records, can also limit the effectiveness of AI (48). Table 9 highlights some applicable quotes around this subject and wider theme. One study accommodated for some human factors, such as ignorance towards AI decisions, to more closely resemble the differences in experiments and real life (38).

A majority of experts believe all AI tools should be designed with the human end-user in mind, as model precision can be negatively impacted by human decisions in real life (ID 2, ID 3, ID 4, ID 5, ID 7, ID 9). As demonstrated in some of the quotes shown in Table 10, almost all experts stated that they believe AI tools should be developed to work alongside humans and not replace them (ID 1, ID 2, ID 3, ID 4, ID 5, ID 7, ID 9). To ensure the successful implementation of AI tools into a healthcare system, most experts feel that there is a need to educate the end-user on the purpose and functioning of the model (ID 3, ID 4, ID 5, ID 7). This was believed to build trust in the technology amongst the users (ID 2, ID 3, ID 4). Some experts maintained that XAI can also establish trust in AI (ID 3, ID 5, ID 8). Enhancing collaboration between contributors on future projects was also seen by some experts as an important step towards improving AI tools (ID 4, ID 5, ID 7, ID 8).

Overall, integrating humans in the decision process of AI models is necessary to ensure successful implementation, but this can limit the accuracy of the model. To minimise the risk of this occurring, all AI models should be developed using human-centred design principles.

Table 11 displays relevant quotes related to this theme from the literature review. To ensure AI models benefit the population they have been built to help, they all should incorporate local context into their design and function. However, this is particularly important for models that deal with spoken language (47, 50). Some studies also discussed the importance of using contextually appropriate model designs (46, 47, 50).

Experts were keen to stress models should only be implemented for the local population they have been trained on, including considerations on important differences in urban and rural populations (ID 5, ID 6, ID 7). Applying the model outside of the local context would require retraining (ID 2, ID 4). Some experts believe social and cultural acceptability must be considered when implementing AI tools (ID 5, ID 6, ID 7, ID 9). Key quotations related to this theme are included in Table 12.

In conclusion, factors such as language and culture cannot be ignored when designing an AI tool. These and other context-specific factors must be taken into consideration to ensure successful implementation.

AI tools in PECS should have the overarching goal to improve patient outcomes by facilitating the treatment of patients as soon as possible after an acute injury or illness occurs. The results of this qualitative analysis suggest that attaining adequate amounts of high-quality data is fundamental to achieving this goal. However, acquiring the necessary level of detail in data can be difficult for some LMICs (51). Healthcare systems in some of these countries do not have access to personnel with adequate training in data collection and analysis (52). This is typically exacerbated in PECS, many of which often lack trained medical personnel (14). Ensuring workers are skilled and available to appropriately use AI tools has been a challenge in adopting AI previously (53). Even when personnel are appropriately trained, they can lack the system-wide technical infrastructure and therefore have limited capacity to use data on the frontline (52, 54). Organisational challenges such as financial limitations and inadequate government policies further limit accurate data collection (52, 55). In addition to facilitating the integration of AI, addressing structural barriers is important as LMICs seeking to improve the delivery of healthcare generally may benefit from improvements in data utilisation through a cyclical positive feedback process (54). By focusing on improving data utilisation as a means to improve a healthcare system, healthcare services seek to gather additional data. This demand can promote improvements in data collection and analysis methods, through technical infrastructure, training of individuals to ensure correct procedural technique, and additional organisational support (52, 56). The digitisation of written records into EHR is one way to support improved data collection and analysis (52). This should be a priority for health systems in LMICs to allow AI technology to be appropriately trained (57). Electronic healthcare systems can also lower costs, boost transparency and lead to efficiency gains (54, 58). These adaptations can result in an increase in information available for stakeholders to make informed decisions. Improving access to this information can allow more stakeholders to use all the available information to make more evidence-based health decisions, thereby promoting evidence-based policies (56). Adopting policies supported by research is seen as essential for the development of emergency care in LMICs (59). Finally, a greater desire for information to assist decision-making stimulates a greater demand for data, allowing the cycle to continue (56).

AI can also widen the scope of information available for evidence-based decisions by improving the efficiency and accuracy of data analysis processes by rapidly identifying patterns from large volumes of data (22). However, it is important that the use of information, and therefore the demand for data (56), is based on the needs of the population and not on the current availability of data and information (5). This can be accomplished by implementing AI models that are designed to address the specific and relevant challenges in prehospital care for a defined population or setting. Aside from the advantages directly associated with using AI in LMICs and the positive feedback mechanism discussed, upgrading existing digital healthcare infrastructure to meet the data demands of AI will likely result in other wider benefits for the health system. One example of this is cloud computing, which can provide leverage for AI and simplify its deployment in a healthcare system (60), while simultaneously improving the reliability and reducing the cost of data storage and management (1). Despite these benefits, many LMICs lack the sustained financial resources for AI implementation, with funding being prioritised for wider EMS upgrade costs instead (61). If executed correctly, AI augmentation may reduce healthcare costs through improvements in operational efficiency and quality care. Resource limited settings should prioritise areas that are lower cost, offer the greatest gains in efficiency and target diseases with the highest burden to maximise AI investment and simultaneously upgrade PECS (58). However, this can be expensive to undertake and countries implementing AI tools in PECS should consider both direct and indirect costs with this technology, ensuring rigorous procedures are in place to allow for sustained cost savings and efficiencies (62).

According to the study results, continued technological innovation is believed to improve AI models in PECS in LMICs. The ability to use XAI without limitations to performance is considered by some to play a central role in future applications by making currently hidden deep learning decision processes visible, promoting model reliability and physician trust (57, 63). Future AI models should not only demonstrate transparency in how the data is used to generate a final decision but should also clearly show where the data has been obtained. Algorithms based on data from a HIC will likely not reflect the population of a LMIC, and risk exacerbating established prejudices (64, 65). Even if data is collected from the population of a selected LMIC, factors such as aggregation bias or representation bias can ensure the collected data is not a true reflection of the population. Moreover, making an assumption about a single patient based on observations of an entire population is unlikely to properly accommodate for individual differences (66, 67). For example, if an AI tool to predict sepsis is based on data from the general population, a patient with a compromised immune system may be screened as healthy by the algorithm but be, in reality, very unwell. The model trained using observations from “typical” patients, i.e., immunocompetent patients, is unlikely to recognise an immunocompromised patient as unwell. Within local populations, vulnerable and marginalised groups often face more significant challenges in accessing technologies such as AI, meaning populations with a perceived higher social standing are more likely to be able to access and therefore benefit from new health innovations (68). Not only does this result in improved outcomes only for these patients, but the resulting algorithm is also trained and refined based on this sample, and therefore introduces bias into the tool and becomes more accurate for the convenient majority, potentially negatively impacting patient outcomes for other historically marginalised demographic groups (66, 69). To minimise this risk, AI tools and the corresponding algorithms must be regularly audited by independent third parties. These auditing processes should be continuous and adapted to appropriately assess the AI tool as it is developed and refined. It is essential that these evaluations assess the model’s performance within these vulnerable communities (70). These processes should follow the framework elements and principles outlined in the World Health Organisation’s guidelines on AI ethics and governance (70).

When correctly integrated into a specific setting, AI can improve the quality of and access to healthcare for some populations (65). For example, using AI may be particularly useful in multicultural settings where patients speak a diverse set of languages by providing real-time accurate translations of discussions between healthcare provider and patient. Some healthcare professionals view AI as potentially beneficial for patients who are normally isolated by language barriers, by improving the standard of care for these patients and providing additional support to the healthcare team, on top of gains in efficiency (71). This function can have further unintended benefits for health systems and staff as well. For dispatch services, AI may also reduce levels of staffing required to address language barriers, relieving emergency care teams in resource-limited settings (72). Moreover, automating certain processes in emergency care can even lead to improvements in job satisfaction amongst staff, by lowering levels of frustration and the perceived time spent documenting notes (73).

Although an essential factor in AI development, ethical considerations for equitable AI use should not be limited to ensuring representative datasets, but should also consider the availability of equipment and staff in the healthcare system (69). For example, should an AI tool advise intubation for a prehospital patient if there are no healthcare personnel available with the skills and knowledge to intubate? What if there are no intensive care beds available for the patient when they arrive at hospital? There is little advantage in improving the ability to diagnose a condition if a particular healthcare system lacks the resources to treat it (1). Ethical implications of AI use in the prehospital setting should be discussed and decisions made with consideration for current system capacity, as well as local culture and customs (69). This point has been previously disregarded in the technological research agendas implemented by Western actors that lack local contextual knowledge, who instead seek to engage and satisfy the needs of international donors more than those of the end-users (74).

Healthcare systems in LMICs that are seeking to implement AI will also need to address data security and patient privacy concerns that accompany the generation of electronically stored datasets (1, 13). AI models are reliant on multiple actors accessing patient information across different databases (13), potentially exposing confidential medical records to malicious threats (52). Previous studies have highlighted the difficulties of establishing privacy policies and enforcing medical law in LMICs (75). Although these issues exist in AI use in all areas of medicine, the diversity of patient presentations in PECS makes it difficult to develop generalised regulations for this domain. In addition, the imminently life-threatening nature of some patient presentations means that the ease of access to vital patient information such as past medical history and drug allergies is an important consideration when developing security measures for EHR. Despite this, the ease of accessing data must never be at the expense of patient confidentiality. AI developers will have to balance these ethical privacy challenges with the accessibility requirements of healthcare staff (1). Governments of LMICs will also have to manage energy security concerns if AI is implemented in the healthcare system. AI systems are reliant on large amounts of energy to process information through its servers, that then require cooling with clean, fresh water. Although some studies suggest AI can improve energy efficiency in LMICs (76), this computational demand combined with the rapid growth of large language models is expected to exceed any potential energy savings in the coming years (77). Balancing this net-positive energy expenditure of AI, coupled with growing international concerns about the effect of greenhouse gas emissions on climate change, against the possible advantages of implementing AI tools, will be a challenge for LMICs as they continue to rely heavily on fossil fuels and struggle to match energy supply with demand (76). Power cuts affecting AI tools that PECS have become reliant on may have devastating consequences for patient outcomes.

There is a need to successfully implement AI now so that it can continue to benefit society in the future. This is particularly important in LMICs, where preliminary research is demonstrating AI’s potential to augment current health service distribution (5). To address some of the current issues with AI utilisation in PECS in LMICs, there must be improved international partnership between stakeholders in prehospital EMS, government organisations, research and the medical technology field (15), to empower leadership from those with local, contextual experience to implement necessary changes. This will promote the safe, ethical and socially acceptable use of AI technology and ensure a smooth transition towards programs funded and run by local people (23). Collaboration should occur early in the design process, allowing governments and investors to make informed decisions about the implementation of this technology and strengthen the sustainability of the intervention (21). These discussions should continue after the implementation of the device, to make sure algorithms are appropriately retrained and adapt to the changing reality of their environment (51). All collaborators should ensure understanding in a tool’s applications and limitations, and should share collective responsibility for any implemented AI models (70). By ensuring humans remain central to decisions made while using AI, healthcare systems can also ease legal liability and ethical considerations in cases of patient harm (1, 57).

Stakeholders involved in these discussions should also develop a regulatory framework for AI use in these settings. This should set out the legal requirements for this technology to prevent models discriminating against marginalised populations (69). This framework should also consider how to ensure data is securely stored (52). Having governmental support for AI-friendly policies can further motivate healthcare systems to embrace AI solutions (78). In addition, prehospital emergency care staff that will avail of this technology should be educated on how any future AI tools will work before they have been implemented and used. Improvements in training should also be extended to those involved in data collection and analysis, with an emphasis on transparency at all stages of data collection, model training and implementation to make sure standards are upheld, evidence-based decision making processes are followed, and future developers learn from previous mistakes (12, 69, 79). There is also a need for more qualitative reviews of key actors, including patients, on the use of AI in prehospital settings. This should include studies on the acceptance of AI amongst key prehospital staff in LMICs, as well as how to ensure contextual acceptability.

This study had several limitations. Firstly, participants may have felt obligated to respond to questions in a socially acceptable way instead of sharing their true opinions. To try and limit this, the interviewer used a semi-structured approach with open and follow-up questions. Secondly, as participants were unable to provide feedback on the findings and therefore validate the results, it must be acknowledged that the results are based on the interviewer’s interpretation of the expert’s spoken word and may not directly mirror the intention of the expert. Using a single coder to identify themes in the data can also introduce the risk of interpretive bias in the analysis. To try and reduce this risk of bias, themes were discussed with all co-authors to ensure inter-author agreement. Finally, the sampling methods of snowball and purposive sampling mean the study sample may not be representative of the range of researchers and clinicians involved in this research area, reducing the external validity of the findings. Moreover, only nine participants were interviewed for the thematic analysis, possibly limiting the depth of identified themes. However, as this research area is only beginning to develop there are only a limited number of experts working on AI in PECS in LMICs, some of whom may not have fully developed strong international research links, making it difficult to identify experts through the selected sampling methods. This is also reflected in the geographical representation of interviewed experts. Most participants originate from Europe and Asia, with other LMIC regions such as Latin America underrepresented. This sampling bias may limit the applicability of the study’s findings in these underrepresented areas. This lack of expert availability is one of the reasons a scoping review method was applied, as it would be difficult to achieve true data saturation.