From scores to signals: evolution and innovations in pediatric early warning systems

Article information

Acute Crit Care. 2026;41(2):252-261

Publication date (electronic) : 2026 May 19

doi : https://doi.org/10.4266/acc.004475

Wonjin Jang^1,2, Bongjin Lee^,1

¹Department of Pediatrics, Seoul National University Hospital, Seoul, Korea

²Department of Biomedical Engineering, Seoul National University College of Medicine, Seoul, Korea

Corresponding author: Bongjin Lee Department of Pediatrics, Seoul National University Hospital, 101 Daehak-ro, Jongno-gu, Seoul 03080, Korea Tel: +82-2-2072-3568 Fax: +82-2-747-5130 Email: pedbjl@snu.ac.kr

Received 2025 September 23; Revised 2026 February 9; Accepted 2026 February 11.

Abstract

Infographics

INTRODUCTION

Early recognition of clinical deterioration in pediatric patients admitted to general wards is associated with improved outcomes by preventing critical events such as cardiac arrest, unplanned intensive care unit (ICU) admission, or even death [1-4]. However, detecting early signs of decline in hospitalized children is often challenging. This difficulty is largely due to physiological variability in pediatric patients and their limited ability to clearly express symptoms [3,5-7]. These factors can lead to missed or misinterpreted warning signs indicating sudden deterioration, resulting in delayed interventions and, consequently, poor outcomes [5-7].

Rapid response systems (RRS) have been introduced in response to these challenges. The primary goal of RRS is to facilitate early identification and prompt intervention before deterioration becomes life-threatening for patients in non-critical care settings. To achieve this goal, hospitals worldwide have invested in implementing, refining, and evaluating RRS over the past two decades, and these systems continue to evolve [8]. This narrative review explores approaches to enhance the early detection of clinical deterioration, focusing on the development of structured systems by examining key historical publications and summarizing recent technological advances in pediatric patient monitoring through a review of the most contemporary studies indexed in PubMed.

HISTORICAL PERSPECTIVE

The development of effective early detection strategies for clinical deterioration in pediatric patients has been a progressive endeavor (Table 1). Early initiatives were motivated by the recognition that observable physiological changes typically precede critical events by several hours in both adult and pediatric populations [5,7,9]. This interval serves as an opportunity for healthcare providers to intervene proactively and potentially alter the clinical trajectory during admission. Furthermore, evidence shows that factors contributing to critical events are often avoidable. For example, a study in the United Kingdom estimated that a significant proportion of pediatric in-hospital deaths involved potentially avoidable contributing factors, with figures approaching 50% [10,11]. Such findings have spurred the development and implementation of structured early warning systems.

Table 1.

Key milestones in the evolution of early warning and rapid response systems

Initial early warning systems were adapted from adult systems, such as the Modified Early Warning Score [5]. However, these early tools proved to be inadequate for pediatric populations due to children’s unique and variable physiology, requiring different thresholds and considerations for children. In response, pediatric-specific systems were developed. The Brighton Pediatric Early Warning Score (PEWS) [5] and Duncan et al.’s [12] PEWS in the early 2000s were among the first to systematically combine vital signs and clinical observations such as behavior and respiratory effort into composite scores. These systems formalized the early warning model, which relies on regular monitoring of patient status, structured assessment, and predefined intervention protocols triggered when concerning patterns appear in hospitalized patients.

As the field progressed, two distinct types of systems became mainstays of operationalizing the early warning approach: parameter-based activation and scoring-based activation [10]. Parameter-based systems, like the Bristol PEWS and the Medical Emergency Team Activation Criteria in Australia, trigger an alert when any physiological measure exceeds a predefined threshold [10,13,14]. In contrast, scoring-based systems, such as the Cardiac Children’s Hospital Early Warning Score and the Bedside PEWS combine physiological measurements and clinical observations into a single composite score that stratifies patient risk and determines the appropriate level of response [10,15,16]. While these systems differ in design, thresholds, and complexity, they share a common goal of improving the timely detection of deterioration of pediatric patients admitted to general wards (Table 2) [3,5,12-16]. Importantly, these tools were developed within heterogeneous clinical environments and therefore differ in intended scope. Some were optimized for general ward surveillance, whereas others reflect population-specific priorities or institutional workflows, such as cardiovascular-focused parameters in cardiac care settings.

Table 2.

Summary of notable PEWS variants and associated clinical outcomes

Increasingly, such scoring systems have been integrated within broader RRS frameworks, implicitly acknowledging that detection alone is insufficient without timely response and system-wide support [17]. Modern RRS frameworks were refined to include four key components: an afferent limb for detection, an efferent limb for response, a supporting structure for ongoing quality improvement, and governance to oversee system function [11]. Examples of RRS implementation in pediatric hospitals include the Acute Response System described by Jeon et al. [3] in South Korea, which significantly reduced life-threatening events outside the ICU, and the Bedside PEWS integrated into a quality improvement initiative in Saudi Arabia, which improved the timeliness of ICU transfers and overall patient outcomes [1]. Additionally, large-scale initiatives like the National Health Service 100,000 Lives Campaign in the United Kingdom further accelerated the adoption of structured RRS incorporating PEWS across multiple hospitals [18]. These collective efforts highlight how early warning tools have evolved not in isolation but as part of broader institutional strategies to enhance pediatric patient safety in non-critical care settings.

EFFECTIVENESS AND CONTROVERSIES

Despite decades of development and widespread adoption, debate persists regarding the real-world effectiveness of early warning systems. While some studies have shown encouraging results following implementation of RRS, systematic reviews and meta-analyses suggest that overall effectiveness remains inconsistent across settings [7,8,10].

Among the studies that reported encouraging results, Jeon et al. [3] found that RRS implementation over a period of 11 years decreased CPR and unplanned ICU admission rates in the general ward. Similarly, Al-Harbi [1] observed that implementing RRS was associated with a reduction in cardiac arrests of 1.64 fewer incidents per unit time, alongside a drop of 9.61 unplanned ICU admissions per unit time with statistical significance. AlZaher et al. [2] found that embedding the Bedside PEWS in daily clinical practice facilitated earlier escalation of care, resulting in reduced mortality rate after unplanned ICU admission from 1.3 per 100 bed days to 0.4 per 100 bed days.

However, systematic reviews by Trubey et al. [7] and Chapman et al. [10] and a meta-analysis done by Maharaj et al. [8] have highlighted variability in RRS effectiveness across hospitals and countries due to differences in staff training, adherence to protocol, and patient demographics. Additionally, over 30 different PEWS variants are in use globally, each with unique scoring methods and thresholds, complicating direct comparisons and limiting standardization [7,19,20]. For these reasons, experts have argued that PEWS should not be viewed as a standalone solution but rather as part of a broader socio-technical system that includes staff engagement, strong governance, and continuous quality improvement [7].

Another persistent limitation beyond these systemic challenges is the trade-off between sensitivity and specificity: highly sensitive PEWS can cause alarm fatigue through excess false positives, whereas systems tuned for specificity may fail to identify early signs of deterioration [7]. Other issues include inconsistent activation of RRS protocols by medical staff even when PEWS thresholds are crossed and concerns that over-reliance on numeric scores may erode clinical judgment [10,11]. Moreover, the dependence of PEWS on manually calculated scores based on static criteria may delay recognition of subtle physiological changes. These inherent constraints have prompted researchers and clinicians to seek more adaptive, data-driven solutions that move beyond intermittent, static score-based assessment and operate in real time (Figure 1).

Figure 1.

Evolution of Pediatric Early Warning System (PEWS): intermittent score-based workflows versus continuous artificial intelligence (AI)-driven monitoring. Schematic comparison of intermittent, score-based pediatric early warning workflows and continuous AI-driven monitoring. Traditional systems rely on periodic manual assessment and static scoring, which may delay recognition of clinical deterioration. In contrast, continuous biosignal acquisition with real-time AI analytics enables dynamic risk assessment, actionable alerts, and earlier clinical intervention.

INNOVATIONS IN ARTIFICAL INTELLIGENCE-BASED EARLY WARNING SYSTEMS

To overcome the limitations of static, rule-based PEWS, recent research has shifted toward artificial intelligence (AI) and machine learning approaches capable of leveraging high-dimensional electronic medical record data. An AI-based early warning system generally refers to a predictive model that continuously analyzes large volumes of clinical data such as vital signs and laboratory results to identify patterns associated with impending deterioration. Unlike traditional scoring systems that rely on fixed thresholds and limited parameters, these models can integrate numerous variables, capture complex and non-linear relationships, and update risk assessments dynamically as new data become available.

Foote et al. [21] developed and internally validated a machine learning model for predicting pediatric deterioration events, achieving an area under the receiver operating curve of 0.85, outperforming traditional PEWS (area under the receiver operating curve [AUROC], 0.69); this model enabled earlier detection of deterioration by up to a median of 8 hours. In Korea, Jeon et al. [22] proposed a deep learning model for predicting critical events in general wards, reporting an AUROC of 0.99 and an area under the precision-recall curve (AUPRC) of 0.9, demonstrating both excellent discrimination and balance between sensitivity and specificity. However, these models have so far only undergone internal validation, with more evidence of their generalizability required.

Park et al. [23] reported an AUROC of 0.92 for cardiopulmonary arrest and 0.91 for unplanned ICU transfers using a deep learning-based pediatric early warning system. Their model also reduced false alarms by up to 82%, improving precision and mitigating alarm fatigue compared to traditional PEWS. An effort to externally validate this model was undertaken by Shin et al. [6], who conducted a multicenter study and demonstrated that the model had an improved AUROC of 0.8 compared to 0.76 for traditional PEWS. While external validation showed some performance drop compared to internal validation, this model still outperformed traditional PEWS, highlighting its potential practical value. This model has since been deployed in clinical practice and is commercially available (DeepCARS; VUNO).

Adult hospitals have a longer and more established history of developing AI-driven RRS, providing an important roadmap for pediatric translation. Over the past decade, several AI-based deterioration prediction models have been created and retrospectively validated at scale in adult inpatient settings [24,25]. Beyond retrospective validation, some adult systems have already progressed to prospective evaluation. For example, Cho et al. [26] prospectively implemented the Deep Learning-based Cardiac Arrest Risk Management System across four hospitals in Korea and reported consistently superior performance compared with conventional early warning systems. The adult studies illustrate both the feasibility and clinical impact of AI-driven RRS, and adult systems can provide an important foundation for developing AI-based models tailored to pediatric care.

Still, pediatric studies lack external validation, raising concerns about generalizability to diverse pediatric populations and clinical environments [8,11]. Moreover, while pediatric AI models often report high AUROC values, corresponding AUPRC metrics, which better reflect performance in detecting rare deterioration events, are sometimes low or omitted altogether. Although AI models show promise due to their use of real-world data, the absence of validated pediatric models using appropriate performance metrics limits clarity about their true impact on false alarms and missed detections. The ability to integrate AI-driven RRS seamlessly into existing pediatric clinical workflows is also essential, especially given that pediatric patient characteristics and ward workflows differ substantially from those in adult care settings.

In addition to these validation and implementation challenges, many AI-based prediction models continue to rely on intermittently updated clinical data feeds rather than continuously streamed physiological data, which means predictions may not capture real-time changes in patient status. Likewise, most models depend on aggregated variables rather than raw, high-resolution waveforms, potentially missing subtle patterns that may precede overt deterioration. These factors point to the need for next-generation innovations that integrate continuous, high-fidelity biosignal monitoring with AI analytics, an approach that has the potential to provide earlier and more precise alerts.

REAL-TIME BIOSIGNAL-BASED AI

Building on this need, recent advancements in real-time biosignal analysis powered by AI are opening new frontiers for the early and accurate detection of clinical deterioration in pediatric patients. Unlike conventional systems that rely primarily on periodically recorded vital signs or observational assessments, continuous biosignal monitoring captures dynamic physiological fluctuations in real time, providing richer and temporally granular data [27,28]. This continuous data stream allows predictive models to detect subtle or transient patterns that may remain undetected in periodically measured data, ultimately improving early warning performance [28].

Continuous, high-fidelity biosignal analytics have traditionally been developed and validated within ICU settings because invasive lines are frequently placed in ICU patients and high sampling-rate monitoring data with detailed physiologic annotations are typically available. The ICU environment has served as an ideal testbed for advancing real-time signal processing and AI methods, and algorithms developed using ICU-acquired biosignals are transferable to the general ward because the underlying signal sources studied in the ICU can be acquired through widely available bedside monitors. Thus, while the methodological roots of biosignal AI lie in the ICU, its practical applications can be expanded toward ward-based early warning systems, helping bridge the gap between intermittent vital sign assessment and real-time physiologic monitoring.

Several studies have illustrated the clinical potential of biosignal AI (Table 3) [29-38]. For example, Duncan et al. [31] developed the Real-time Adaptive Predictive Indicator of Deterioration (RAPID) Index, an early warning system that continuously monitors heart rate and respiratory rate. While the traditional PEWS showed better overall predictive discrimination, the RAPID Index identified a higher proportion of significant clinical deterioration events, with a sensitivity of 97.2% compared to the 86.1% sensitivity of PEWS [31]. Similarly, Lee et al. [35] applied deep learning to continuous heart rate variability data in the ICU, achieving an AUROC of 0.89 for predicting impending cardiac arrest 12 hours in advance.

Table 3.

Recent application of biosignal-based AI models for critical care

Transfer learning, a machine learning technique that leverages pre-trained models on large biomedical signal datasets, has further improved model performance. According to Jafari et al. [39], transfer learning has shown notable success in domains such as electrocardiography, electroencephalography, and photoplethysmography, with consistently high classification accuracy reported. This technique enables models to capture complex temporal and morphological features from rich source datasets, which can then be adapted to target pediatric populations, even when labeled data are limited. Importantly, Jafari et al. [39] emphasized that transfer learning not only improves predictive accuracy, but also supports better generalization across diverse patient cohorts, device types, and clinical environments.

Beyond improving predictive accuracy, biosignal AI holds promise for non-invasive monitoring of traditionally invasive parameters. For instance, deep learning models applied to photoplethysmogram signals have been used to estimate arterial blood pressure non-invasively with high accuracy (mean error of –0.006) [29]. Additionally, algorithms trained on capnography-derived end-tidal carbon dioxide data have been shown to estimate arterial carbon dioxide levels, with an R2 coefficient of determination of 0.85 reported using a multilayer perceptron model [32]. Such advances could reduce procedural risks and discomfort for pediatric patients, enabling continuous, high-fidelity monitoring even outside intensive care settings.

However, despite these promising developments, challenges remain. Similar to other AI models based on electronic medical record data, biosignal AI models lack external validation across diverse pediatric populations, require integration with existing hospital workflows, and raise data privacy concerns as well as user acceptability issues such as limited model interpretability [28,39]. Overcoming these barriers is essential to shape the next phase of development.

FUTURE DIRECTIONS

Among these barriers, limited interpretability stands out as one of the critical challenges for safe clinical adoption, as clinicians need to understand how model predictions are generated in order to act on them with confidence. Without insight into the factors driving a model’s output, there is a risk of relying on opaque algorithms or, conversely, reluctance to act on potentially meaningful alerts. To address this issue, explainable AI could bridge the gap between complex model output and reliable clinical decision-making. As Maurer et al. [40] emphasized, incorporating explainable models into biosignal-based clinical decision support can help clinicians understand why a model makes certain predictions, thereby fostering trust and appropriate action. For example, visualization techniques such as heatmaps and saliency maps that highlight which parts of a biosignal contribute most to a prediction can make AI models more transparent.

In terms of integration into real-world workflows, large language models (LLMs) are being explored as tools to improve communication between AI systems and clinical users. LLMs have the potential to enhance workflow efficiency by translating complex physiological data into structured, clinician-friendly reports. Liu et al. [41] demonstrated that their system, BioSignal Copilot, which integrates LLMs for real-time report generation, can reduce documentation time and improve clinician’s understanding of biosignal-derived insights.

Additionally, the miniaturization of sensors and advances in mobile computing technology are starting to allow wearable and mobile biosignal applications, extending early warning capabilities even beyond hospital walls. Future devices may not only monitor biosignals but interact with AI models on-device, allowing for real-time alerts in home or ambulatory settings. This concept of ubiquitous, intelligent biosignal monitoring has vast implications for reducing preventable deteriorations, especially in resource-limited environments [42].

Lastly, several practical barriers continue to limit real-world implementation. Heterogeneity among electronic monitoring device vendors and the use of proprietary, undisclosed data acquisition protocols create challenges for data interoperability and reliable real-time data flow. Even with improved AI algorithms, alert fatigue remains a major concern, as poorly calibrated or poorly integrated AI systems can unintentionally disrupt clinical workflows. Staffing limitations, specifically the scarcity of clinicians with sufficient understanding of AI-driven analytics, may further hinder widespread and effective utilization. Medico-legal responsibility also remains opaque in many jurisdictions, with unresolved questions about liability when automated predictions influence clinical decisions. Accordingly, the integration of biosignal AI into regulatory, ethical, and clinical governance frameworks is essential, and future systems must ensure robust data security, patient consent, equitable access, and clear clinical accountability. Addressing these operational, technical, and legal barriers will require interdisciplinary collaboration among engineers, clinicians, and regulatory bodies to ensure safe, fair, and impactful deployment of these technologies [43].

CONCLUSIONS

Early warning systems in pediatric care have evolved from static scoring tools to dynamic, AI-driven biosignal models. While PEWS and RRS provided the foundation for structured detection and response, emerging technologies offer the potential for earlier, more accurate, and more actionable predictions. Challenges such as generalizability, alarm fatigue, and workflow integration remain, but advances in explainable AI, LLM-based clinical interfaces, and wearable biosignal monitoring devices hold promise for the next generation of pediatric early warning systems. With careful implementation and governance, these innovations may transform pediatric practice and contribute to saving more lives worldwide.

KEY MESSAGES

Article information Continued