Real-time prediction of cardiorespiratory deterioration during paediatric critical care transport using interpretable machine learning

Real-time prediction of cardiorespiratory deterioration during paediatric critical care transport using interpretable machine learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Real-time prediction of cardiorespiratory deterioration during paediatric critical care transport using interpretable machine learning Milan Kapur, Kezhi Li, Alexander Brown, Zhiqiang Huo, John Booth, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7319827/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Interhospital transport of critically ill children carries inherent risks, including unexpected respiratory and cardiovascular deterioration. Early warning of impending deterioration may allow physicians to avert a more serious issue. We developed and evaluated lightweight, explainable machine learning models to forecast adverse physiological events up to 15 minutes in advance using continuously streamed vital signs and clinical data. Models were trained and evaluated on 1,519 transports conducted by a specialist paediatric critical care team in London (2016–2021). Transformer-based models incorporating vital sign time-series and vector-embedded diagnoses outperformed simpler models, achieving AUROC scores of 0.851 for respiratory and 0.792 for cardiovascular deterioration. Model interpretability was provided using Integrated Gradients, revealing alignment with clinical reasoning. Designed for deployment on edge devices, these models offer real-time, interpretable risk predictions in resource-limited transport settings. These results demonstrate that real-time, explainable machine learning models can accurately predict deterioration during interhospital paediatric transport using routinely collected data, supporting their potential role in enhancing early clinical intervention. Health sciences/Cardiology Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Health care Physical sciences/Mathematics and computing Health sciences/Medical research Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction In the United Kingdom, around 20,000 children require admission to Paediatric Intensive Care Units (PICUs) annually. 1 Many initially present to general hospitals that lack on-site PICU services, making urgent interhospital transfer essential for them to receive necessary intensive care. In 2023, more than 4,000 such transfers were conducted. 1 Despite being conducted by specialised Paediatric Critical Care Transport (PCCT) teams of doctors, nurses, and paramedics, interhospital transfers remain high-risk. 2 Transport-related physiological stress elevates the risk of en-route clinical deterioration, with adverse events reported in 12.3% of over 8,000 interhospital transport episodes in one large study. 3 A recent systematic review categorised transport associated adverse events into four main groups: respiratory (e.g., hypoxaemia), cardiovascular (e.g., hypotension, tachycardia/bradycardia), equipment-related (e.g., monitor failure), and other (e.g., medication error). 4 These events, even when promptly identified and managed by PCCT teams, can lead to significant morbidity and even mortality, underscoring the critical need for advancements that move beyond reactive care to proactive intervention. During transport, vital signs are continuously monitored to detect deterioration. However, PCCT services largely rely on manual interpretation against static, age-based reference ranges or clinical judgment. This approach has limitations, as "normal" vital sign ranges for critically unwell children vary widely with age, diagnosis, and illness severity. However, this approach requires well-characterized normal vital sign ranges for the patient. Whilst these ranges exist for stable paediatric patients, there are no “normal” ranges for critically unwell children during transport; the expected values vary significantly depending on age, diagnosis, severity of illness and ongoing interventions. 5 – 8 This often leads to either over-alerting (false alarms due to transient fluctuations in already abnormal physiology) or under-detection of subtle but significant changes. To address this, our prior work explored dynamic thresholding approaches that better account for inter-individual variability, allowing for personalised and adaptive detection of patient deterioration. 9 While detecting deterioration in real time is critical for timely intervention, predicting future deterioration enables clinicians to act pre-emptively before a patient’s condition worsens. Recent machine learning (ML) advancements allow sophisticated predictive tools to leverage large patient datasets, analysing continuous physiological data to forecast acute deterioration. 10 – 13 Despite these advancements, to the best of our knowledge, no models exist capable of predicting real-time risk of acute deterioration during paediatric critical care transport. To address this gap, we introduce two novel, lightweight ML models capable of predicting respiratory and cardiovascular deterioration up to 15 minutes in advance, in real-time. Trained on a large dataset of 1200 transport episodes with a lightweright design suitable for edge computation, these models have the potential to provide timely clinical decision support during transport, which may in turn enable earlier intervention and contribute to improved patient outcomes. Results Cohort Description : his retrospective study analysed continuously monitored vital signs and linked electronic health record (EHR) data from critically ill children transported by the Children’s Acute Transport Service (CATS) in North London, UK, between January 2016 and May 2021. Of the 6,471 transport episodes during this period, data capture was significantly limited by operational constraints: only two recording devices were available for three transport teams, making full coverage unfeasible. Additional data loss occurred when devices were forgotten or due to technical issues maintaining stable monitor-device-server connections. Consequently, 1,781 episodes had sufficiently complete monitoring and EHR data for initial consideration. After further excluding patients over 18 years old and those with < 30 minutes of monitoring, a final cohort of 1,519 episodes (23.5% of the total 6,471) met inclusion criteria (Fig. 1 ). Table 1 shows the included cohort was demographically and clinically representative of the overall cohort. Table 2 summarizes the availability of key variables; not all parameters were recorded for every patient due to clinical variability (e.g., EtCO₂ only for invasively ventilated patients). The median duration of a transport episode was 113 minutes (IQR: 75–160) (Supplementary Fig. 1). For our predictive analysis, each transport was segmented into 15-minute time windows, within which we aimed to predict the occurrence of at least one adverse event. Adverse events were uncommon within these windows (Table 3 ), suggesting patients were generally stable during transport. Table 1 Breakdown of characteristics of all transport episodes and included episodes Characteristics All Transport Episodes (n = 6471) Included Episodes (n = 1519) Age Group Distribution ≤ 1 month 2219 (34.3%) 568 (37.4%) 1– ≤12 months 1330 (20.6%) 296 (19.5%) 1– ≤4 years 1198 (18.5%) 264 (17.4%) 4– ≤11 years 1017 (15.7%) 239 (15.7%) 11– ≤18 years 678 (10.5%) 151 (9.9%) > 18 years 27 (0.4%) 0 (0%) Gender Distribution Male 3603 (55.7%) 825 (54.3%) Female 2861 (44.2%) 692 (45.6%) Diagnosis Group Distribution Respiratory 2233 (34.5%) 498 (32.8%) Cardiovascular 1385 (21.4%) 364 (24.0%) Neurological 921 (14.2%) 227 (14.9%) Infection 888 (13.7%) 202 (13.3%) Gastrointestinal 368 (5.7%) 89 (5.9%) Metabolic 161 (2.5%) 43 (2.8%) Trauma 81 (1.3%) 21 (1.4%) Other 434 (6.7%) 75 (4.9%) PIM3 Risk of Mortality ≤ 1% 502 (7.8%) 92 (6.1%) 1– ≤3% 2345 (36.2%) 510 (33.6%) 3– ≤5% 2060 (31.8%) 518 (34.1%) 5– ≤10% 1038 (16.0%) 264 (17.4%) 10– ≤15% 206 (3.2%) 59 (3.9%) 15– ≤30% 182 (2.8%) 45 (3.0%) > 30% 126 (1.9%) 30 (2.0%) Respiratory Support Self-ventilating (Room Air) 1292 (20.0%) 257 (16.9%) Self-ventilating (supplemental O₂) 141 (2.2%) 27 (1.8%) Self-ventilating (HFNC) 250 (3.9%) 62 (4.1%) Self-ventilating (CPAP) 278 (4.3%) 55 (3.6%) Self-ventilating (BIPAP) 54 (0.8%) 12 (0.8%) Invasive ventilation (ETT) 4214 (65.1%) 1073 (70.6%) Invasive ventilation (Tracheostomy) 93 (1.4%) 19 (1.3%) Invasive ventilation (Other airway) 10 (0.2%) 3 (0.2%) Cardiovascular Support Adrenaline 815 (12.6%) 199 (13.1%) Dobutamine 31 (0.5%) 9 (0.6%) Dopamine 580 (9.0%) 137 (9.0%) Milrinone 58 (0.9%) 12 (0.8%) Noradrenaline 507 (7.8%) 131 (8.6%) Any agent 2011 (31.1%) 488 (32.1%) Overall Transport Time, minutes ≤ 60 16 (0.2%) 0 (0.0%) 60–120 598 (9.2%) 105 (6.9%) 120–180 1523 (23.5%) 345 (22.7%) 180–240 2067 (31.9%) 545 (35.9%) 240–300 1282 (19.8%) 322 (21.2%) 300–360 556 (8.6%) 142 (9.3%) > 360 261 (4.0%) 55 (3.6%) Demographic, clinical, and transport characteristics of the study population compared to all transported children during the study period. Respiratory support and cardiovascluar support refer to support received during transport. HFNC = High-Flow Nasal Cannula; CPAP = Continuous Positive Airway Pressure; BIPAP = Bilevel Positive Airway Pressure; ETT = Endotracheal Tube. Table 2 List of collected Electronic Health Care (EHR) and vital sign data in the cohort Category EHR and vital signs data Data type Range Missing (%) Patient demographics Age (years), Weight (kg) Sex Ethnicity Numerical Numerical Categorical Categorical 0–18 1–90 NA NA 0.00% 0.26% 0.07% 51.68% Transport details Referring hospital Destination hospital Time of transfer Destination care area (PICU, NICU, HDU Ward, other) Categorical Categorical Numerical Categorical NA NA NA NA 0.00% 0.26% 0.00% 0.20% Medical Diagnosis Primary diagnosis, Paediatric index of mortality 3 (PIM3), Existing medical conditions (respiratory, cardiac, renal, genetic, metabolic/endocrine, haematological/oncological, other) Categorical Numerical Categorical NA 0–1 NA 0.00% 0.26% 0.00% Interventions by local team prior to transport Intubation (primary/re-intubation/ETT repositioning), mechanical ventilation (invasive, non-invasive and HFNC), suctioning, chest drain insertion, vascular access (peripheral, central, arterial, intraosseous), vasoactive support (inotropes/vasopressors, prostaglandin), blood product transfusion, urinary catheterisation, nasogastric/orogastric tube placement, imaging (CT scan), C-spine immobilisation, osmotherapy, CPR/defibrillation, ECMO, and ICP monitoring Binary NA 0.79% Intra-transport respiratory support commenced prior to transport Self-ventilating: Room air, Supplemental O₂, HFNC, CPAP, BiPAP; Invasive ventilation: ETT, Tracheostomy, Other airway Categorical NA 0.72% Intra-transport cardiovascular support commenced prior to transport Adrenaline, noradrenaline, dobutamine, dopamine, milrinone, prostaglandin, inhaled nitric oxide Categorical NA 0.00% Vital Signs SpO2 Heart rate (3-lead ECG) End tidal CO2 (minimum value in 1-second period) End tidal CO2 (maximum value in 1-second period) Airway derived respiratory rate Impedance pneumography derived respiratory rate Mean systolic blood pressure (non-invasive) Mean arterial blood pressure (non-invasive) Mean diastolic blood pressure (non-invasive) Mean systolic blood pressure (invasive) Mean arterial blood pressure (invasive) Mean diastolic blood pressure (invasive) Temperature (oesophageal) Temperature (skin) Temperature (core) Temperature (unspecified) Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical Numerical 0-100 0-301 0-19.1 0.3–20.1 0-164 0-171 0-264 0-242 0-247 0-361 0-361 0-340 2.9–40.8 6.6–43.4 8.2–44.6 5.8–43.1 0.07% 0.59% 29.00% 27.32% 27.39% 4.08% 5.99% 5.99% 5.99% 63.98% 63.98% 63.98% 78.74% 72.55% 85.71% 64.05% Missing rates were calculated on the 1519 patients included in the study and captures transport episodes where there were zero recorded values of a given parameter. PICU = Paediatric Intensive Care Unit, NICU = Neonatal Intensive Care Unit, HDU = High Dependency Unit, ETT = Endotracheal tube, HFNC = High Flow Nasal Canula, CPR = Cardiopulmonary Resuscitation, ECMO = Extra-Corporeal Membrane Oxygenation, ICP = Intracranial Pressure, CPAP = Continuous Positive Airway Pressure, BiPAP = Bi-level Positive Airway Pressure. Table 3 Frequency of Adverse Events in Time Windows. Metric Train Tune Test Total Windows 8109 1022 1031 Respiratory Deterioration (%) 2.48% 2.25% 3.1% Cardiac Deterioration (%) 4.06% 5.19% 4.36% Both Respiratory and Cardiac Deterioration (%) 0.57% 0.59% 0.87% No Deterioration (%) 92.9% 91.98% 91.66% This table summarizes the total number of time windows analysed in the train, tune, and test sets, along with the percentages of windows exhibiting respiratory deterioration, cardiac deterioration, both respiratory and cardiac deterioration, and no deterioration. Notably, adverse events are rare, with windows showing any type of deterioration collectively accounting for less than 10% of the total. Time windows from a given patient belong exclusively to one of the three groups. Model Performance : Performance was evaluated on the holdout test set. Models were provided with the 15-minutes of high-resolution vital signs data preceding the start of the 15-minute prediction window, alongside averaged historical context from up to 120 minutes prior and liked EHR data. Six models were developed for each prediction task (respiratory and cardiovascular deterioration), progressively increasing in feature richness and computational complexity (Supplementary Table 1). Models were then assessed on their ability to predict the occurrence of at least one adverse respiratory or cardiovascular event in the subsequent 15-minute period. Performance was evaluated on a random label-stratified holdout test set using standard metrics: Area Under the Receiver Operating Characteristic Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), Sensitivity (Recall), Specificity, Positive Predictive Value (PPV, Precision), Negative Predictive Value (NPV), Balanced Accuracy, and F1-score. A fixed decision threshold was chosen on the tuning set to cap the false positive rate at 20% (i.e., ≥ 80% specificity). ROC and PR curves are shown in Fig. 2 . Respiratory Events : The baseline-only model, relying solely on demographic and pre-transport features, achieved an AUROC of 0.669 (95% CI: 0.578–0.757) and an AUPRC of 0.099 (95% CI: 0.064–0.145), indicating modest predictive capability. Enhancing the model by incorporating time-series vital signs through a simple feed-forward network (Combined FF) improved performance to an AUROC of 0.800 (95% CI: 0.746–0.850) and an AUPRC of 0.139 (95% CI: 0.090–0.215). Further gains were observed when transformer-based architectures were applied. A transformer-based model, using only the time-series vital signs (“Vitals-Only Transformer”) achieved an AUROC of 0.846 (95% CI: 0.799–0.891) and an AUPRC of 0.201 (95% CI: 0.124–0.300), with a sensitivity of 0.707 (95% CI: 0.561–0.829) and balanced accuracy of 0.755 (95% CI: 0.684–0.823). We also tested combining time-series vital signs with differing combinations of diagnosis and baseline variables. Among the combined transformer models, the approach using one-hot encoded diagnosis with a reduced baseline feature set reached an AUROC of 0.850 (95% CI: 0.805–0.892), an AUPRC of 0.173 (95% CI: 0.119–0.247), balanced accuracy of 0.764 (95% CI: 0.693–0.832), and an F1-score of 0.235 (95% CI: 0.188–0.282). The variant using vector-embedded diagnosis (with the same reduced baseline feature set) showed a comparable AUROC of 0.851 (95% CI: 0.805–0.894) but achieved a higher AUPRC of 0.200 (95% CI: 0.127–0.298) at the expense of a slightly lower F1-score (0.216, 95% CI: 0.174–0.257). Notably, including the full set of baseline features in the transformer (Vector Diagnosis, Full Baseline) did not improve performance, yielding an AUROC of 0.841 (95% CI: 0.789–0.889) and balanced accuracy of 0.754 (95% CI: 0.685–0.823). Cardiovascular Events : For cardiovascular deterioration, the baseline-only model again demonstrated limited performance, with an AUROC of 0.652 (95% CI: 0.594–0.709) and an AUPRC of 0.076 (95% CI: 0.064–0.091). The Combined FF model, which incorporated time-series vital signs, improved these metrics to an AUROC of 0.763 (95% CI: 0.697–0.823) and an AUPRC of 0.149 (95% CI: 0.110–0.200). The Vitals-Only Transformer produced similar results (AUROC 0.759, 95% CI: 0.704–0.813; AUPRC 0.149, 95% CI: 0.102–0.213). Among the transformer-based approaches, the Combined Transformer using one-hot encoded diagnosis with a reduced baseline feature set achieved an AUROC of 0.772 (95% CI: 0.720–0.823) and an AUPRC of 0.144 (95% CI: 0.103–0.204). The best performance for cardiovascular prediction was obtained with the Combined Transformer employing vector-embedded diagnosis (with the same reduced baseline feature set), which achieved an AUROC of 0.792 (95% CI: 0.739–0.842), an AUPRC of 0.183 (95% CI: 0.122–0.261), a balanced accuracy of 0.677 (95% CI: 0.611–0.747), and an F1-score of 0.242 (95% CI: 0.182–0.308). As seen in the respiratory models, incorporating the full set of baseline features (Vector Diagnosis, Full Baseline) did not offer additional benefits, with this configuration attaining an AUROC of 0.786 (95% CI: 0.735–0.833) and an AUPRC of 0.135 (95%CI: 0.105–0.173). Table 4 Results of evaluating the model on the holdout test set Model AUROC AUPRC Sensitivity Specificity PPV NPV Balanced Accuracy F1-Score Respiratory Models Baseline-Only FF 0.669 (0.578–0.757) 0.099 (0.064–0.145) 0.413 (0.268–0.561) 0.892 (0.873–0.911) 0.137 (0.090–0.185) 0.973 (0.967–0.980) 0.653 (0.579–0.729) 0.206 (0.136–0.277) Combined FF 0.800 (0.746–0.850) 0.139 (0.090–0.215) 0.610 (0.463–0.756) 0.803 (0.778–0.828) 0.114 (0.086–0.142) 0.980 (0.973–0.988) 0.706 (0.629–0.780) 0.192 (0.145–0.237) Vitals-Only Transformer 0.846 (0.799–0.891) 0.201 (0.124-0.300) 0.707 (0.561–0.829) 0.804 (0.779–0.828) 0.130 (0.104–0.157) 0.985 (0.978–0.991) 0.755 (0.684–0.823) 0.220 (0.177–0.263) Combined Transformer (One-Hot Diagnosis, Reduced Baseline) 0.850 (0.805–0.892) 0.173 (0.119–0.247) 0.707 (0.561–0.829) 0.821 (0.797–0.844) 0.141 (0.112–0.170) 0.985 (0.979–0.992) 0.764 (0.693–0.832) 0.235 (0.188–0.282) Combined Transformer (Vector Diagnosis, Reduced Baseline) 0.851 (0.805–0.894) 0.200 (0.127–0.298) 0.730 (0.585–0.854) 0.792 (0.766–0.817) 0.127 (0.102–0.152) 0.986 (0.979–0.993) 0.761 (0.690–0.828) 0.216 (0.174–0.257) Combined Transformer (Vector Diagnosis, Full Baseline) 0.841 (0.789–0.889) 0.177 (0.121–0.251) 0.70 7 (0.561–0.830) 0.802 (0.777–0.826) 0.129 (0.104–0.155) 0.985 (0.978–0.992) 0.754 (0.685–0.823) 0.218 (0.176–0.261) Cardiovascular Models Baseline-Only FF 0.652 (0.594–0.709) 0.076 (0.064–0.091) 0.037 (0.000-0.093) 0.892 (0.871–0.911) 0.019 (0.000-0.047) 0.944 (0.941–0.947) 0.464 (0.441–0.494) 0.025 (0.000-0.062) Combined FF 0.763 (0.697–0.823) 0.149 (0.110–0.200) 0.481 (0.352–0.611) 0.833 (0.810–0.856) 0.138 (0.102–0.175) 0.967 (0.959–0.975) 0.657 (0.591–0.727) 0.214 (0.158–0.271) Vitals-Only Transformer 0.759 (0.704–0.813) 0.149 (0.102–0.213) 0.500 (0.370–0.630) 0.822 (0.798–0.845) 0.135 (0.101–0.170) 0.968 (0.959–0.976) 0.661 (0.595–0.728) 0.212 (0.159–0.266) Combined Transformer (One-Hot Diagnosis, Reduced Baseline) 0.772 (0.720–0.823) 0.144 (0.103–0.204) 0.463 (0.333–0.593) 0.845 (0.822–0.866) 0.142 (0.103–0.181) 0.966 (0.958–0.974) 0.654 (0.586–0.721) 0.217 (0.157–0.277) Combined Transformer (Vector Diagnosis, Reduced Baseline) 0.792 (0.739–0.842) 0.183 (0.122–0.261) 0.500 (0.370–0.630) 0.855 (0.832–0.876) 0.160 (0.120–0.204) 0.969 (0.961–0.977) 0.677 (0.611–0.747) 0.242 (0.182–0.308) Combined Transformer (Vector Diagnosis, Full Baseline) 0.786 (0.735–0.833) 0.135 (0.105–0.173) 0.501 (0.370–0.630) 0.831 (0.808–0.855) 0.141 (0.105–0.178) 0.968 (0.960–0.976) 0.666 (0.599–0.734) 0.220 (0.164–0.275) Table 4 reports the results of evaluating the trained models for each of the two prediction tasks on the holdout test set. The decision threshold was determined on the tuning set by selecting the value that yielded a maximum false positive rate (FPR) of 20% (i.e., a minimum specificity of 80%). The values in brackets represent the 95% confidence interval calculated by performing stratified bootstrapping with replacement on the test set. "FF" refers to a feed-forward neural network. AUROC - Area Under the Receiver Operating Characteristic curve, AUPRC - Area Under the Precision-Recall Curve, PPV - Positive Predictive Value, NPV - Negative Predictive Value. Model Explainability We employed the Integrated Gradients method to identify which input features most influenced model predictions. This analysis focused on the best-performing models (Combined Transformer (Vector Diagnosis, Reduced Baseline)) for both respiratory and cardiovascular deterioration, based on AUROC. Feature attributions are shown in Fig. 3 . In the respiratory model, the most influential features were non-invasive blood pressure, previous cardiovascular deteriorations, end-tidal CO₂ (EtCO₂), diagnosis, and previous respiratory deteriorations, followed by blood oxygen saturation (SpO₂) and respiratory rate. Positive predictions relied more on SpO₂, while negative predictions emphasised diagnosis and prior respiratory deteriorations. For the cardiovascular model, EtCO₂, non-invasive and arterial blood pressure, heart rate, and prior deteriorations were most important, with arterial pressure weighted more in positive predictions and diagnosis in negative ones. Overall, time-series vital sign features were more important than baseline demographic features, though both informed the final predictions. The same feature attribution method can be applied dynamically to individual prediction windows, offering real-time insight into how each input influences the model’s output. For each 15-minute window, the model generates a probability (0–1) of an adverse event occurring in the following 15 minutes. In Fig. 4 a, a high predicted probability (0.91) for respiratory deterioration was followed by an actual event; key contributing features included arterial and non-invasive blood pressure, prior cardiovascular events, EtCO₂, respiratory rate, SpO₂, and prior respiratory events. In contrast, Fig. 4 b shows a low predicted probability (0.18) for cardiovascular deterioration, with no event occurring. The most influential features in this case included blood pressure, EtCO₂, heart rate, and temperature. Discussion To our knowledge, this is the first study to use continuously collected vital signs and linked EHR data during interhospital transport of critically ill children to predict real-time risk of deterioration. We present two lightweight, explainable machine learning models that integrate live-streamed physiology with baseline clinical data to forecast respiratory and cardiovascular events up to 15 minutes in advance, potentially enabling earlier intervention and preventing decline. Previous studies have largely focused on predicting mortality during transport or deterioration in stable, non-critically ill paediatric inpatients. 11 , 13 – 16 While several of these models have demonstrated strong performance in predicting ward-to-PICU transfers or ICU mortality, none have been specifically developed to predict imminent deterioration in critically ill children during interhospital transport Our work addresses this gap, introducing a real-time risk prediction tool for one of the most high-risk settings in paediatric care. The models were developed using a diverse dataset of over 1,500 interhospital transports conducted by the CATS team in central London. We adopted a systematic approach, starting with models based solely on demographic and pre-transport features, then incorporating high-resolution time-series data. Initial models used simple feed-forward networks, progressing to transformer-based architectures for more advanced temporal modelling. We also evaluated multiple strategies for diagnosis representation, comparing one-hot encoding against vector embeddings derived from a pretrained clinical language model. 17 This iterative, constructive approach allowed us to quantify how predictive performance scaled with increasing model complexity and feature richness, while identifying the minimal configuration needed to maintain high accuracy. High-frequency vital-sign data, particularly when processed using transformer-based models capable of capturing complex temporal dynamics significantly boosted accuracy (Table 4 ). Our best-performing models for both tasks used the Combined Transformer (Vector Diagnosis, Reduced Baseline) architecture. On the holdout test set, the respiratory deterioration model achieved an AUROC of 0.851 and AUPRC of 0.200, with a sensitivity of 0.730 and specificity of 0.792. The cardiovascular model yielded an AUROC of 0.792 and AUPRC of 0.183, with a sensitivity of 0.500 and specificity of 0.855. These results are particularly notable given the low baseline incidence of events in the test set: only 3.1% of time windows included respiratory deterioration, 4.4% included cardiovascular deterioration, and just 0.87% involved both. This extreme class imbalance underscores the models’ ability to extract clinically salient signals from noisy transport data. Beyond overall performance, we used Integrated Gradients to understand how the best-performing models made predictions, assessing feature importance at both the cohort and individual level. 18 In the respiratory model, the top contributors were non-invasive blood pressure, prior cardiovascular events, and EtCO₂, followed by diagnosis, prior respiratory events, and SpO₂, likely reflecting the interdependent nature of respiratory and cardiovascular physiology, where cardiovascular instability is often associated with respiratory instability. The cardiovascular model simliaryly emphasised EtCO₂, non-invasive and arterial blood pressure, heart rate and previous episodes of respiratory or cardiovascular instability. The high weighting of EtCO₂ may again reflect the interdependent nature of respiratory and cardiovascular physiology, where CO2 is not only a marker of ventilation but also perfusion. 19 Furthermore, the presence of invasive blood pressure, available in ~ 36% of transport episodes, was associated with positive predictions, likely reflecting the fact that patients requiring arterial lines are more haemodynamically unstable and thus at higher risk of deterioration. Interestingly, diagnosis was more influential in negative predictions, suggesting the model learned that certain conditions confer lower deterioration risk. These insights highlight the models’ ability to provide interpretable outputs that align with clinical reasoning, a key factor in promoting clinician trust and potential real-world adoption . A further key strength of our models is their computational efficiency and suitability for real-time deployment in resource-constrained settings. Both are lightweight and would be capable of running on edge devices, used by transport teams. Once loaded, both models can generate predictions for a given time window in well under one second on a standard laptop CPU, supporting real-time risk assessment without requiring high-end hardware. Integrated gradients, used for interpretability, is similarly efficient and can run in real time on the same device, providing patient-specific explanations alongside risk scores. Crucially, all computations can be performed locally, without relying on continuous server access; an advantage in transport environments where connectivity may be limited. Together, these features make real-world deployment in paediatric critical care transport both practical and feasible. One limitation is the availability and continuity of monitoring data: only 23.7% (1,519 of 6,471) of transports met inclusion criteria. However, included cases closely matched the overall cohort across key demographic and clinical variables. Second, although our models achieved strong discrimination with high AUROCs, their positive predictive values (PPVs) and F1-scores were modest; largely due to the low prevalence of adverse events (4–5% of windows). As a result, even with high specificity, false positives outnumber true positives, leading to lower PPV and F1-scores. This challenge is not unique to our study and is a recognised limitation in predictive modelling for rare but critical events. Similar findings have been reported in other early warning systems, where models predicting rare but critical deterioration events (such as ICU transfers, in-hospital cardiac arrests, or ward-based decompensation) achieved high AUROCs (> 0.85) but relatively low PPVs (typically ~ 10–15%), due to the low event prevalence. 20 – 22 Despite appearing suboptimal in absolute terms, such PPVs can still be clinically acceptable, particularly in the absence of alternative systems or when coupled with high sensitivity and fewer false alarms than traditional scoring methods. Further testing in prospective trials should include detailed examination of human factors and clinician preferences in managing the balance between sensitivity and false alarm rates. Looking ahead, external validation is essential before real-world deployment. Testing on datasets from other paediatric transport services (both nationally and internationally) will be crucial to assess generalisability across different patient populations, transport systems, monitoring technologies, and clinical practices. Multicentre validation would also enable retraining or fine-tuning on more heterogeneous data, improving robustness and broader applicability. Following this, prospective deployment studies will be needed to evaluate real-world performance, clinical decision-making and ultimately, patient outcomes. Furthermore, the general framework presented here could be adapted to in-hospital PICU applications relating to risk of clinical deterioration. Here, similar models could serve as adjunctive tools to support continuous risk stratification and enable proactive intervention in children already receiving intensive care. Conclusion We present the first real-time machine learning models capable of predicting acute deterioration in critically ill children during interhospital transport using routinely collected high-frequency vital sign and clinical data. These lightweight, explainable models perform well and can run on edge devices, making them practical for resource-limited settings. Our findings suggest significant potential to enhance early intervention during transport. Methods Study approval The study was approved by Great Ormond Street Hospital's (GOSH) Research and Innovation Department, under ethical approvals for use of routine de-identified healthcare and operational hospital data (Research Database, NHS REC reference 21/LO/0646). Data Sources This retrospective study analysed continuously monitored vital signs of critically ill children transported by the Children’s Acute Transport Service (CATS), a regional paediatric critical care team in North London, UK, between July 2016 and May 2021. Since 2016, CATS has used SwiftCare (Kinseed Limited, UK) to collect one data point per second vital signs, including heart rate, respiratory rate, blood pressure, oxygen saturation, and end-tidal carbon dioxide (EtCO2). Data collection starts at transfer initiation, with ambulance staff using a SwiftCare-enabled smartphone to connect wirelessly to a Philips Intellivue MP5 monitor, recording continuously until patient handover at the destination unit (Fig. 5 ). These data were linked with deidentified electronic health records (EHRs), which included demographics, transport details, diagnoses, pre-transport interventions, intra-transport support, and time-series vital signs. A complete list of recorded parameters is provided in Table 2 . While most transport episodes involved unique patients, a small proportion may reflect repeat transports. Due to record anonymisation, we could not identify repeated episodes for the same patient. However, as each transport was clinically distinct, all episodes were treated as independent records. Inclusion Criteria Episodes were included if the patient was ≤ 18 years old with at least 30 minutes of recorded vital signs. Continuous data throughout transport was not mandated, acknowledging real-world interruptions such as sensor dropouts or new measurements (e.g., EtCO2 upon intubation during transport). Data Processing Vital sign data, collected at one-second resolution (with the exception of non-invasive blood pressure, where the last recorded value was carried forward until a new measurement), underwent initial preprocessing. Physiologically implausible values were first removed from the vital sign data (heart rate 300 bpm, EtCO₂ >15 kPa, blood pressure < 5 mmHg, temperature 45°C). To account for age-related variability, all vital signs were standardised to improve comparability and enhance the reliability of model training. Respiratory rate, heart rate, and blood pressure were z-score normalised following methodology outlined in our prior work. 5 SpO₂ was scaled using (SpO₂ − 97)/6, ETCO₂ using (ETCO₂ − 5.25)/1.5, and temperature using (Temp − 36.75)/1.5. Missing values (data due real-world interruptions such as sensor dropouts) were mean-imputed (set to zero post-standardisation), and a corresponding missingness mask was generated for model input. Prior to imputation a missingness mask was created to track imputed values; this was used later as part of the feature set to train the models. Demographic features were preprocessed for model compatibility. Age, weight, and the Paediatric Index of Mortality 3 (PIM3) were scaled to a 0–1 range, with missing values imputed as − 1. 23 Sex was binary encoded, and ethnicity one-hot encoded using NHS Digital schema. 24 Referring and destination hospitals were also one-hot encoded, with units having < 10 retrievals grouped as "Other." One-hot encoding was applied to destination care area, pre-transport interventions, intra-transport respiratory and cardiovascular support, and pre-existing conditions. Team arrival time was used to label day vs night shifts. Primary diagnosis was encoded in two ways: via one-hot encoding into clinical categories (respiratory, cardiovascular, neurological, infection, gastroenterology, metabolic, trauma), and via 768-dimensional embeddings from BioClinicalBERT, pre-trained on MIMIC-III. 17 Embeddings were precomputed and cached in advance to avoid inference-time overhead. Both encoding strategies were tested during model development. Label Creation Our dataset did not include timestamped, clinician-annotated deterioration labels. To address this, we implemented an automated, data-driven approach adapting Bollinger bands to continuous vital sign data, as detailed in our previous work. 9 For each transport episode, one-minute mean-averaged data were used to compute an exponentially weighted moving average (EMA) and standard deviation (EWMSTD), generating patient-specific upper and lower thresholds that reflect each individual's evolving baseline. To prevent these dynamic bands from becoming infinitesimally narrow during periods of prolonged stability or low variability if a parameter’s EWMSTD was within 5% of its current EMA, a fixed threshold of ± 5% of the current EMA was enforced. For oxygen saturation (SpO₂), due to its inherently lower variability, a stricter fixed boundary of ± 2.5% of its current EMA was applied. Respiratory deteriorations were flagged when oxygen saturation (SpO₂) fell below either a fixed threshold of 94% or the dynamic lower bound (whichever was lower) alongside a concurrent abnormality in at least one other respiratory parameter (impedance pneumography derived respiratory rate, airway-derived respiratory rate, or EtCO₂). Cardiovascular deteriorations were identified by simultaneous deviations in heart rate and blood pressure. To reduce false positives from transient fluctuations or artefacts, events were required to persist for at least one minute and be flanked by five minutes of continuous data. These criteria were applied retrospectively to label minute-level or cumulative periods of respiratory and cardiovascular deterioration. Crucially, our automated labelling method does not carry the inherent risk of a model "rediscovering" the labelling logic. This is because the model's prediction relies exclusively on preceding 15-minute vital sign data and historical context, not on the vital sign data within the 15-minute prediction window itself. This strict temporal separation ensures that the model cannot simply identify the conditions that trigger a band-based alert at the time of the event. Instead, the model is designed to learn and predict future deterioration based on subtle physiological shifts and patterns that precede a deterioration event, rather than merely re-identifying the criteria used to define the event retrospectively. Data Split : We applied a label-stratified split, allocating 80% of episodes (1,214) to training, 10% (150) to tuning, and 10% (155) to testing. The slight imbalance between tuning and test set sizes reflects the stratification process. Episodes were first grouped into four categories based on deterioration patterns: no deterioration, only respiratory deterioration, only cardiovascular deterioration, and both respiratory and cardiovascular deterioration. Each category was then split into 80/10/10 subsets, which were subsequently combined to form the final stratified dataset. Window Creation : To facilitate model training and evaluation time series vital signs data was first discretised into 15-minute windows. This ensured that each period was used only once as a potential label window. For each window the model was tasked to use the data from the current window ( t − 14:59 - t 00:00 ), to predict whether any adverse event would occur in the next window ( t 00:01 - t + 15:00 ). Timestamped adverse events from the current window were also included as features. If data was available prior the current window (as would occur with data accumulation over the course of a long transport episode) additional context was provided as input to the model by summarizing data from t − 02:15:00 up to t − 00:15:00 . This additional data was mean-averaged over 5-minute intervals and each interval was annotated with the occurrence of respiratory or cardiovascular events, further enriching the feature set. Strict controls were applied to prevent future information leakage, and all windows from a given episode remained within the same data split (train, tune, or test). Figure 6 illustrates this windowing strategy. We chose a 15-minute prediction horizon and a 15-minute high-resolution input window due to pragmatic and clinical considerations. This future horizon offers clinically relevant lead time for proactive intervention without exceeding the model's predictive capability or providing insights that are not sufficiently actionable. The 15-minute input window effectively establishes a robust clinical baseline and captures meaningful physiological trends.12 Shorter windows (e.g., 10 minutes) might lack sufficient context, while longer ones (e.g., 20 minutes) would increase computational complexity and latency, hindering real-time deployment without substantial predictive gains. This duration also supports frequent predictions during typical transport times (median 113 minutes). Additionally, 120 minutes of averaged historical context (at 5-minute intervals) was included to offer a broader clinical trajectory while maintaining computational efficiency for edge device deployment. Model Development: We developed two sets of models to forecast respiratory and cardiovascular deteriorations within a 15-minute horizon. For each, model complexity and feature richness were gradually increased to identify the minimal configuration needed to sustain high performance while minimising user input and computational load (see Supplementary Table 1). We used a transformer architecture for time-series analysis of vital signs, leveraging its superior ability to model long-range dependencies and its computational efficiency over recurrent approaches such as recurrent neural networks or long short-term memory networks. 25 To optimize temporal encoding, transformer blocks were implemented with rotary positional embeddings. 26 Transformer blocks were implemented with a decoder-only architecture to ensure data points could only attend to preceding data, thereby guaranteeing that predictions were based solely on historical information. 25 Given the class imbalance (adverse events being rare) we applied class-weighted loss during training to penalize false negatives more heavily to maintain balanced performance across classes. We performed random search hyperparameter optimization over the training and tuning sets exploring parameters including learning rate, batch size, number of epochs, positional weighting, hidden layer dimensions, transformer heads and layers, batch normalization, dropout rates, L2 regularization, and max-norm constraints to identify the optimal training configuration. The best-performing model on the tuning set was then evaluated on the test set. This approach yielded separate models for predicting respiratory and cardiovascular events. While a combined multiclass model was considered, the non-exclusive nature of events and differing optimal architectures (e.g., hidden layer sizes, transformer depth, attention heads) made separate models more practical. Model Evaluation Model performance was evaluated on the holdout test set using AUROC, AUPRC, sensitivity (recall), specificity, PPV (precision), NPV, balanced accuracy, and F1-score. To enable consistent comparisons and reflect real-world clinical constraints such as alarm fatigue, a fixed decision threshold was selected using the tuning set; specifically, the threshold that capped the false positive rate at 20% (i.e., ensured at least 80% specificity). This threshold was then applied unchanged to the test set. It is important to note that threshold-dependent metrics (sensitivity, specificity, PPV, NPV, balanced accuracy, and F1-score) may vary if a different threshold is used, for example one that enforces a minimum sensitivity of 80%. Model Explainability To understand the decision-making process of our models, we employed Integrated Gradients. 18 We focused on the absolute (magnitude) of each feature’s attribution, rather than its signed contribution. Because time-series features can fluctuate, sometimes pushing the prediction higher and sometimes lower, summing signed attributions could lead to misleading cancellations. By taking the absolute value, we preserved each feature’s overall contribution across time and avoided under-representing features whose positive and negative influences might otherwise cancel out. We applied this approach to both the best respiratory and cardiovascular models at individual- patient and cohort-wide levels. At the individual level, we identified the most influential features for specific predictions; at the cohort level, we examined which features mattered most on average, providing deeper insight into the models’ overall behaviour. Research Environment All experiments were conducted within a secure Digital Research Environment provided by Aridhia Informatics Ltd in Glasgow, Scotland. The computational resources included an Intel® Xeon® Platinum 8272CL CPU with 64 GB of RAM; no GPU was utilized. Model development was performed using Python 3.12, leveraging the PyTorch framework (v2.6). Additional utilised libraries included Pandas, NumPy, X-transformers, Scikit-learn, Captum, Matplotlib, Seaborn, BioClinicalBert. Declarations Data Availability: The patient physiological data underlying the findings of this study subject to ethical and legal restrictions due to their sensitive patient information content. This dataset was obtained from routine clinical records held by the Children’s Acute Transport Service (CATS) and Great Ormond Street Hospital (GOSH). Our research use of this dataset was approved by GOSH Research and Innovation Division, under ethical approvals for use of routine de-identified healthcare and operational hospital data (Research Database, NHS REC reference 21/LO/0646). The authors did not receive any special privileges in accessing the data that other researchers would not have. Researchers seeking access to this dataset for secondary analysis must obtain permission directly from the data custodians. Interested researchers should contact the Research and Innovation Division at Great Ormond Street Hospital for Children NHS Foundation Trust ( [email protected] ). Data access applications are subject to a formal review process, including confirmation of appropriate ethical approval and the establishment of a Data Sharing Agreement with the requesting institution. Code Availability All code and trained models associated with this project are available at https://github.com/MilanKapur1/paediatric_tranport_deterioration_prediction. Acknowledgements We acknowledge the support from Kinseed for engineering the SwiftCare system used for the real-time extraction and secure server upload of high-frequency vital sign data directly from patient monitors during transport. We also thank,Great Ormond Street Hospital (GOSH) and the GOSH Digital Research Environment team for data curation.All research at Great Ormond Street Hospital NHS Foundation Trust and UCL Great Ormond Street Institute of Child Health is made possible by the NIHR Great Ormond Street Hospital Biomedical Research Centre. Funding MK is supported by an NIHR Academic Clinical Fellowship at University College London. KL is supported by UKRI Centre for Doctoral Training in AI-enabled healthcare systems. GD is supported by a UKRI Future Leaders Fellowship [MR/T041285/1]. PR is in receipt of grant support from the National Institute of Health Research, Rosetrees Trust and BMA Foundation. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health and Social Care. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Conflicts of Interest GD reports speaker honoraria from Vertex Pharmaceuticals and Chiesi Ltd, and advisory board and clinical trial leadership roles with Vertex, unrelated to the current manuscript. PR reports travel support for conference attendance from Fisher and Paykel Healthcare Limited. All other authors declare no conflicts of interest. Author Contribution M.K. and K.L. contributed equally as joint first authors. G.D. and P.R. contributed equally as joint senior authors. M.K., K.L., G.D., and P.R. conceived the study. P.K., P.R., Z.H. and J.B. were responsible for data acquisition and pre-processing. M.K. developed the models and performed the analysis. M.K., K.L., A.B., G.D., and P.R. contributed to the interpretation of the results. M.K. and K.L. wrote the manuscript. All authors reviewed and approved the final manuscript. References Paediatric Intensive Care Audit Network (PICANet). National Paediatric Critical Care Audit State of the Nation Report 2024 . https://www.picanet.org.uk/annual-reporting-and-publications/ (2024). Orr, R. A. et al. Pediatric specialized transport teams are associated with improved outcomes. Pediatrics 124 , 40–48 (2009). Singh, J. M., Gunz, A. C., Dhanani, S., Aghari, M. & MacDonald, R. D. Frequency, Composition, and Predictors of In-Transit Critical Events During Pediatric Critical Care Transport*. Pediatric Critical Care Medicine 17 , 984 (2016). Haydar, B. et al. Adverse Events During Intrahospital Transport of Critically Ill Children: A Systematic Review. Anesthesia & Analgesia 131 , 1135 (2020). Huo, Z. et al. Distribution and trajectory of vital signs from high-frequency continuous monitoring during pediatric critical care transport. Intensive Care Med. Paediatr. Neonatal 1 , 18 (2023). Fleming, S. et al. Normal ranges of heart rate and respiratory rate in children from birth to 18 years of age: a systematic review of observational studies. Lancet 377 , 1011–1018 (2011). Bonafide, C. P. et al. Development of heart and respiratory rate percentile curves for hospitalized children. Pediatrics 131 , e1150-1157 (2013). Nielsen, V. M. L., Kløjgård, T., Bruun, H., Søvsø, M. B. & Christensen, E. F. Progression of vital signs during ambulance transport categorised by a paediatric triage model: a population-based historical cohort study. BMJ Open 10 , e042401 (2020). Kapur, M. et al. Identification of physiological adverse events using continuous vital signs monitoring during paediatric critical care transport: a novel data-driven approach. 2025.03.11.25323742 Preprint at https://doi.org/10.1101/2025.03.11.25323742 (2025). Shamout, F. E. et al. An artificial intelligence system for predicting the deterioration of COVID-19 patients in the emergency department. npj Digit. Med. 4 , 1–11 (2021). Huo, Z. et al. Dynamic mortality prediction in critically Ill children during interhospital transports to PICUs using explainable AI. npj Digit. Med. 8 , 1–11 (2025). Sundrani, S. et al. Predicting patient decompensation from continuous physiologic monitoring in the emergency department. npj Digit. Med. 6 , 1–10 (2023). Jeon, Y., Kim, Y. S., Jang, W., Park, J. D. & Lee, B. Development of a deep learning model that predicts critical events of pediatric patients admitted to general wards. Sci Rep 14 , 4707 (2024). Mayampurath, A. et al. A Vital Sign-Based Model to Predict Clinical Deterioration in Hospitalized Children. Pediatr Crit Care Med 21 , 820–826 (2020). Rust, L. O. H. et al. The Deterioration Risk Index: Developing and Piloting a Machine Learning Algorithm to Reduce Pediatric Inpatient Deterioration. Pediatr Crit Care Med 24 , 322–333 (2023). Early Risk Prediction of Pediatric Cardiac Arrest from Electronic Health Records via Multimodal Fused Transformer. https://arxiv.org/html/2502.07158v2. Alsentzer, E. et al. Publicly Available Clinical BERT Embeddings. in Proceedings of the 2nd Clinical Natural Language Processing Workshop (eds. Rumshisky, A., Roberts, K., Bethard, S. & Naumann, T.) 72–78 (Association for Computational Linguistics, Minneapolis, Minnesota, USA, 2019). doi:10.18653/v1/W19-1909. Sundararajan, M., Taly, A. & Yan, Q. Axiomatic Attribution for Deep Networks. Preprint at https://doi.org/10.48550/arXiv.1703.01365 (2017). Gavelli, F., Teboul, J.-L. & Monnet, X. How can CO2-derived indices guide resuscitation in critically ill patients? Journal of Thoracic Disease 11 , (2019). Steitz, B. D. et al. Development and Validation of a Machine Learning Algorithm Using Clinical Pages to Predict Imminent Clinical Deterioration. J Gen Intern Med 39 , 27–35 (2024). Kwon, J.-M., Lee, Y., Lee, Y., Lee, S. & Park, J. An Algorithm Based on Deep Learning for Predicting In-Hospital Cardiac Arrest. J Am Heart Assoc 7 , e008678 (2018). Churpek, M. M. et al. Multicenter Development and Prospective Validation of eCARTv5: A Gradient-Boosted Machine-Learning Early Warning Score. Critical Care Explorations 7 , e1232 (2025). Straney, L. et al. Paediatric index of mortality 3: an updated model for predicting mortality in pediatric intensive care*. Pediatr Crit Care Med 14 , 673–681 (2013). NHS Digital. Data quality of protected characteristics and other vulnerable groups: Ethnicity. NHS England Digital https://digital.nhs.uk/data-and-information/data-collections-and-data-sets/data-sets/mental-health-services-data-set/submit-data/data-quality-of-protected-characteristics-and-other-vulnerable-groups/ethnicity. Vaswani, A. et al. Attention Is All You Need. Preprint at https://doi.org/10.48550/arXiv.1706.03762 (2023). Su, J. et al. RoFormer: Enhanced transformer with Rotary Position Embedding. Neurocomput. 568 , (2024). Additional Declarations No competing interests reported. Supplementary Files supplementarymaterial.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7319827","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":506164058,"identity":"134da596-2bd5-408c-b55a-672dbdb2d8c8","order_by":0,"name":"Milan Kapur","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAwElEQVRIiWNgGAWjYFACNjApx8DA3ABmSRCrxZiBgZFELYkNRGvRbWBL/FzYdi99w/GDDQw/ahgSZzYQ0GJ2gO2w9My24twNZxIbGHuOMSTOJmSL2QH2BmnetoTcDTeADuNtYEicR4SW5t9ALekGQC2Mf4nTwnYMZEsCSAszyBbCDjvMlmY941yC4UygXw7LHJMwJuz9423GtwvKEuT5jh8++PBNjY3sjAOErGEGIwg4QFREQnWNglEwCkbBKMADAJAEPj3MxUwfAAAAAElFTkSuQmCC","orcid":"","institution":"University College London","correspondingAuthor":true,"prefix":"","firstName":"Milan","middleName":"","lastName":"Kapur","suffix":""},{"id":506164059,"identity":"60e9e7e3-d6b6-410a-89c5-f67904492cd3","order_by":1,"name":"Kezhi Li","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"Kezhi","middleName":"","lastName":"Li","suffix":""},{"id":506164060,"identity":"fbb60820-8625-42ad-93b9-82832ec69b0f","order_by":2,"name":"Alexander Brown","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"Alexander","middleName":"","lastName":"Brown","suffix":""},{"id":506164061,"identity":"89e789d4-1427-4f9a-bb66-6a67bdcc08e6","order_by":3,"name":"Zhiqiang Huo","email":"","orcid":"","institution":"Wolfson Institute of Population Health, Queen Mary University of London","correspondingAuthor":false,"prefix":"","firstName":"Zhiqiang","middleName":"","lastName":"Huo","suffix":""},{"id":506164062,"identity":"2a08ce87-e6fd-4238-959b-16f5aa822dfa","order_by":4,"name":"John Booth","email":"","orcid":"","institution":"Digital Research Innovation and Virtual Environment (DRIVE), Great Ormond Street Hospital","correspondingAuthor":false,"prefix":"","firstName":"John","middleName":"","lastName":"Booth","suffix":""},{"id":506164063,"identity":"eb4e9ace-ed63-40fc-97a9-b314dc96a5d4","order_by":5,"name":"Philip Knight","email":"","orcid":"","institution":"Children’s Acute Transport Service (CATS), Great Ormond Street Hospital for Children NHS Foundation Trust","correspondingAuthor":false,"prefix":"","firstName":"Philip","middleName":"","lastName":"Knight","suffix":""},{"id":506164064,"identity":"78377b5b-5ecc-4b86-affb-ef9b1c2b446d","order_by":6,"name":"Gwyneth Davies","email":"","orcid":"","institution":"University College London","correspondingAuthor":false,"prefix":"","firstName":"Gwyneth","middleName":"","lastName":"Davies","suffix":""},{"id":506164065,"identity":"1fe88ce2-8ccd-41e9-9e6e-35c97090cc44","order_by":7,"name":"Padmanabhan Ramnarayan","email":"","orcid":"","institution":"Imperial College London","correspondingAuthor":false,"prefix":"","firstName":"Padmanabhan","middleName":"","lastName":"Ramnarayan","suffix":""}],"badges":[],"createdAt":"2025-08-07 14:38:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7319827/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7319827/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90487241,"identity":"39a33e1d-fe07-4dea-a386-e7875c4e0caa","added_by":"auto","created_at":"2025-09-03 09:01:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":43060,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of inclusion criteria. EHR: Electronic Health Record.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7319827/v1/9d081d68543b919c5e418649.png"},{"id":90485047,"identity":"15e57bce-0389-4e61-917e-cdd8756034e5","added_by":"auto","created_at":"2025-09-03 08:45:09","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":145535,"visible":true,"origin":"","legend":"\u003cp\u003ea) Stacked Receiver Operating Characteristic curves for the six respiratory. b) Stacked Precision-Recall curves for the trained respiratory models. c) Stacked ROC curves for the six cardiovascular. d) Stacked PR curves for the trained cardiovascular models. In all graphs we see that incorporating the vital sign data processing it with a transformer network boosts performance in both AUROC and AUPRC.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7319827/v1/274da4d175d11b663497890f.png"},{"id":90485037,"identity":"2149cd55-7720-4d75-bdbc-ef7323f232ba","added_by":"auto","created_at":"2025-09-03 08:45:08","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2634249,"visible":true,"origin":"","legend":"\u003cp\u003ea) Relative importance of the features used by the respiratory model in predicting adverse respiratory events. b) Relative importance of the features used by the cardiovascular model in predicting cardiovascular events. Feature importance for both graphs calculated on the holdout test set.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7319827/v1/d40923ba6b0a54353c0c75df.png"},{"id":90485038,"identity":"a7b11eee-a33c-455c-8f31-ecd49623ad32","added_by":"auto","created_at":"2025-09-03 08:45:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1847615,"visible":true,"origin":"","legend":"\u003cp\u003eFeature importance in predictions made for a given data window in the same transport episode. (a) For the respiratory model, the predicted probability of an adverse respiratory event was 0.91 (and the event did occur). The most influential features were arterial and non-invasive blood pressure, previous cardiovascular deteriorations, EtCO₂, respiratory rate, SpO₂, and previous respiratory deteriorations. (b) For the cardiovascular model, the predicted probability of an adverse cardiovascular event was 0.18 (and the event did not occur). Non-invasive and arterial blood pressure, EtCO₂, heart rate, and temperature had the greatest impact on the prediction.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7319827/v1/6d2125c1a755238f4d8663ce.png"},{"id":90485035,"identity":"6873bb4b-853c-4c71-aee5-eaa797898b39","added_by":"auto","created_at":"2025-09-03 08:45:08","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":356777,"visible":true,"origin":"","legend":"\u003cp\u003ea) Vital sign data were collected during the interhospital transfer. b) Corresponding demographic and clinical information were extracted from the EHR at GOSH. c) Data was cleaned with physiologically implausible values removed. d) Respiratory and cardiovascular events were labelled using a Bollinger Bands-based method, applying patient-specific dynamic thresholds to identify sustained combinations of significant deviations in vital signs. e) Vital signs were normalised/standardised to allow efficient training. f) Missing vital sign values were mean imputed. g) 15-minute windows of data were extracted to convert each transport episode into multiple windows of data. These windows were combined with the demographic and clinical features derived from the EHR to make a complete feature set for the window. Care was taken to ensure data from a given transport episode belonged exclusively to either the train, tune or test sets. h) Models were then trained for the classification task: to predict whether an adverse respiratory or cardiovascular event occurred in the 15-minute label period. i) Model performance was evaluated using a range of metrics. j) Integrated Gradients were used to understand the features that drive the model to make predictions at a given time-point at a cohort and individual level.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7319827/v1/c4fd6287ca97ff2fb49e6ffd.png"},{"id":90488304,"identity":"7715c278-665f-422f-8607-384d52985800","added_by":"auto","created_at":"2025-09-03 09:09:08","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":21983,"visible":true,"origin":"","legend":"\u003cp\u003eThis diagram illustrates how each transport episode is divided into sequential time windows for predictive modelling of adverse events. The 15-minute interval highlighted in red is the prediction window, during which adverse events may occur. The 15-minute interval in blue represents the immediately preceding 15-minute period where vital signs are captured at 1-second resolution. The green section indicates an extended historical context (up to a maximum of 120 minutes prior), where vital signs are down-sampled to 5-minute averages to reduce computational overhead while preserving key trends. Each horizontal “track” corresponds to a different 15-minute prediction window, demonstrating how multiple windows are formed across the full duration of a transport episode.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7319827/v1/a15ddb098a152d2f9c833b10.png"},{"id":94988833,"identity":"2dcf8436-eea7-4725-bc7b-2757d8d48e28","added_by":"auto","created_at":"2025-11-03 07:11:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":8482838,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7319827/v1/8207cd6b-2434-47e6-bffc-534c9896eadc.pdf"},{"id":90485031,"identity":"f615cede-ac5e-4a93-88a7-792221333de9","added_by":"auto","created_at":"2025-09-03 08:45:08","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":967584,"visible":true,"origin":"","legend":"","description":"","filename":"supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-7319827/v1/ff34f59a423bf65f55ea88d2.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Real-time prediction of cardiorespiratory deterioration during paediatric critical care transport using interpretable machine learning","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIn the United Kingdom, around 20,000 children require admission to Paediatric Intensive Care Units (PICUs) annually.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e Many initially present to general hospitals that lack on-site PICU services, making urgent interhospital transfer essential for them to receive necessary intensive care. In 2023, more than 4,000 such transfers were conducted.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e Despite being conducted by specialised Paediatric Critical Care Transport (PCCT) teams of doctors, nurses, and paramedics, interhospital transfers remain high-risk.\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e Transport-related physiological stress elevates the risk of en-route clinical deterioration, with adverse events reported in 12.3% of over 8,000 interhospital transport episodes in one large study.\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e A recent systematic review categorised transport associated adverse events into four main groups: respiratory (e.g., hypoxaemia), cardiovascular (e.g., hypotension, tachycardia/bradycardia), equipment-related (e.g., monitor failure), and other (e.g., medication error).\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e These events, even when promptly identified and managed by PCCT teams, can lead to significant morbidity and even mortality, underscoring the critical need for advancements that move beyond reactive care to proactive intervention.\u003c/p\u003e\u003cp\u003eDuring transport, vital signs are continuously monitored to detect deterioration. However, PCCT services largely rely on manual interpretation against static, age-based reference ranges or clinical judgment. This approach has limitations, as \"normal\" vital sign ranges for critically unwell children vary widely with age, diagnosis, and illness severity. However, this approach requires well-characterized normal vital sign ranges for the patient. Whilst these ranges exist for stable paediatric patients, there are no \u0026ldquo;normal\u0026rdquo; ranges for critically unwell children during transport; the expected values vary significantly depending on age, diagnosis, severity of illness and ongoing interventions.\u003csup\u003e\u003cspan additionalcitationids=\"CR6 CR7\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e This often leads to either over-alerting (false alarms due to transient fluctuations in already abnormal physiology) or under-detection of subtle but significant changes. To address this, our prior work explored dynamic thresholding approaches that better account for inter-individual variability, allowing for personalised and adaptive detection of patient deterioration.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eWhile detecting deterioration in real time is critical for timely intervention, predicting future deterioration enables clinicians to act pre-emptively before a patient\u0026rsquo;s condition worsens. Recent machine learning (ML) advancements allow sophisticated predictive tools to leverage large patient datasets, analysing continuous physiological data to forecast acute deterioration. \u003csup\u003e\u003cspan additionalcitationids=\"CR11 CR12\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e Despite these advancements, to the best of our knowledge, no models exist capable of predicting real-time risk of acute deterioration during paediatric critical care transport. To address this gap, we introduce two novel, lightweight ML models capable of predicting respiratory and cardiovascular deterioration up to 15 minutes in advance, in real-time. Trained on a large dataset of 1200 transport episodes with a lightweright design suitable for edge computation, these models have the potential to provide timely clinical decision support during transport, which may in turn enable earlier intervention and contribute to improved patient outcomes.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cb\u003eCohort Description\u003c/b\u003e: his retrospective study analysed continuously monitored vital signs and linked electronic health record (EHR) data from critically ill children transported by the Children\u0026rsquo;s Acute Transport Service (CATS) in North London, UK, between January 2016 and May 2021. Of the 6,471 transport episodes during this period, data capture was significantly limited by operational constraints: only two recording devices were available for three transport teams, making full coverage unfeasible. Additional data loss occurred when devices were forgotten or due to technical issues maintaining stable monitor-device-server connections. Consequently, 1,781 episodes had sufficiently complete monitoring and EHR data for initial consideration. After further excluding patients over 18 years old and those with \u0026lt;\u0026thinsp;30 minutes of monitoring, a final cohort of 1,519 episodes (23.5% of the total 6,471) met inclusion criteria (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the included cohort was demographically and clinically representative of the overall cohort. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e summarizes the availability of key variables; not all parameters were recorded for every patient due to clinical variability (e.g., EtCO₂ only for invasively ventilated patients). The median duration of a transport episode was 113 minutes (IQR: 75\u0026ndash;160) (Supplementary Fig.\u0026nbsp;1). For our predictive analysis, each transport was segmented into 15-minute time windows, within which we aimed to predict the occurrence of at least one adverse event. Adverse events were uncommon within these windows (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), suggesting patients were generally stable during transport.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eBreakdown of characteristics of all transport episodes and included episodes\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCharacteristics\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAll Transport Episodes (n\u0026thinsp;=\u0026thinsp;6471)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIncluded Episodes \u003cem\u003e(n\u0026thinsp;=\u0026thinsp;1519)\u003c/em\u003e\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge Group Distribution\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026le;\u0026thinsp;1 month\u003c/p\u003e \u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2219 (34.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e568 (37.4%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u0026ndash; \u0026le;12 months\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1330 (20.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e296 (19.5%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u0026ndash; \u0026le;4 years\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1198 (18.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e264 (17.4%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4\u0026ndash; \u0026le;11 years\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1017 (15.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e239 (15.7%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e11\u0026ndash; \u0026le;18 years\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e678 (10.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e151 (9.9%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026gt;\u0026thinsp;18 years\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e27 (0.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e0 (0%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eGender Distribution\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3603 (55.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e825 (54.3%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFemale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2861 (44.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e692 (45.6%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDiagnosis Group Distribution\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRespiratory\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2233 (34.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e498 (32.8%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCardiovascular\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1385 (21.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e364 (24.0%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNeurological\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e921 (14.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e227 (14.9%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInfection\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e888 (13.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e202 (13.3%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGastrointestinal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e368 (5.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e89 (5.9%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMetabolic\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e161 (2.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e43 (2.8%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTrauma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e81 (1.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e21 (1.4%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOther\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e434 (6.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e75 (4.9%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003ePIM3 Risk of Mortality\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026le;\u0026thinsp;1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e502 (7.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e92 (6.1%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u0026ndash; \u0026le;3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2345 (36.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e510 (33.6%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u0026ndash; \u0026le;5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2060 (31.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e518 (34.1%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5\u0026ndash; \u0026le;10%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1038 (16.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e264 (17.4%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e10\u0026ndash; \u0026le;15%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e206 (3.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e59 (3.9%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e15\u0026ndash; \u0026le;30%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e182 (2.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e45 (3.0%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026gt;\u0026thinsp;30%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e126 (1.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e30 (2.0%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eRespiratory Support\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSelf-ventilating (Room Air)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1292 (20.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e257 (16.9%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSelf-ventilating (supplemental O₂)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e141 (2.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e27 (1.8%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSelf-ventilating (HFNC)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e250 (3.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e62 (4.1%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSelf-ventilating (CPAP)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e278 (4.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e55 (3.6%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSelf-ventilating (BIPAP)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e54 (0.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e12 (0.8%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInvasive ventilation (ETT)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4214 (65.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e1073 (70.6%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInvasive ventilation (Tracheostomy)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e93 (1.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e19 (1.3%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInvasive ventilation (Other airway)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e10 (0.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e3 (0.2%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCardiovascular Support\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAdrenaline\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e815 (12.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e199 (13.1%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDobutamine\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e31 (0.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e9 (0.6%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDopamine\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e580 (9.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e137 (9.0%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMilrinone\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e58 (0.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e12 (0.8%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNoradrenaline\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e507 (7.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e131 (8.6%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAny agent\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2011 (31.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e488 (32.1%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eOverall Transport Time, minutes\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026le;\u0026thinsp;60\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e16 (0.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e0 (0.0%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e60\u0026ndash;120\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e598 (9.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e105 (6.9%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e120\u0026ndash;180\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1523 (23.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e345 (22.7%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e180\u0026ndash;240\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2067 (31.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e545 (35.9%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e240\u0026ndash;300\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1282 (19.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e322 (21.2%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e300\u0026ndash;360\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e556 (8.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e142 (9.3%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026gt;\u0026thinsp;360\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e261 (4.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cem\u003e55 (3.6%)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eDemographic, clinical, and transport characteristics of the study population compared to all transported children during the study period. Respiratory support and cardiovascluar support refer to support received during transport. HFNC\u0026thinsp;=\u0026thinsp;High-Flow Nasal Cannula; CPAP\u0026thinsp;=\u0026thinsp;Continuous Positive Airway Pressure; BIPAP\u0026thinsp;=\u0026thinsp;Bilevel Positive Airway Pressure; ETT\u0026thinsp;=\u0026thinsp;Endotracheal Tube.\u003c/p\u003e\u003ctable id=\"Tab2\" border=\"1\" width=\"120%\" class=\"fr-table-selection-hover\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eList of collected Electronic Health Care (EHR) and vital sign data in the cohort\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCategory\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 43.3142%;\"\u003e\n \u003cp\u003eEHR and vital signs data\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 15.7254%;\"\u003e\n \u003cp\u003eData type\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 13.8236%;\"\u003e\n \u003cp\u003eRange\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMissing (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePatient demographics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 43.3142%;\"\u003e\n \u003cp\u003eAge (years),\u003c/p\u003e\n \u003cp\u003eWeight (kg)\u003c/p\u003e\n \u003cp\u003eSex\u003c/p\u003e\n \u003cp\u003eEthnicity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 15.7254%;\"\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 13.8236%;\"\u003e\n \u003cp\u003e0\u0026ndash;18\u003c/p\u003e\n \u003cp\u003e1\u0026ndash;90\u003c/p\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.00%\u003c/p\u003e\n \u003cp\u003e0.26%\u003c/p\u003e\n \u003cp\u003e0.07%\u003c/p\u003e\n \u003cp\u003e51.68%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTransport details\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 43.3142%;\"\u003e\n \u003cp\u003eReferring hospital\u003c/p\u003e\n \u003cp\u003eDestination hospital\u003c/p\u003e\n \u003cp\u003eTime of transfer\u003c/p\u003e\n \u003cp\u003eDestination care area (PICU, NICU, HDU Ward, other)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 15.7254%;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 13.8236%;\"\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.00%\u003c/p\u003e\n \u003cp\u003e0.26%\u003c/p\u003e\n \u003cp\u003e0.00%\u003c/p\u003e\n \u003cp\u003e0.20%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMedical Diagnosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 43.3142%;\"\u003e\n \u003cp\u003ePrimary diagnosis,\u003c/p\u003e\n \u003cp\u003ePaediatric index of mortality 3 (PIM3),\u003c/p\u003e\n \u003cp\u003eExisting medical conditions (respiratory, cardiac, renal, genetic, metabolic/endocrine, haematological/oncological, other)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 15.7254%;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 13.8236%;\"\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003cp\u003e0\u0026ndash;1\u003c/p\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.00%\u003c/p\u003e\n \u003cp\u003e0.26%\u003c/p\u003e\n \u003cp\u003e0.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInterventions by local team prior to transport\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 43.3142%;\"\u003e\n \u003cp\u003eIntubation (primary/re-intubation/ETT repositioning), mechanical ventilation (invasive, non-invasive and HFNC), suctioning, chest drain insertion, vascular access (peripheral, central, arterial, intraosseous), vasoactive support (inotropes/vasopressors, prostaglandin), blood product transfusion, urinary catheterisation, nasogastric/orogastric tube placement, imaging (CT scan), C-spine immobilisation, osmotherapy, CPR/defibrillation, ECMO, and ICP monitoring\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 15.7254%;\"\u003e\n \u003cp\u003eBinary\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 13.8236%;\"\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.79%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIntra-transport respiratory support commenced prior to transport\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 43.3142%;\"\u003e\n \u003cp\u003eSelf-ventilating: Room air, Supplemental O₂, HFNC, CPAP, BiPAP; Invasive ventilation: ETT, Tracheostomy, Other airway\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 15.7254%;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 13.8236%;\"\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.72%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIntra-transport cardiovascular support commenced prior to transport\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 43.3142%;\"\u003e\n \u003cp\u003eAdrenaline, noradrenaline, dobutamine, dopamine, milrinone, prostaglandin, inhaled nitric oxide\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 15.7254%;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 13.8236%;\"\u003e\n \u003cp\u003eNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.00%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVital Signs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 43.3142%;\"\u003e\n \u003cp\u003eSpO2\u003c/p\u003e\n \u003cp\u003eHeart rate (3-lead ECG)\u003c/p\u003e\n \u003cp\u003eEnd tidal CO2 (minimum value in 1-second period)\u003c/p\u003e\n \u003cp\u003eEnd tidal CO2 (maximum value in 1-second period)\u003c/p\u003e\n \u003cp\u003eAirway derived respiratory rate\u003c/p\u003e\n \u003cp\u003eImpedance pneumography derived respiratory rate\u003c/p\u003e\n \u003cp\u003eMean systolic blood pressure (non-invasive)\u003c/p\u003e\n \u003cp\u003eMean arterial blood pressure (non-invasive)\u003c/p\u003e\n \u003cp\u003eMean diastolic blood pressure (non-invasive)\u003c/p\u003e\n \u003cp\u003eMean systolic blood pressure (invasive)\u003c/p\u003e\n \u003cp\u003eMean arterial blood pressure (invasive)\u003c/p\u003e\n \u003cp\u003eMean diastolic blood pressure (invasive)\u003c/p\u003e\n \u003cp\u003eTemperature (oesophageal)\u003c/p\u003e\n \u003cp\u003eTemperature (skin)\u003c/p\u003e\n \u003cp\u003eTemperature (core)\u003c/p\u003e\n \u003cp\u003eTemperature (unspecified)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 15.7254%;\"\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003cp\u003eNumerical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 13.8236%;\"\u003e\n \u003cp\u003e0-100\u003c/p\u003e\n \u003cp\u003e0-301\u003c/p\u003e\n \u003cp\u003e0-19.1\u003c/p\u003e\n \u003cp\u003e0.3\u0026ndash;20.1\u003c/p\u003e\n \u003cp\u003e0-164\u003c/p\u003e\n \u003cp\u003e0-171\u003c/p\u003e\n \u003cp\u003e0-264\u003c/p\u003e\n \u003cp\u003e0-242\u003c/p\u003e\n \u003cp\u003e0-247\u003c/p\u003e\n \u003cp\u003e0-361\u003c/p\u003e\n \u003cp\u003e0-361\u003c/p\u003e\n \u003cp\u003e0-340\u003c/p\u003e\n \u003cp\u003e2.9\u0026ndash;40.8\u003c/p\u003e\n \u003cp\u003e6.6\u0026ndash;43.4\u003c/p\u003e\n \u003cp\u003e8.2\u0026ndash;44.6\u003c/p\u003e\n \u003cp\u003e5.8\u0026ndash;43.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.07%\u003c/p\u003e\n \u003cp\u003e0.59%\u003c/p\u003e\n \u003cp\u003e29.00%\u003c/p\u003e\n \u003cp\u003e27.32%\u003c/p\u003e\n \u003cp\u003e27.39%\u003c/p\u003e\n \u003cp\u003e4.08%\u003c/p\u003e\n \u003cp\u003e5.99%\u003c/p\u003e\n \u003cp\u003e5.99%\u003c/p\u003e\n \u003cp\u003e5.99%\u003c/p\u003e\n \u003cp\u003e63.98%\u003c/p\u003e\n \u003cp\u003e63.98%\u003c/p\u003e\n \u003cp\u003e63.98%\u003c/p\u003e\n \u003cp\u003e78.74%\u003c/p\u003e\n \u003cp\u003e72.55%\u003c/p\u003e\n \u003cp\u003e85.71%\u003c/p\u003e\n \u003cp\u003e64.05%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e\u003cp\u003eMissing rates were calculated on the 1519 patients included in the study and captures transport episodes where there were zero recorded values of a given parameter. PICU\u0026thinsp;=\u0026thinsp;Paediatric Intensive Care Unit, NICU\u0026thinsp;=\u0026thinsp;Neonatal Intensive Care Unit, HDU\u0026thinsp;=\u0026thinsp;High Dependency Unit, ETT\u0026thinsp;=\u0026thinsp;Endotracheal tube, HFNC\u0026thinsp;=\u0026thinsp;High Flow Nasal Canula, CPR\u0026thinsp;=\u0026thinsp;Cardiopulmonary Resuscitation, ECMO\u0026thinsp;=\u0026thinsp;Extra-Corporeal Membrane Oxygenation, ICP\u0026thinsp;=\u0026thinsp;Intracranial Pressure, CPAP\u0026thinsp;=\u0026thinsp;Continuous Positive Airway Pressure, BiPAP\u0026thinsp;=\u0026thinsp;Bi-level Positive Airway Pressure.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eFrequency of Adverse Events in Time Windows.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMetric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTrain\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTune\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTest\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTotal Windows\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e8109\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1022\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1031\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRespiratory Deterioration (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2.48%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2.25%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e3.1%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCardiac Deterioration (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4.06%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e5.19%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e4.36%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBoth Respiratory and Cardiac Deterioration (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.57%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.59%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.87%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNo Deterioration (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e92.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e91.98%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e91.66%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThis table summarizes the total number of time windows analysed in the train, tune, and test sets, along with the percentages of windows exhibiting respiratory deterioration, cardiac deterioration, both respiratory and cardiac deterioration, and no deterioration. Notably, adverse events are rare, with windows showing any type of deterioration collectively accounting for less than 10% of the total. Time windows from a given patient belong exclusively to one of the three groups.\u003c/p\u003e\u003cp\u003e\u003cb\u003eModel Performance\u003c/b\u003e: Performance was evaluated on the holdout test set. Models were provided with the 15-minutes of high-resolution vital signs data preceding the start of the 15-minute prediction window, alongside averaged historical context from up to 120 minutes prior and liked EHR data. Six models were developed for each prediction task (respiratory and cardiovascular deterioration), progressively increasing in feature richness and computational complexity (Supplementary Table\u0026nbsp;1). Models were then assessed on their ability to predict the occurrence of at least one adverse respiratory or cardiovascular event in the subsequent 15-minute period. Performance was evaluated on a random label-stratified holdout test set using standard metrics: Area Under the Receiver Operating Characteristic Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), Sensitivity (Recall), Specificity, Positive Predictive Value (PPV, Precision), Negative Predictive Value (NPV), Balanced Accuracy, and F1-score. A fixed decision threshold was chosen on the tuning set to cap the false positive rate at 20% (i.e., \u0026ge;\u0026thinsp;80% specificity). ROC and PR curves are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cb\u003eRespiratory Events\u003c/b\u003e: The baseline-only model, relying solely on demographic and pre-transport features, achieved an AUROC of 0.669 (95% CI: 0.578\u0026ndash;0.757) and an AUPRC of 0.099 (95% CI: 0.064\u0026ndash;0.145), indicating modest predictive capability. Enhancing the model by incorporating time-series vital signs through a simple feed-forward network (Combined FF) improved performance to an AUROC of 0.800 (95% CI: 0.746\u0026ndash;0.850) and an AUPRC of 0.139 (95% CI: 0.090\u0026ndash;0.215). Further gains were observed when transformer-based architectures were applied. A transformer-based model, using only the time-series vital signs (\u0026ldquo;Vitals-Only Transformer\u0026rdquo;) achieved an AUROC of 0.846 (95% CI: 0.799\u0026ndash;0.891) and an AUPRC of 0.201 (95% CI: 0.124\u0026ndash;0.300), with a sensitivity of 0.707 (95% CI: 0.561\u0026ndash;0.829) and balanced accuracy of 0.755 (95% CI: 0.684\u0026ndash;0.823). We also tested combining time-series vital signs with differing combinations of diagnosis and baseline variables. Among the combined transformer models, the approach using one-hot encoded diagnosis with a reduced baseline feature set reached an AUROC of 0.850 (95% CI: 0.805\u0026ndash;0.892), an AUPRC of 0.173 (95% CI: 0.119\u0026ndash;0.247), balanced accuracy of 0.764 (95% CI: 0.693\u0026ndash;0.832), and an F1-score of 0.235 (95% CI: 0.188\u0026ndash;0.282). The variant using vector-embedded diagnosis (with the same reduced baseline feature set) showed a comparable AUROC of 0.851 (95% CI: 0.805\u0026ndash;0.894) but achieved a higher AUPRC of 0.200 (95% CI: 0.127\u0026ndash;0.298) at the expense of a slightly lower F1-score (0.216, 95% CI: 0.174\u0026ndash;0.257). Notably, including the full set of baseline features in the transformer (Vector Diagnosis, Full Baseline) did not improve performance, yielding an AUROC of 0.841 (95% CI: 0.789\u0026ndash;0.889) and balanced accuracy of 0.754 (95% CI: 0.685\u0026ndash;0.823).\u003c/p\u003e\u003cp\u003e\u003cb\u003eCardiovascular Events\u003c/b\u003e: For cardiovascular deterioration, the baseline-only model again demonstrated limited performance, with an AUROC of 0.652 (95% CI: 0.594\u0026ndash;0.709) and an AUPRC of 0.076 (95% CI: 0.064\u0026ndash;0.091). The Combined FF model, which incorporated time-series vital signs, improved these metrics to an AUROC of 0.763 (95% CI: 0.697\u0026ndash;0.823) and an AUPRC of 0.149 (95% CI: 0.110\u0026ndash;0.200). The Vitals-Only Transformer produced similar results (AUROC 0.759, 95% CI: 0.704\u0026ndash;0.813; AUPRC 0.149, 95% CI: 0.102\u0026ndash;0.213). Among the transformer-based approaches, the Combined Transformer using one-hot encoded diagnosis with a reduced baseline feature set achieved an AUROC of 0.772 (95% CI: 0.720\u0026ndash;0.823) and an AUPRC of 0.144 (95% CI: 0.103\u0026ndash;0.204). The best performance for cardiovascular prediction was obtained with the Combined Transformer employing vector-embedded diagnosis (with the same reduced baseline feature set), which achieved an AUROC of 0.792 (95% CI: 0.739\u0026ndash;0.842), an AUPRC of 0.183 (95% CI: 0.122\u0026ndash;0.261), a balanced accuracy of 0.677 (95% CI: 0.611\u0026ndash;0.747), and an F1-score of 0.242 (95% CI: 0.182\u0026ndash;0.308). As seen in the respiratory models, incorporating the full set of baseline features (Vector Diagnosis, Full Baseline) did not offer additional benefits, with this configuration attaining an AUROC of 0.786 (95% CI: 0.735\u0026ndash;0.833) and an AUPRC of 0.135 (95%CI: 0.105\u0026ndash;0.173).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eResults of evaluating the model on the holdout test set\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAUROC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAUPRC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003ePPV\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eNPV\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eBalanced Accuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eF1-Score\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRespiratory Models\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBaseline-Only FF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.669 (0.578\u0026ndash;0.757)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.099 (0.064\u0026ndash;0.145)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.413 (0.268\u0026ndash;0.561)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.892 (0.873\u0026ndash;0.911)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.137 (0.090\u0026ndash;0.185)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.973 (0.967\u0026ndash;0.980)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.653 (0.579\u0026ndash;0.729)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.206 (0.136\u0026ndash;0.277)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCombined FF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.800 (0.746\u0026ndash;0.850)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.139 (0.090\u0026ndash;0.215)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.610 (0.463\u0026ndash;0.756)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.803 (0.778\u0026ndash;0.828)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.114 (0.086\u0026ndash;0.142)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.980 (0.973\u0026ndash;0.988)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.706 (0.629\u0026ndash;0.780)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.192 (0.145\u0026ndash;0.237)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVitals-Only Transformer\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.846 (0.799\u0026ndash;0.891)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.201 (0.124-0.300)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.707 (0.561\u0026ndash;0.829)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.804 (0.779\u0026ndash;0.828)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.130 (0.104\u0026ndash;0.157)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.985 (0.978\u0026ndash;0.991)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.755 (0.684\u0026ndash;0.823)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.220 (0.177\u0026ndash;0.263)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCombined Transformer (One-Hot Diagnosis, Reduced Baseline)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.850 (0.805\u0026ndash;0.892)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.173 (0.119\u0026ndash;0.247)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.707 (0.561\u0026ndash;0.829)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.821 (0.797\u0026ndash;0.844)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.141 (0.112\u0026ndash;0.170)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.985 (0.979\u0026ndash;0.992)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.764 (0.693\u0026ndash;0.832)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.235 (0.188\u0026ndash;0.282)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCombined Transformer (Vector Diagnosis, Reduced Baseline)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.851 (0.805\u0026ndash;0.894)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.200 (0.127\u0026ndash;0.298)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.730 (0.585\u0026ndash;0.854)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.792 (0.766\u0026ndash;0.817)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.127 (0.102\u0026ndash;0.152)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.986 (0.979\u0026ndash;0.993)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.761 (0.690\u0026ndash;0.828)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.216 (0.174\u0026ndash;0.257)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCombined Transformer (Vector Diagnosis, Full Baseline)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.841 (0.789\u0026ndash;0.889)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.177 (0.121\u0026ndash;0.251)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.70 7 (0.561\u0026ndash;0.830)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.802 (0.777\u0026ndash;0.826)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.129 (0.104\u0026ndash;0.155)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.985 (0.978\u0026ndash;0.992)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.754 (0.685\u0026ndash;0.823)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.218 (0.176\u0026ndash;0.261)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCardiovascular Models\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBaseline-Only FF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.652 (0.594\u0026ndash;0.709)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.076 (0.064\u0026ndash;0.091)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.037 (0.000-0.093)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.892 (0.871\u0026ndash;0.911)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.019 (0.000-0.047)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.944 (0.941\u0026ndash;0.947)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.464 (0.441\u0026ndash;0.494)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.025 (0.000-0.062)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCombined FF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.763 (0.697\u0026ndash;0.823)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.149 (0.110\u0026ndash;0.200)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.481 (0.352\u0026ndash;0.611)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.833 (0.810\u0026ndash;0.856)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.138 (0.102\u0026ndash;0.175)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.967 (0.959\u0026ndash;0.975)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.657 (0.591\u0026ndash;0.727)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.214 (0.158\u0026ndash;0.271)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVitals-Only Transformer\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.759 (0.704\u0026ndash;0.813)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.149 (0.102\u0026ndash;0.213)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.500 (0.370\u0026ndash;0.630)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.822 (0.798\u0026ndash;0.845)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.135 (0.101\u0026ndash;0.170)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.968 (0.959\u0026ndash;0.976)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.661 (0.595\u0026ndash;0.728)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.212 (0.159\u0026ndash;0.266)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCombined Transformer (One-Hot Diagnosis, Reduced Baseline)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.772 (0.720\u0026ndash;0.823)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.144 (0.103\u0026ndash;0.204)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.463 (0.333\u0026ndash;0.593)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.845 (0.822\u0026ndash;0.866)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.142 (0.103\u0026ndash;0.181)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.966 (0.958\u0026ndash;0.974)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.654 (0.586\u0026ndash;0.721)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.217 (0.157\u0026ndash;0.277)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCombined Transformer (Vector Diagnosis, Reduced Baseline)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.792 (0.739\u0026ndash;0.842)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.183 (0.122\u0026ndash;0.261)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.500 (0.370\u0026ndash;0.630)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.855 (0.832\u0026ndash;0.876)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.160 (0.120\u0026ndash;0.204)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.969 (0.961\u0026ndash;0.977)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.677 (0.611\u0026ndash;0.747)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.242 (0.182\u0026ndash;0.308)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCombined Transformer (Vector Diagnosis, Full Baseline)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.786 (0.735\u0026ndash;0.833)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.135 (0.105\u0026ndash;0.173)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.501 (0.370\u0026ndash;0.630)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.831 (0.808\u0026ndash;0.855)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.141 (0.105\u0026ndash;0.178)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.968 (0.960\u0026ndash;0.976)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.666 (0.599\u0026ndash;0.734)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.220 (0.164\u0026ndash;0.275)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e reports the results of evaluating the trained models for each of the two prediction tasks on the holdout test set. The decision threshold was determined on the tuning set by selecting the value that yielded a maximum false positive rate (FPR) of 20% (i.e., a minimum specificity of 80%). The values in brackets represent the 95% confidence interval calculated by performing stratified bootstrapping with replacement on the test set. \"FF\" refers to a feed-forward neural network. AUROC - Area Under the Receiver Operating Characteristic curve, AUPRC - Area Under the Precision-Recall Curve, PPV - Positive Predictive Value, NPV - Negative Predictive Value.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eModel Explainability\u003c/strong\u003e\u003cp\u003eWe employed the Integrated Gradients method to identify which input features most influenced model predictions. This analysis focused on the best-performing models (Combined Transformer (Vector Diagnosis, Reduced Baseline)) for both respiratory and cardiovascular deterioration, based on AUROC. Feature attributions are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\u003c/p\u003e\u003cp\u003eIn the respiratory model, the most influential features were non-invasive blood pressure, previous cardiovascular deteriorations, end-tidal CO₂ (EtCO₂), diagnosis, and previous respiratory deteriorations, followed by blood oxygen saturation (SpO₂) and respiratory rate. Positive predictions relied more on SpO₂, while negative predictions emphasised diagnosis and prior respiratory deteriorations. For the cardiovascular model, EtCO₂, non-invasive and arterial blood pressure, heart rate, and prior deteriorations were most important, with arterial pressure weighted more in positive predictions and diagnosis in negative ones. Overall, time-series vital sign features were more important than baseline demographic features, though both informed the final predictions.\u003c/p\u003e\u003cp\u003eThe same feature attribution method can be applied dynamically to individual prediction windows, offering real-time insight into how each input influences the model\u0026rsquo;s output. For each 15-minute window, the model generates a probability (0\u0026ndash;1) of an adverse event occurring in the following 15 minutes. In Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea, a high predicted probability (0.91) for respiratory deterioration was followed by an actual event; key contributing features included arterial and non-invasive blood pressure, prior cardiovascular events, EtCO₂, respiratory rate, SpO₂, and prior respiratory events. In contrast, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb shows a low predicted probability (0.18) for cardiovascular deterioration, with no event occurring. The most influential features in this case included blood pressure, EtCO₂, heart rate, and temperature.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eTo our knowledge, this is the first study to use continuously collected vital signs and linked EHR data during interhospital transport of critically ill children to predict real-time risk of deterioration. We present two lightweight, explainable machine learning models that integrate live-streamed physiology with baseline clinical data to forecast respiratory and cardiovascular events up to 15 minutes in advance, potentially enabling earlier intervention and preventing decline.\u003c/p\u003e\u003cp\u003ePrevious studies have largely focused on predicting mortality during transport or deterioration in stable, non-critically ill paediatric inpatients. \u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan additionalcitationids=\"CR14 CR15\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e–\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e While several of these models have demonstrated strong performance in predicting ward-to-PICU transfers or ICU mortality, none have been specifically developed to predict imminent deterioration in critically ill children during interhospital transport Our work addresses this gap, introducing a real-time risk prediction tool for one of the most high-risk settings in paediatric care.\u003c/p\u003e\u003cp\u003eThe models were developed using a diverse dataset of over 1,500 interhospital transports conducted by the CATS team in central London. We adopted a systematic approach, starting with models based solely on demographic and pre-transport features, then incorporating high-resolution time-series data. Initial models used simple feed-forward networks, progressing to transformer-based architectures for more advanced temporal modelling. We also evaluated multiple strategies for diagnosis representation, comparing one-hot encoding against vector embeddings derived from a pretrained clinical language model.\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e This iterative, constructive approach allowed us to quantify how predictive performance scaled with increasing model complexity and feature richness, while identifying the minimal configuration needed to maintain high accuracy. High-frequency vital-sign data, particularly when processed using transformer-based models capable of capturing complex temporal dynamics significantly boosted accuracy (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eOur best-performing models for both tasks used the Combined Transformer (Vector Diagnosis, Reduced Baseline) architecture. On the holdout test set, the respiratory deterioration model achieved an AUROC of 0.851 and AUPRC of 0.200, with a sensitivity of 0.730 and specificity of 0.792. The cardiovascular model yielded an AUROC of 0.792 and AUPRC of 0.183, with a sensitivity of 0.500 and specificity of 0.855. These results are particularly notable given the low baseline incidence of events in the test set: only 3.1% of time windows included respiratory deterioration, 4.4% included cardiovascular deterioration, and just 0.87% involved both. This extreme class imbalance underscores the models’ ability to extract clinically salient signals from noisy transport data.\u003c/p\u003e\u003cp\u003eBeyond overall performance, we used Integrated Gradients to understand how the best-performing models made predictions, assessing feature importance at both the cohort and individual level.\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e In the respiratory model, the top contributors were non-invasive blood pressure, prior cardiovascular events, and EtCO₂, followed by diagnosis, prior respiratory events, and SpO₂, likely reflecting the interdependent nature of respiratory and cardiovascular physiology, where cardiovascular instability is often associated with respiratory instability. The cardiovascular model simliaryly emphasised EtCO₂, non-invasive and arterial blood pressure, heart rate and previous episodes of respiratory or cardiovascular instability. The high weighting of EtCO₂ may again reflect the interdependent nature of respiratory and cardiovascular physiology, where CO2 is not only a marker of ventilation but also perfusion.\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e Furthermore, the presence of invasive blood pressure, available in ~ 36% of transport episodes, was associated with positive predictions, likely reflecting the fact that patients requiring arterial lines are more haemodynamically unstable and thus at higher risk of deterioration. Interestingly, diagnosis was more influential in negative predictions, suggesting the model learned that certain conditions confer lower deterioration risk. These insights highlight the models’ ability to provide interpretable outputs that align with clinical reasoning, a key factor in promoting clinician trust and potential real-world adoption .\u003c/p\u003e\u003cp\u003eA further key strength of our models is their computational efficiency and suitability for real-time deployment in resource-constrained settings. Both are lightweight and would be capable of running on edge devices, used by transport teams. Once loaded, both models can generate predictions for a given time window in well under one second on a standard laptop CPU, supporting real-time risk assessment without requiring high-end hardware. Integrated gradients, used for interpretability, is similarly efficient and can run in real time on the same device, providing patient-specific explanations alongside risk scores. Crucially, all computations can be performed locally, without relying on continuous server access; an advantage in transport environments where connectivity may be limited. Together, these features make real-world deployment in paediatric critical care transport both practical and feasible.\u003c/p\u003e\u003cp\u003eOne limitation is the availability and continuity of monitoring data: only 23.7% (1,519 of 6,471) of transports met inclusion criteria. However, included cases closely matched the overall cohort across key demographic and clinical variables. Second, although our models achieved strong discrimination with high AUROCs, their positive predictive values (PPVs) and F1-scores were modest; largely due to the low prevalence of adverse events (4–5% of windows). As a result, even with high specificity, false positives outnumber true positives, leading to lower PPV and F1-scores. This challenge is not unique to our study and is a recognised limitation in predictive modelling for rare but critical events. Similar findings have been reported in other early warning systems, where models predicting rare but critical deterioration events (such as ICU transfers, in-hospital cardiac arrests, or ward-based decompensation) achieved high AUROCs (\u0026gt; 0.85) but relatively low PPVs (typically ~ 10–15%), due to the low event prevalence. \u003csup\u003e\u003cspan additionalcitationids=\"CR21\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e–\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e Despite appearing suboptimal in absolute terms, such PPVs can still be clinically acceptable, particularly in the absence of alternative systems or when coupled with high sensitivity and fewer false alarms than traditional scoring methods. Further testing in prospective trials should include detailed examination of human factors and clinician preferences in managing the balance between sensitivity and false alarm rates.\u003c/p\u003e\u003cp\u003eLooking ahead, external validation is essential before real-world deployment. Testing on datasets from other paediatric transport services (both nationally and internationally) will be crucial to assess generalisability across different patient populations, transport systems, monitoring technologies, and clinical practices. Multicentre validation would also enable retraining or fine-tuning on more heterogeneous data, improving robustness and broader applicability. Following this, prospective deployment studies will be needed to evaluate real-world performance, clinical decision-making and ultimately, patient outcomes. Furthermore, the general framework presented here could be adapted to in-hospital PICU applications relating to risk of clinical deterioration. Here, similar models could serve as adjunctive tools to support continuous risk stratification and enable proactive intervention in children already receiving intensive care.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe present the first real-time machine learning models capable of predicting acute deterioration in critically ill children during interhospital transport using routinely collected high-frequency vital sign and clinical data. These lightweight, explainable models perform well and can run on edge devices, making them practical for resource-limited settings. Our findings suggest significant potential to enhance early intervention during transport.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eStudy approval\u003c/strong\u003e\u003cp\u003e The study was approved by Great Ormond Street Hospital's (GOSH) Research and Innovation Department, under ethical approvals for use of routine de-identified healthcare and operational hospital data (Research Database, NHS REC reference 21/LO/0646).\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eData Sources\u003c/strong\u003e\u003cp\u003eThis retrospective study analysed continuously monitored vital signs of critically ill children transported by the Children\u0026rsquo;s Acute Transport Service (CATS), a regional paediatric critical care team in North London, UK, between July 2016 and May 2021. Since 2016, CATS has used SwiftCare (Kinseed Limited, UK) to collect one data point per second vital signs, including heart rate, respiratory rate, blood pressure, oxygen saturation, and end-tidal carbon dioxide (EtCO2). Data collection starts at transfer initiation, with ambulance staff using a SwiftCare-enabled smartphone to connect wirelessly to a Philips Intellivue MP5 monitor, recording continuously until patient handover at the destination unit (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). These data were linked with deidentified electronic health records (EHRs), which included demographics, transport details, diagnoses, pre-transport interventions, intra-transport support, and time-series vital signs. A complete list of recorded parameters is provided in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. While most transport episodes involved unique patients, a small proportion may reflect repeat transports. Due to record anonymisation, we could not identify repeated episodes for the same patient. However, as each transport was clinically distinct, all episodes were treated as independent records.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eInclusion Criteria\u003c/strong\u003e\u003cp\u003eEpisodes were included if the patient was \u0026le;\u0026thinsp;18 years old with at least 30 minutes of recorded vital signs. Continuous data throughout transport was not mandated, acknowledging real-world interruptions such as sensor dropouts or new measurements (e.g., EtCO2 upon intubation during transport).\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eData Processing\u003c/strong\u003e\u003cp\u003eVital sign data, collected at one-second resolution (with the exception of non-invasive blood pressure, where the last recorded value was carried forward until a new measurement), underwent initial preprocessing. Physiologically implausible values were first removed from the vital sign data (heart rate\u0026thinsp;\u0026lt;\u0026thinsp;30 or \u0026gt;\u0026thinsp;300 bpm, EtCO₂ \u0026gt;15 kPa, blood pressure\u0026thinsp;\u0026lt;\u0026thinsp;5 mmHg, temperature\u0026thinsp;\u0026lt;\u0026thinsp;25\u0026deg;C or \u0026gt;\u0026thinsp;45\u0026deg;C). To account for age-related variability, all vital signs were standardised to improve comparability and enhance the reliability of model training. Respiratory rate, heart rate, and blood pressure were z-score normalised following methodology outlined in our prior work.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e SpO₂ was scaled using (SpO₂ \u0026minus; 97)/6, ETCO₂ using (ETCO₂ \u0026minus; 5.25)/1.5, and temperature using (Temp\u0026thinsp;\u0026minus;\u0026thinsp;36.75)/1.5. Missing values (data due real-world interruptions such as sensor dropouts) were mean-imputed (set to zero post-standardisation), and a corresponding missingness mask was generated for model input. Prior to imputation a missingness mask was created to track imputed values; this was used later as part of the feature set to train the models.\u003c/p\u003e\u003c/p\u003e\u003cp\u003eDemographic features were preprocessed for model compatibility. Age, weight, and the Paediatric Index of Mortality 3 (PIM3) were scaled to a 0\u0026ndash;1 range, with missing values imputed as \u0026minus;\u0026thinsp;1.\u003csup\u003e23\u003c/sup\u003e Sex was binary encoded, and ethnicity one-hot encoded using NHS Digital schema.\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e Referring and destination hospitals were also one-hot encoded, with units having\u0026thinsp;\u0026lt;\u0026thinsp;10 retrievals grouped as \"Other.\" One-hot encoding was applied to destination care area, pre-transport interventions, intra-transport respiratory and cardiovascular support, and pre-existing conditions. Team arrival time was used to label day vs night shifts. Primary diagnosis was encoded in two ways: via one-hot encoding into clinical categories (respiratory, cardiovascular, neurological, infection, gastroenterology, metabolic, trauma), and via 768-dimensional embeddings from BioClinicalBERT, pre-trained on MIMIC-III.\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e Embeddings were precomputed and cached in advance to avoid inference-time overhead. Both encoding strategies were tested during model development.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eLabel Creation\u003c/strong\u003e\u003cp\u003eOur dataset did not include timestamped, clinician-annotated deterioration labels. To address this, we implemented an automated, data-driven approach adapting Bollinger bands to continuous vital sign data, as detailed in our previous work.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e For each transport episode, one-minute mean-averaged data were used to compute an exponentially weighted moving average (EMA) and standard deviation (EWMSTD), generating patient-specific upper and lower thresholds that reflect each individual's evolving baseline. To prevent these dynamic bands from becoming infinitesimally narrow during periods of prolonged stability or low variability if a parameter\u0026rsquo;s EWMSTD was within 5% of its current EMA, a fixed threshold of \u0026plusmn;\u0026thinsp;5% of the current EMA was enforced. For oxygen saturation (SpO₂), due to its inherently lower variability, a stricter fixed boundary of \u0026plusmn;\u0026thinsp;2.5% of its current EMA was applied. Respiratory deteriorations were flagged when oxygen saturation (SpO₂) fell below either a fixed threshold of 94% or the dynamic lower bound (whichever was lower) alongside a concurrent abnormality in at least one other respiratory parameter (impedance pneumography derived respiratory rate, airway-derived respiratory rate, or EtCO₂). Cardiovascular deteriorations were identified by simultaneous deviations in heart rate and blood pressure. To reduce false positives from transient fluctuations or artefacts, events were required to persist for at least one minute and be flanked by five minutes of continuous data. These criteria were applied retrospectively to label minute-level or cumulative periods of respiratory and cardiovascular deterioration.\u003c/p\u003e\u003c/p\u003e\u003cp\u003eCrucially, our automated labelling method does not carry the inherent risk of a model \"rediscovering\" the labelling logic. This is because the model's prediction relies exclusively on preceding 15-minute vital sign data and historical context, not on the vital sign data within the 15-minute prediction window itself. This strict temporal separation ensures that the model cannot simply identify the conditions that trigger a band-based alert at the time of the event. Instead, the model is designed to learn and predict future deterioration based on subtle physiological shifts and patterns that precede a deterioration event, rather than merely re-identifying the criteria used to define the event retrospectively.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData Split\u003c/b\u003e: We applied a label-stratified split, allocating 80% of episodes (1,214) to training, 10% (150) to tuning, and 10% (155) to testing. The slight imbalance between tuning and test set sizes reflects the stratification process. Episodes were first grouped into four categories based on deterioration patterns: no deterioration, only respiratory deterioration, only cardiovascular deterioration, and both respiratory and cardiovascular deterioration. Each category was then split into 80/10/10 subsets, which were subsequently combined to form the final stratified dataset.\u003c/p\u003e\u003cp\u003e\u003cb\u003eWindow Creation\u003c/b\u003e: To facilitate model training and evaluation time series vital signs data was first discretised into 15-minute windows. This ensured that each period was used only once as a potential label window. For each window the model was tasked to use the data from the current window (\u003cem\u003et\u003c/em\u003e\u003csub\u003e\u0026minus;\u0026thinsp;14:59\u003c/sub\u003e - \u003cem\u003et\u003c/em\u003e\u003csub\u003e00:00\u003c/sub\u003e), to predict whether any adverse event would occur in the next window (\u003cem\u003et\u003c/em\u003e\u003csub\u003e00:01\u003c/sub\u003e - \u003cem\u003et\u003c/em\u003e\u003csub\u003e+\u0026thinsp;15:00\u003c/sub\u003e). Timestamped adverse events from the current window were also included as features. If data was available prior the current window (as would occur with data accumulation over the course of a long transport episode) additional context was provided as input to the model by summarizing data from \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u0026minus;\u0026thinsp;02:15:00\u003c/sub\u003e up to \u003cem\u003et\u003c/em\u003e\u003csub\u003e\u0026minus;\u0026thinsp;00:15:00\u003c/sub\u003e. This additional data was mean-averaged over 5-minute intervals and each interval was annotated with the occurrence of respiratory or cardiovascular events, further enriching the feature set. Strict controls were applied to prevent future information leakage, and all windows from a given episode remained within the same data split (train, tune, or test). Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e illustrates this windowing strategy.\u003c/p\u003e\u003cp\u003eWe chose a 15-minute prediction horizon and a 15-minute high-resolution input window due to pragmatic and clinical considerations. This future horizon offers clinically relevant lead time for proactive intervention without exceeding the model's predictive capability or providing insights that are not sufficiently actionable. The 15-minute input window effectively establishes a robust clinical baseline and captures meaningful physiological trends.12 Shorter windows (e.g., 10 minutes) might lack sufficient context, while longer ones (e.g., 20 minutes) would increase computational complexity and latency, hindering real-time deployment without substantial predictive gains. This duration also supports frequent predictions during typical transport times (median 113 minutes). Additionally, 120 minutes of averaged historical context (at 5-minute intervals) was included to offer a broader clinical trajectory while maintaining computational efficiency for edge device deployment.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eModel Development:\u003c/h3\u003e\n\u003cp\u003eWe developed two sets of models to forecast respiratory and cardiovascular deteriorations within a 15-minute horizon. For each, model complexity and feature richness were gradually increased to identify the minimal configuration needed to sustain high performance while minimising user input and computational load (see Supplementary Table\u0026nbsp;1).\u003c/p\u003e\u003cp\u003eWe used a transformer architecture for time-series analysis of vital signs, leveraging its superior ability to model long-range dependencies and its computational efficiency over recurrent approaches such as recurrent neural networks or long short-term memory networks.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e To optimize temporal encoding, transformer blocks were implemented with rotary positional embeddings.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e Transformer blocks were implemented with a decoder-only architecture to ensure data points could only attend to preceding data, thereby guaranteeing that predictions were based solely on historical information.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e Given the class imbalance (adverse events being rare) we applied class-weighted loss during training to penalize false negatives more heavily to maintain balanced performance across classes.\u003c/p\u003e\u003cp\u003eWe performed random search hyperparameter optimization over the training and tuning sets exploring parameters including learning rate, batch size, number of epochs, positional weighting, hidden layer dimensions, transformer heads and layers, batch normalization, dropout rates, L2 regularization, and max-norm constraints to identify the optimal training configuration. The best-performing model on the tuning set was then evaluated on the test set. This approach yielded separate models for predicting respiratory and cardiovascular events. While a combined multiclass model was considered, the non-exclusive nature of events and differing optimal architectures (e.g., hidden layer sizes, transformer depth, attention heads) made separate models more practical.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eModel Evaluation\u003c/strong\u003e\u003cp\u003eModel performance was evaluated on the holdout test set using AUROC, AUPRC, sensitivity (recall), specificity, PPV (precision), NPV, balanced accuracy, and F1-score. To enable consistent comparisons and reflect real-world clinical constraints such as alarm fatigue, a fixed decision threshold was selected using the tuning set; specifically, the threshold that capped the false positive rate at 20% (i.e., ensured at least 80% specificity). This threshold was then applied unchanged to the test set. It is important to note that threshold-dependent metrics (sensitivity, specificity, PPV, NPV, balanced accuracy, and F1-score) may vary if a different threshold is used, for example one that enforces a minimum sensitivity of 80%.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eModel Explainability\u003c/strong\u003e\u003cp\u003eTo understand the decision-making process of our models, we employed Integrated Gradients.\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e We focused on the absolute (magnitude) of each feature\u0026rsquo;s attribution, rather than its signed contribution. Because time-series features can fluctuate, sometimes pushing the prediction higher and sometimes lower, summing signed attributions could lead to misleading cancellations. By taking the absolute value, we preserved each feature\u0026rsquo;s overall contribution across time and avoided under-representing features whose positive and negative influences might otherwise cancel out.\u003c/p\u003e\u003c/p\u003e\u003cp\u003eWe applied this approach to both the best respiratory and cardiovascular models at individual- patient and cohort-wide levels. At the individual level, we identified the most influential features for specific predictions; at the cohort level, we examined which features mattered most on average, providing deeper insight into the models\u0026rsquo; overall behaviour.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eResearch Environment\u003c/strong\u003e\u003cp\u003eAll experiments were conducted within a secure Digital Research Environment provided by Aridhia Informatics Ltd in Glasgow, Scotland. The computational resources included an Intel\u0026reg; Xeon\u0026reg; Platinum 8272CL CPU with 64 GB of RAM; no GPU was utilized. Model development was performed using Python 3.12, leveraging the PyTorch framework (v2.6). Additional utilised libraries included Pandas, NumPy, X-transformers, Scikit-learn, Captum, Matplotlib, Seaborn, BioClinicalBert.\u003c/p\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData Availability:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe patient physiological data underlying the findings of this study subject to ethical and legal restrictions due to their sensitive patient information content. This dataset was obtained from routine clinical records held by the Children’s Acute Transport Service (CATS) and Great Ormond Street Hospital (GOSH). Our research use of this dataset was approved by GOSH Research and Innovation Division, under ethical approvals for use of routine de-identified healthcare and operational hospital data (Research Database, NHS REC reference 21/LO/0646). The authors did not receive any special privileges in accessing the data that other researchers would not have.\u003c/p\u003e\n\u003cp\u003eResearchers seeking access to this dataset for secondary analysis must obtain permission directly from the data custodians. Interested researchers should contact the Research and Innovation Division at Great Ormond Street Hospital for Children NHS Foundation Trust ([email protected]). Data access applications are subject to a formal review process, including confirmation of appropriate ethical approval and the establishment of a Data Sharing Agreement with the requesting institution.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll code and trained models associated with this project are available at https://github.com/MilanKapur1/paediatric_tranport_deterioration_prediction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe acknowledge the support from Kinseed for engineering the SwiftCare system used for the real-time extraction and secure server upload of high-frequency vital sign data directly from patient monitors during transport. We also thank,Great Ormond Street Hospital (GOSH) and the GOSH Digital Research Environment team for data curation.All research at Great Ormond Street Hospital NHS Foundation Trust and UCL Great Ormond Street Institute of Child Health is made possible by the NIHR Great Ormond Street Hospital Biomedical Research Centre.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMK is supported by an NIHR Academic Clinical Fellowship at University College London. KL is supported by UKRI Centre for Doctoral Training in AI-enabled healthcare systems. GD is supported by a UKRI Future Leaders Fellowship [MR/T041285/1]. \u0026nbsp;PR is in receipt of grant support from the National Institute of Health Research, Rosetrees Trust and BMA Foundation. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health and Social Care.\u0026nbsp;The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGD reports speaker honoraria from Vertex Pharmaceuticals and Chiesi Ltd, and advisory board and clinical trial leadership roles with Vertex, unrelated to the current manuscript. PR reports travel support for conference attendance from Fisher and Paykel Healthcare Limited. All other authors declare no conflicts of interest.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eM.K. and K.L. contributed equally as joint first authors. G.D. and P.R. contributed equally as joint senior authors. M.K., K.L., G.D., and P.R. conceived the study. P.K., P.R., Z.H. and J.B. were responsible for data acquisition and pre-processing. M.K. developed the models and performed the analysis. M.K., K.L., A.B., G.D., and P.R. contributed to the interpretation of the results. M.K. and K.L. wrote the manuscript. All authors reviewed and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePaediatric Intensive Care Audit Network (PICANet). \u003cem\u003eNational Paediatric Critical Care Audit State of the Nation Report 2024\u003c/em\u003e. https://www.picanet.org.uk/annual-reporting-and-publications/ (2024).\u003c/li\u003e\n\u003cli\u003eOrr, R. A. \u003cem\u003eet al.\u003c/em\u003e Pediatric specialized transport teams are associated with improved outcomes. \u003cem\u003ePediatrics\u003c/em\u003e \u003cstrong\u003e124\u003c/strong\u003e, 40\u0026ndash;48 (2009).\u003c/li\u003e\n\u003cli\u003eSingh, J. M., Gunz, A. C., Dhanani, S., Aghari, M. \u0026amp; MacDonald, R. D. Frequency, Composition, and Predictors of In-Transit Critical Events During Pediatric Critical Care Transport*. \u003cem\u003ePediatric Critical Care Medicine\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 984 (2016).\u003c/li\u003e\n\u003cli\u003eHaydar, B. \u003cem\u003eet al.\u003c/em\u003e Adverse Events During Intrahospital Transport of Critically Ill Children: A Systematic Review. \u003cem\u003eAnesthesia \u0026amp; Analgesia\u003c/em\u003e \u003cstrong\u003e131\u003c/strong\u003e, 1135 (2020).\u003c/li\u003e\n\u003cli\u003eHuo, Z. \u003cem\u003eet al.\u003c/em\u003e Distribution and trajectory of vital signs from high-frequency continuous monitoring during pediatric critical care transport. \u003cem\u003eIntensive Care Med. Paediatr. Neonatal\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, 18 (2023).\u003c/li\u003e\n\u003cli\u003eFleming, S. \u003cem\u003eet al.\u003c/em\u003e Normal ranges of heart rate and respiratory rate in children from birth to 18 years of age: a systematic review of observational studies. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e377\u003c/strong\u003e, 1011\u0026ndash;1018 (2011).\u003c/li\u003e\n\u003cli\u003eBonafide, C. P. \u003cem\u003eet al.\u003c/em\u003e Development of heart and respiratory rate percentile curves for hospitalized children. \u003cem\u003ePediatrics\u003c/em\u003e \u003cstrong\u003e131\u003c/strong\u003e, e1150-1157 (2013).\u003c/li\u003e\n\u003cli\u003eNielsen, V. M. L., Kl\u0026oslash;jg\u0026aring;rd, T., Bruun, H., S\u0026oslash;vs\u0026oslash;, M. B. \u0026amp; Christensen, E. F. Progression of vital signs during ambulance transport categorised by a paediatric triage model: a population-based historical cohort study. \u003cem\u003eBMJ Open\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, e042401 (2020).\u003c/li\u003e\n\u003cli\u003eKapur, M. \u003cem\u003eet al.\u003c/em\u003e Identification of physiological adverse events using continuous vital signs monitoring during paediatric critical care transport: a novel data-driven approach. 2025.03.11.25323742 Preprint at https://doi.org/10.1101/2025.03.11.25323742 (2025).\u003c/li\u003e\n\u003cli\u003eShamout, F. E. \u003cem\u003eet al.\u003c/em\u003e An artificial intelligence system for predicting the deterioration of COVID-19 patients in the emergency department. \u003cem\u003enpj Digit. Med.\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 1\u0026ndash;11 (2021).\u003c/li\u003e\n\u003cli\u003eHuo, Z. \u003cem\u003eet al.\u003c/em\u003e Dynamic mortality prediction in critically Ill children during interhospital transports to PICUs using explainable AI. \u003cem\u003enpj Digit. Med.\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 1\u0026ndash;11 (2025).\u003c/li\u003e\n\u003cli\u003eSundrani, S. \u003cem\u003eet al.\u003c/em\u003e Predicting patient decompensation from continuous physiologic monitoring in the emergency department. \u003cem\u003enpj Digit. Med.\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 1\u0026ndash;10 (2023).\u003c/li\u003e\n\u003cli\u003eJeon, Y., Kim, Y. S., Jang, W., Park, J. D. \u0026amp; Lee, B. Development of a deep learning model that predicts critical events of pediatric patients admitted to general wards. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 4707 (2024).\u003c/li\u003e\n\u003cli\u003eMayampurath, A. \u003cem\u003eet al.\u003c/em\u003e A Vital Sign-Based Model to Predict Clinical Deterioration in Hospitalized Children. \u003cem\u003ePediatr Crit Care Med\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 820\u0026ndash;826 (2020).\u003c/li\u003e\n\u003cli\u003eRust, L. O. H. \u003cem\u003eet al.\u003c/em\u003e The Deterioration Risk Index: Developing and Piloting a Machine Learning Algorithm to Reduce Pediatric Inpatient Deterioration. \u003cem\u003ePediatr Crit Care Med\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 322\u0026ndash;333 (2023).\u003c/li\u003e\n\u003cli\u003eEarly Risk Prediction of Pediatric Cardiac Arrest from Electronic Health Records via Multimodal Fused Transformer. https://arxiv.org/html/2502.07158v2.\u003c/li\u003e\n\u003cli\u003eAlsentzer, E. \u003cem\u003eet al.\u003c/em\u003e Publicly Available Clinical BERT Embeddings. in \u003cem\u003eProceedings of the 2nd Clinical Natural Language Processing Workshop\u003c/em\u003e (eds. Rumshisky, A., Roberts, K., Bethard, S. \u0026amp; Naumann, T.) 72\u0026ndash;78 (Association for Computational Linguistics, Minneapolis, Minnesota, USA, 2019). doi:10.18653/v1/W19-1909.\u003c/li\u003e\n\u003cli\u003eSundararajan, M., Taly, A. \u0026amp; Yan, Q. Axiomatic Attribution for Deep Networks. Preprint at https://doi.org/10.48550/arXiv.1703.01365 (2017).\u003c/li\u003e\n\u003cli\u003eGavelli, F., Teboul, J.-L. \u0026amp; Monnet, X. How can CO2-derived indices guide resuscitation in critically ill patients? \u003cem\u003eJournal of Thoracic Disease\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, (2019).\u003c/li\u003e\n\u003cli\u003eSteitz, B. D. \u003cem\u003eet al.\u003c/em\u003e Development and Validation of a Machine Learning Algorithm Using Clinical Pages to Predict Imminent Clinical Deterioration. \u003cem\u003eJ Gen Intern Med\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, 27\u0026ndash;35 (2024).\u003c/li\u003e\n\u003cli\u003eKwon, J.-M., Lee, Y., Lee, Y., Lee, S. \u0026amp; Park, J. An Algorithm Based on Deep Learning for Predicting In-Hospital Cardiac Arrest. \u003cem\u003eJ Am Heart Assoc\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, e008678 (2018).\u003c/li\u003e\n\u003cli\u003eChurpek, M. M. \u003cem\u003eet al.\u003c/em\u003e Multicenter Development and Prospective Validation of eCARTv5: A Gradient-Boosted Machine-Learning Early Warning Score. \u003cem\u003eCritical Care Explorations\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, e1232 (2025).\u003c/li\u003e\n\u003cli\u003eStraney, L. \u003cem\u003eet al.\u003c/em\u003e Paediatric index of mortality 3: an updated model for predicting mortality in pediatric intensive care*. \u003cem\u003ePediatr Crit Care Med\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 673\u0026ndash;681 (2013).\u003c/li\u003e\n\u003cli\u003eNHS Digital. Data quality of protected characteristics and other vulnerable groups: Ethnicity. \u003cem\u003eNHS England Digital\u003c/em\u003e https://digital.nhs.uk/data-and-information/data-collections-and-data-sets/data-sets/mental-health-services-data-set/submit-data/data-quality-of-protected-characteristics-and-other-vulnerable-groups/ethnicity.\u003c/li\u003e\n\u003cli\u003eVaswani, A. \u003cem\u003eet al.\u003c/em\u003e Attention Is All You Need. Preprint at https://doi.org/10.48550/arXiv.1706.03762 (2023).\u003c/li\u003e\n\u003cli\u003eSu, J. \u003cem\u003eet al.\u003c/em\u003e RoFormer: Enhanced transformer with Rotary Position Embedding. \u003cem\u003eNeurocomput.\u003c/em\u003e \u003cstrong\u003e568\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7319827/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7319827/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eInterhospital transport of critically ill children carries inherent risks, including unexpected respiratory and cardiovascular deterioration. Early warning of impending deterioration may allow physicians to avert a more serious issue. We developed and evaluated lightweight, explainable machine learning models to forecast adverse physiological events up to 15 minutes in advance using continuously streamed vital signs and clinical data. Models were trained and evaluated on 1,519 transports conducted by a specialist paediatric critical care team in London (2016\u0026ndash;2021). Transformer-based models incorporating vital sign time-series and vector-embedded diagnoses outperformed simpler models, achieving AUROC scores of 0.851 for respiratory and 0.792 for cardiovascular deterioration. Model interpretability was provided using Integrated Gradients, revealing alignment with clinical reasoning. Designed for deployment on edge devices, these models offer real-time, interpretable risk predictions in resource-limited transport settings. These results demonstrate that real-time, explainable machine learning models can accurately predict deterioration during interhospital paediatric transport using routinely collected data, supporting their potential role in enhancing early clinical intervention.\u003c/p\u003e","manuscriptTitle":"Real-time prediction of cardiorespiratory deterioration during paediatric critical care transport using interpretable machine learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-03 08:45:03","doi":"10.21203/rs.3.rs-7319827/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"65a93ca8-e256-44f9-a8ae-ac70a5e7a737","owner":[],"postedDate":"September 3rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":53762704,"name":"Health sciences/Cardiology"},{"id":53762705,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":53762706,"name":"Health sciences/Diseases"},{"id":53762707,"name":"Health sciences/Health care"},{"id":53762708,"name":"Physical sciences/Mathematics and computing"},{"id":53762709,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2025-11-01T18:23:20+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-03 08:45:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7319827","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7319827","identity":"rs-7319827","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}