Artificial intelligence: Revolutionizing pediatric emergency care - A narrative review

ABSTRACT

Artificial intelligence (AI) refers to machines capable of imitating human cognition, with abilities to learn, apply logic and reasoning, and adapt to new information. The scope of AI in medicine ranges from prehospital triage to assisting in diagnosis and prognosticating patients. AI has shown incredible potential in pediatric emergency department by focusing on the development of clinical prediction models, triage systems, and diagnostic aids, contributing to higher accuracy and efficiency in patient management, along with hospital management, medical education, and training. Our review article discusses the current applications of AI in pediatric emergency and explores the barriers to AI in health care and ways to circumnavigate them moving forward. We aim to offer an insight into this less-explored world where technology meets the unpredictable and fast-paced environment of pediatric emergency medicine, building a future with a promise of innovation and redefining standards of care.

INTRODUCTION

Pediatric emergency care is a fast-paced field that demands quick triaging, prompt disease recognition, and precise treatment, all while minimizing unnecessary tests to save time and resources without compromising patient outcomes. In the conventional pediatric emergency setting, even after triaging is done, patients have to wait for an initial clinical assessment, preliminary tests, and reassessment by a clinician. These waiting periods increase crowding in the emergency department (ED), especially in hectic environments and lead to significant delays in initiating emergency treatment of patients in a conventional ED. In high-pressure environments, the likelihood of human error increases due to fatigue, stress, and distractions clinicians face. There are also biases that might influence judgment and could lead to misdiagnosis and further delays in the process.[1]

Artificial intelligence (AI) has emerged as an innovative approach and is increasingly finding applications in various medical specialties. Integrating AI algorithms into pediatric emergency protocols can complement traditional systems while reducing the strain on human resources. The widespread availability of electronic health records (EHRs) can be leveraged to develop machine learning (ML) models capable of assisting in potential diagnoses, thereby decreasing emergency room waiting times. However, a strong understanding of the fundamentals is essential before applying AI principles.

AI broadly refers to machines that are capable of exhibiting “intelligence” similar to that of humans. This can range from having certain traits that can make a machine “intelligent”, to possessing all of these traits and reaching a theoretical state where computers can essentially achieve or even exceed human intelligence. A typical artificially intelligent machine possesses not only mental skills such as learning, and reasoning but also, high acuity skills like self-correction and creativity. A better-suited term is “augmented intelligence” instead of “artificial intelligence,” which emphasizes the concept of an assistive role to the human brain.[2]

The idea of AI first emerged in 1949, when the English mathematician and computer scientist Alan Turing devised an “Imitation Game,” also known as the “Turing test.” The idea was to make a human interrogator have a conversation with a computer and another human. This test became the basis for work on AI in the decades to come. Subsequently, the term “Artificial Intelligence” was coined by John McCarthy in 1956.[3]

Since then, the field of AI has experienced both progress and setbacks. With every breakthrough, there has been a fair share of hindrances. However, the year 2020 marked the beginning of a massive AI boom, driven by the development of advanced algorithms and architectures. This in turn has led to the development and widespread popularity of large language models (LLMs) like ChatGPT which are capable of exhibiting human-like traits required for a machine to be considered “intelligent.”

This increased awareness has also led to the application of AI in other fields including medicine. While there are several studies exploring the potential of AI in the medical field, understanding of its utility is still deficient among health professionals. This can be primarily attributed to the knowledge gap between the domains of medicine and computer science. Many clinicians acknowledged their own and their colleagues’ limited understanding of AI. They also felt the need to address fundamental shortcomings of the current information systems of their workplaces such as EHRs before the integration of AI. In addition, concerns about reliability, transparency, and bias in AI models pose significant challenges to their adoption in healthcare.[4]

Some common terminologies need to be understood before moving ahead and filling these gaps [Table 1].

Table 1.

Common terminologies used with respect to artificial intelligence

Weak AI: Uses algorithms based on large datasets to perform specific, narrow tasks, such as medical diagnosis or interventions. It can simulate human behavior and cognitive processes but does not possess actual understanding or consciousness. Weak AI is limited to being a simulation and is not a true cognitive process[5]

Strong AI: Can perfectly emulate the actions of the human brain. Apart from performing narrow tasks, it also possesses true understanding and consciousness, which are higher mental abilities[5]

Big data: Describes datasets of large size and complexity, due to which they cannot be analyzed by traditional means of data-processing techniques. The characteristic features of big data include six versus; namely, volume, velocity, variety, veracity, variability, and value[6,7]

ML: The process of using large datasets to train a program and enable it to learn patterns and relationships from the data. By using ML, the program can make predictions, recognize patterns, or solve specific tasks without being overtly programmed for individual scenarios. Over time, the model refines its algorithms through continuous learning, improving its performance and accuracy in handling new or unseen data[8]

DNN: Computational models similar to the human brain, consisting of interconnected neurons (nodes), which receive inputs, calculate output, and pass it to other nodes. DNNs are trained on large, labeled datasets to refine their connections and understand the patterns in the data. For instance, a DNN used for diagnosing diseases might analyze datasets of symptoms, test results, images, and diagnoses to recognize patterns associated with specific conditions. Once trained, a DNN can predict outcomes for new and unseen data[9]

ML: Machine learning, DNN: Deep neural networks, AI: Artificial intelligence

Role of artificial intelligence in pediatric emergency medicine

AI can cause a remarkable impact in pediatric emergency care by improving diagnostic accuracy and streamlining workflows. From triage algorithms for prioritizing patient care to decision-making tools, AI can be a revolutionary asset. Apart from clinical management, it can prove valuable in education and research models, creating an indirect impact on the care of the vulnerable population [Figure 1].

Figure 1.

Applications of artificial intelligence in pediatric emergency

Clinical management

Triaging of patients

Patients coming into the emergency require targeted and timely medical interventions to maximize their chances of survival. However, overcrowding, long wait times, and overburdened healthcare professionals often hinder achieving good health outcomes in such settings. The leading pediatric triage systems include the Manchester Triage System, Paediatric Canadian Triage and Acuity Scale, Emergency Severity Index (ESI) version 4, Australasian Triage Scale, and modified pediatric early warning sign (PEWS).[5,6,7,8,10] These tests largely depend on the clinician’s judgment and their ability to discriminate remains suboptimal, with over-triage and under-triage as frequently encountered issues.[9]

In recent years, the advent of EHRs coupled with ML and AI has pushed the boundaries further with regard to ED triage. A large-scale prognostic study by Goto et al. demonstrated the superior discrimination ability of ML-based triage models, for clinical outcomes and disposition, compared with the conventional approaches.[11] Another study assessed the performance of ML-based medical directives for predicting and ordering tests to streamline patient flow in the ED. The proposed model was found to streamline care for a large proportion of patients (22.3%) and significantly reduce testing time by 165 min per patient.[1] Levin et al. used ML to develop an ED triage system (e-triage) to be used as a supporting aid to physicians. It was more accurate in predicting the need for critical care, emergency procedures, and hospitalization, compared to the widely used ESI. This resulted in better resource allocation and reduced wait time for patients.[12] A study on prehospital triage using random forest computer algorithm highlighted a reduction in over-triage from 66% to 42% when tested against traditional systems.[13]

The utility of AI-powered language processors such as GPT-4 has also been evaluated in this context. A study found almost perfect agreement between the triage team, GPT-4, and the gold standard, suggesting its use as a potential support tool.[14] AI and ML have also been used to devise new triage scores[15] and predictive algorithms[16] with promising results.

Clinical prediction models and clinical decision support

Clinical prediction models (CPM) use multiple predictors to assess the probability of a disease, its progression, or the occurrence of any adverse event in individual patients. In modern-day medicine, CPMs have become essential, aiding in early diagnosis, risk stratification, and overall better health outcomes. The development of these models mirrors our evolving understanding of disease processes and concurrent technological advancement. For example, the sequential organ failure assessment score has been adapted for the pediatric population as a basis for early recognition of organ dysfunction in a child with suspected sepsis.[17] CPMs have also been developed to effectively distinguish between close differential diagnoses, as demonstrated in a study to differentiate Multisystem inflammatory syndrome in children and Kawasaki disease.[18] Other robust prediction scores include signs of inflammation in children that can kill, PEWS resource-limited settings, pediatric index of mortality, pediatric early death index for Africa-immediate score, and pediatric logistic organ dysfunction.[19]

Conventionally, such models and risk stratification have been based on logistic regression applied to prospectively collected data. The advent of AI has made it possible to develop newer models using larger datasets with more complex relationships among predictors.[20] Studies report that AI and ML use fewer predictors while maintaining accuracy at par with traditional predictive methods.[21,22,23] A pediatric-electrocardiogram (ECG) based deep learning (DL) model was shown to outperform the expert cardiologist’s benchmark in predicting left-ventricular hypertrophy.[24] Another DL-based model for predicting the need for invasive mechanical ventilation in neonates proved superior in its predictive, recall, and alarm performances.[25] Incorporating these models in clinical practice has been shown to assist in the development and application of clinical decision support systems (CDS/CDSS).

CDS/CDSS assists physicians in clinical diagnoses by providing a repository of systematic medical knowledge in conjunction with an in-depth analysis of EHRs, greatly improving upon the accuracy of diagnosis and standard of care being provided.[26] Previously developed systems follow a rule-based approach, while recent advances have prompted the inclusion of AI in these systems. AI and ML-based systems have an advantage over their predecessors through increased model accuracy, with fewer missed patients and false alerts.[27] Figure 2 shows an example of the utilization of AI in clinical applications.

Figure 2.

Example of utilization of artificial intelligence in clinical application

In the context of the pediatric population, one pivotal example is the Electronic Pediatrician, a patented Romanian software that is a non-machine-learning AI system, simulating the pathophysiological reasoning of a physician, working with a human-generated knowledge base. It analyses the clinical findings and paraclinical picture of the patient and reports possible “clusters” (pathophysiological macro-links) and diagnoses. It is primarily used for infectious diseases but is flexible and has widespread application even in the ED.[28]

Interpretation of investigations

In the emergency setting, every minute is crucial. After a thorough evaluation, appropriate and timely investigations decide the future course for the patient. Over the years, AI has found various applications in the field of medical imaging, given its ability to recognize patterns and continuously learn from the large volume of information available.[29] These AI-based models are capable of both detecting anomalies and diagnoses. Earlier models were more rule-oriented, with little scope for any automatic improvement. With the advent of DL-based systems, algorithms are capable of modeling abstractions with high accuracy and no predetermined inputs.[30,31]

A systematic review on AI-based computer-aided design in pediatric radiology found extensive application, with brain, respiratory, and musculoskeletal imaging being the most popular areas.[32] Point of care ultrasound is largely skill-based and regular assessment of competence helps maintain quality of care. AI has now been incorporated to aid new users of this modality. It has shown promising results in training users, helping produce higher-quality images.[33,34] The role of AI in the interpretation of images is also an area of focus. It is shown to aid in earlier and more accurate diagnosis.[29,30] There is also proof of substantial improvement in clinical workflow.[35]

Another simple, cost-effective investigation that greatly benefits from AI and ML is the ECG. It has predictive, monitoring, and diagnostic value. It is capable of identifying abnormal ECG and combining it with clinical analysis with high accuracy.[36,37,38,39] It also shortens data processing time, making it a highly efficient system.[40,41]

Documentation

Healthcare workers (HCWs) are required to document their observations, decisions, actions, and interpretations in a systematic manner. This ensures ease of communication between HCWs, strict quality control of care being provided, and legal liability on a professional to provide justice to their patient. A phenomenon commonly associated with documentation in the healthcare setting is Documentation Burden or “DocBurden,” which leads to significant adverse effects on the HCWs and the patients in the healthcare system.[42,43,44]

In the ED, a foundational element is the discharge summary, which provides important information about patient transfer and further instructions for management. Since this process has become extensive and time-consuming, often the discharge summaries are less than adequate or incomplete.[45] The recent development of LLMs like ChatGPT has introduced the possibility of their use in the clinical context, particularly text summarization. Studies have evaluated various LLMs and their rates of error in generating discharge summaries and have found a reasonable amount of accuracy in their performance.[46,47] Similar studies have also been done to analyze the performance of these LLMs in generating a history of presenting illness and summarizing doctor–patient interaction.[48] One of these studies also found these models to be superior to clinicians in this regard.[49] A study done at Stanford Healthcare detailed the utility of “ambient AI scribes” from a physician’s perspective. The results proved favorable in terms of usability, efficiency, quality of documentation, and decreased perceived burnout.[50] Conversational AI provides an ideal solution to decrease the effort of documentation while upholding quality. This provides compelling evidence for the case of AI assistance in the context of documentation.

Patient disposition from the emergency department

Being in one of the most high-strung areas, HCWs often find themselves making split-second decisions without substantial information due to paucity of time. ED disposition is the process of deciding whether the patient is to be discharged, or to be managed as an in-patient, once they have been adequately assessed and managed in the ED. Challenges such as overcrowding,[51] suboptimal triaging,[52,53] incomplete documentation,[45] diagnostic error,[54] and time and resource constraints[55] often lead to unsafe or failed ED disposition.

AI and ML have gained traction due to their superior ability to predict outcomes in various clinical settings as compared to conventional methods, with agreeable accuracy.[55,56] Accurate disposition predictive models can help in reducing “boarding delays” by proactive initiation of admission processes.[57] Another interesting proposition has been to use natural language processors to predict disposition outcomes based on ED triage notes.[58] Kuo et al. demonstrated the development of an ensemble model that can make predictions of more than two possibilities of disposition (discharge vs. admission), building upon the existing ML-based ED disposition models.[59] It is also useful to look at ED readmission after discharge, which is also a significant burden in the ED.[60]

Equipment and drug dosage optimization

Drug dose optimization is required to ensure the risk-benefit balance to the patient. Patient factors, disease, and healthcare resources all influence this process. The challenge is even greater for the pediatric population since a significant proportion of drug testing is done on the adult population, leading to lack of evidence in the younger age groups. The concepts of therapeutic drug monitoring and model-informed precision dosing, though not new, have always benefited from the advancements in computational technology, the latest being the amalgamation of AI and ML in healthcare.[61] The current standard for describing drug exposure over time in patients has been the population pharmacokinetic model. The integration of these existing models with AI takes advantage of the biological plausibility of the former and superior predictive ability of the latter. Woillard et al. shared their construction of a concentration and exposure predictive model using AI, that fared better than the traditional Bayesian estimation approach.[62] Another area that has benefited from AI is drug dose prediction. This is especially beneficial for drugs requiring close monitoring, such as vancomycin and warfarin.[63,64,65]

A novel AI-derived platform called CURATE. AI is a noteworthy example of dose optimization strategy, which functions by predicting the effect of the input (dose) and the output (effect of dose) for an individual. It is a dynamic platform, changing according to disease progression, medical interventions, differing regimes, and other changes, throughout the course of treatment. It is thus well-suited to for drug dose optimization, polypharmacy, immunotherapy, and prospective dosing.[66] In the domain of personalized medicine, polypharmacy remains a big challenge. Another study assessed an artificial pharmacology intelligence system, which was able to accurately provide clinical insights to reduce medicine-related side effects.[67] There is evidence to suggest that genetic testing backed by AI can be used to predict adverse drug reactions, and subsequently help in choosing the safer and more suited drug for individual patients.[68]

AI has also shown promising results in optimizing equipment used in the healthcare setting for better patient outcomes. For example, Bachtiger et al. discussed the ability of a DL system applied to an ECG-enabled stethoscope examination in being a point-of-care screening test for heart failure.[69] Liu et al. outlined the utility of an AI-enabled alarm system to detect acute myocardial infarction early, thereby fast-tracking the initiation of treatment.[70] On a personalized level, the risk for cardiovascular events is being predicted by applying AI on wearable devices.[71] In the realm of imaging, the development of AI has improved attenuation and scatter correction to produce better-quality images in PET and SPECT scans.[72] Nimri et al. described optimizing insulin pumps with AI led to a similar glycemic control as that achieved by specialists.[73]

Hospital management

Globally, there has been a steady rise in the need for higher quality of care owing to the rising clinical burden. This has made medical resources scarce, especially in developing countries. Therefore, it becomes imperative to optimize the allocation of these resources. A study from China highlighted a commonly faced problem in tertiary care centers, “three longs and one short” (long registration time, long wait times, long time to pick up drugs, and short visiting time). Their survey described the pediatric emergency setting as particularly susceptible to burden. The authors proposed a convolutional neural network to study congestion in the ED and base the workflow on its findings.[74] Proposed uses also include efficient use of the hospital staff time by automating tasks such as managing patient flow, scheduling appointments, tracking bed availability, generating bills, and managing insurance claims.[75] Algorithms have also been successful in analyzing and predicting hospital demand and allocating healthcare personnel accordingly.[76] There is evidence to suggest that AI-driven decision-making can result in decreased operational costs.[77] The predictive nature of AI makes its role pivotal in analysis and strategizing for maximizing efficiency in management.

Research

AI is both a subject and a modality of research in the present scenario. It is being actively integrated into the research process because of its ability to assist researchers in formulating hypotheses, designing experiments, and analyzing and interpreting data.[78] ChatGPT, an example of LLMs, has been extensively studied in this regard. Corsello and Santangelo published an “interview” of the language processor, wherein it successfully highlighted its ability to change the world of pediatric research.[79] Another study found that ChatGPT is a useful assistive tool in scientific writing, improving clarity, grammar, and coherence of the text. It also effectively eliminates language barriers.[80] The utility of ChatGPT in summarizing research abstracts for laypersons to understand, thereby increasing accessibility to scientific literature has also been emphasized.[81] It was also found to be helpful in reviewing and summarizing medical literature, pharmaceutical research, and development.[82,83]

Medical education

Like many other fields, medical education has been revolutionized by the advent of AI. A scoping review into the utility of AI in medical education highlights the various applications of AI in this field, such as college admissions, teaching, assessment of students, and clinical exposure. Budding physicians are excited about the prospect of using AI in everyday practice as doctors, which needs to be met with a standardized curriculum devised to educate about AI and discuss its applications.

Xu et al. proposed an “AI-assisted” and an “AI-integrated” model of education to train students and practicing physicians. Their idea is to make the curriculum more knowledge-dense in a shorter duration and democratize medical knowledge. This serves as an upgrade on the traditional learning model.[84] ChatGPT and other LLMs have been explored in their use to train physicians in cracking clinical scenarios. In a study, the researchers were able to create inclusive clinical vignettes which saved a lot of labor and time on the part of the educator, while delivering high-quality results.[85] Using generative AI in medical imaging education also substantially benefits students’ understanding.[86]

The use of AI in medical education has also given room to provide one-on-one personalized training, tailored to the learners’ needs. A striking example is the use of virtual patient simulators to polish diagnostic skill and treatment, without risking patient harm. This is especially valuable in the emergency setting, where critical patients are the norm, thus making the environment nonconducive to learning.

Barriers to artificial intelligence

Despite the immense prospects of the integration of AI into pediatric emergencies, its adoption is not without challenges. Unique complexities of pediatric emergencies such as variabilities in age, weight, and developmental stages of patients complicate the standardization of AI tools. In addition, the need for robust validation, clinician trust, and integration into existing workflows further impede progress.

Gap in understanding of computers among physicians

ML relies on computers analyzing past data to devise algorithms that are used to solve problems, guided by the training data provided. Consequently, these techniques can inherit biases from the data and lack the ability to reason like humans when faced with unfamiliar conditions.[87] In the case of unsupervised DL, the model is essentially a “black box”, with no human record or history of how the algorithm has been developed.[88] As the process behind this decision-making is not transparent, there is bound to be a reluctance in acceptance of the technology by physicians. Therefore, collaboration between physicians and computers is essential. These models can only be helpful in the long run if physicians apply their human judgment, accounting for personal and societal factors, to assess and manage the risks associated with a certain outcome.

Bad/inaccurate dataset

Developing effective ML algorithms requires extensive, high-quality datasets that are accurate, reliable, and appropriately pre-processed for algorithm development.[89] Having a good electronic medical record data set is challenging, as many healthcare setups are still transitioning from traditional methods of recording data to the digitalization of health records. Specifically, in the case of emergency medicine, the high-paced environment increases the risk of inaccuracies in data entry.[88] Issues such as errors, missing values, and incomplete records can lead to incorrect conclusions. Furthermore, there is no straightforward way for humans to correct these errors during the training of an ML model, other than reviewing and preprocessing the training data using traditional methods, which can be time-consuming and resource-intensive.[90]

Bias

With the increased use of ML algorithms in different fields, there is ample evidence that suggests statistical as well as social bias in AI.[91] These may arise due to various reasons such as missing data and patients not identified by algorithms, misclassification, observational error, and misapplication.[92] As an example, AI is prone to inaccurately estimating risks for patients with incomplete data in EHRs. For instance, according to a study conducted by McCarthy et al., among women with breast cancer, black women were less likely than White women to undergo testing for high-risk germline mutations, despite having a comparable risk of carrying such mutations.[93] Consequently, an AI algorithm reliant on genetic test results is more likely to mischaracterize the risk of breast cancer for Black patients compared to White patients.[91] According to,[92] even if we assume the formulation of a “fair model” it could have the problem of “latent bias” which are biases that are not present at the moment but can happen at any time, due to the nature of AI. An algorithm could learn from existing social biases and enter a feedback loop, which can lead to biases worsening over time.

Liability/ethical concerns

Liability is a major concern when utilizing AI for medical diagnoses, as no algorithm can ever be entirely perfect and some degree of error is inevitable. The potential for errors or misdiagnoses raises critical questions about accountability and responsibility in clinical decision-making. For example, if an algorithm that reports ED imaging studies misses any abnormal findings, several parties, including emergency physicians, radiologists, algorithm developers, and healthcare institutions, could be held responsible.[88] Although case law is still under development for physician use of AI/ML, existing legal precedents suggest that physicians are typically held accountable for errors arising from AI/ML outputs.[94]

CONCLUSION

AI can be incorporated into pediatric emergency systems at various stages, driving significant changes. By analyzing extensive datasets, AI can predict outcomes and assist in clinical decision-making. Its potential to greatly improve care for children in emergency situations is evident. However, to achieve successful integration, efforts to strengthen implementation are necessary. Building comprehensive datasets through strong collaboration between specialties, institutions, and policymakers is crucial. Randomized trials should be carried out to provide evidence-based comparisons between AI-driven algorithms and traditional systems. Given the constant scrutiny of pediatric EDs and physicians, AI-based algorithms present additional challenges. Clear guidelines and frameworks are essential to address regulatory and ethical concerns. Policymakers must work to set standards for the safe and equitable application of AI in pediatric care. While the process is complex, it holds tremendous potential for advancing pediatric emergency medicine and child health.

Research quality and ethics statement

This study did not require approval by the Institutional Review Board/Ethics Committee. No applicable EQUATOR Network (http://www.equator-network.org/) guideline is available for narrative reviews.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.