Development and Validation of a Model to Predict Pediatric Septic Shock Using Data Known Two Hours after Hospital Arrival

Abstract

Objective:

To use Electronic Health Record (EHR) data from the first two hours of care to derive and validate a model to predict hypotensive septic shock in children with infection.

Design:

Derivation-validation study using an existing registry

Setting:

Six emergency care sites within a regional pediatric healthcare system. Three datasets of unique visits were designated:

1. training set (5 sites, 4/1/13-12/31/16)

2. temporal test set (5 sites, 1/1/17-6/30/18)

3. geographic test set (6^th site, 4/1/13-6/30/18)

Patients:

Patients in whom clinicians were concerned about serious infection from 60 days-17 years were included; those with septic shock in the first two hours were excluded. There were 2318 included visits; 197 developed septic shock (8.5%).

Interventions:

Lasso with tenfold cross-validation was used for variable selection; logistic regression was then used to construct a model from those variables in the training set. Variables were derived from EHR data known in the first two hours, including vital signs, medical history, demographics, laboratory information. Test characteristics at two thresholds were evaluated: 1) optimizing sensitivity and specificity, 2) set to 90% sensitivity.

Measurements and Main Results:

Septic shock was defined as systolic hypotension and vasoactive use or ≥30 ml/kg isotonic crystalloid administration in the first 24 hours. A model was created using twenty predictors, with an area under the receiver operating curve in the training set of 0.85 (0.82-0.88); 0.83 [0.78-0.89] in the temporal test set; 0.83 [0.60-1.00] in the geographic test set. Sensitivity and specificity varied based on cutpoint; when sensitivity in the training set was set to 90% (83%, 94%), specificity was 62% (60%, 65%).

Conclusions:

This model predicted risk of septic shock in children with suspected infection 2 hours after arrival, a critical timepoint for emergent treatment and transfer decisions. Varied cutpoints could be used to customize sensitivity to clinical context.

Introduction

Septic shock is a leading cause of death in children. Early recognition and treatment can prevent morbidity and mortality, yet studies demonstrate that many children do not receive early sepsis care that is guideline-concordant.^1-4 Failure to deliver early resuscitation including timely antibiotics and intravenous (IV) fluid in the emergency department (ED), and identify children with sepsis who require intensive care (ICU), can lead to organ dysfunction and intractable shock.^{1, 3} Few children who present for emergency care with fever will deteriorate to septic shock, and identification of this cohort is difficult before hypotension develops.^{1, 5} Most of these children will receive emergency care in hospitals without pediatric ICUs, highlighting the importance of early recognition to facilitate prompt resuscitation and transfer.^6-8

Models exist to predict death in children with sepsis, however, there is a need to predict the risk of shock as early as possible to prevent death.^9-11 Diagnostic aides have been developed to identify children with sepsis, but there are few models for septic shock in children outside ICU settings.^12-14

We previously developed a model to predict septic shock among children in whom clinicians were concerned for sepsis, using only data routinely available in the Electronic Health Record (EHR) at the time of ED arrival.¹⁵ This model provided useful initial risk-stratification immediately upon arrival. However, additional data such as laboratory results and multiple vital sign measurements become available shortly after arrival that could improve initial risk stratification. We hypothesized that a model using data from the first two hours of care would enable reliable prediction of the probability of septic shock, leading to earlier recognition, treatment, disposition and improved outcome. Thus, we sought to develop and validate a model of risk for hypotensive shock using EHR data that is available in the first two hours of emergency care.

Materials and Methods

Objective

Our objectives were to 1) develop a model of the risk of hypotensive septic shock among patients in whom sepsis was suspected by clinicians, using variables available in the EHR in the first two hours of care, and 2) validate the model by testing model performance in two distinct test sets.

Study Design

This was a derivation and validation study of a predictive model, following the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD 2) guidelines, and approved by the Colorado Multiple Institutions Review Board with a waiver of informed consent.¹⁶

Setting

The setting was six emergency care sites within a regional children’s healthcare system including pediatric ED, Urgent Care (UC) and dual-track ED/UC sites with >150,000 annual visits combined. Three sites were within a hospital with inpatient pediatric beds, and only one site included a pediatric intensive care unit (PICU) on site. All sites shared the same electronic health record and formulary, however there were differences among sites in laboratory, radiology, and medication availability based on the size, acuity, and ED or UC designation. There were approximately 150 attending staff among all sites who were members of the same academic department, with varied training, including pediatric emergency physicians, general pediatricians, nurse practitioners and physician assistants. Three of the six sites included staffing with subspecialty pediatric emergency physicians. A common sepsis pathway existed at all sites, including an expedited evaluation tier for patients with suspected infection, and a higher-acuity resuscitation tier for patients with confirmed severe sepsis or septic shock.

This observational study used a sepsis registry extracted from the EHR and loaded monthly into REDCap electronic data capture tools.¹⁷ Three unique datasets were created from the registry, based on two different non-random data splits, which is considered a TRIPOD type 2b, or pseudo-external, validation strategy. The datasets were:

1. A training set, from 5 sites, April 1, 2013-December 31, 2016

2. A temporal holdout test set from the 5 original sites, January 1, 2017-June 30, 2018

3. A geographic holdout test set from a small pediatric community-based ED site, April 1, 2013-June 30, 2018

The temporal test set was intended as the primary test set, while the geographic test set was designed to be a small, preliminary evaluation of the performance of the model in a non-tertiary ED setting.

Participants

Included in the registry were patients who presented to one of the sites for emergency care in whom a physician suspected infection, with a high-risk condition or signs of decreased mental status or perfusion. These patients were identified in the EHR by presence of a sepsis orderset or activation event. In addition, a monthly standardized chart review of patients admitted to the ICU within 24 hours of emergency care was conducted to identify and include missed cases of severe sepsis and septic shock in the registry. This chart review was conducted for institutional quality improvement purposes prior to this study, and was carried out by two physicians and three nurses monthly. Inter-rater reliability for identifying missed severe sepsis cases was tested every 6 months with a kappa >0.8 maintained.

Patients were excluded who were <60 days or ≥18 years-old, had >80% of all EHR data missing indicating a registration or EHR downtime error, or were transferred to a hospital outside of the study sites. Patients with multiple visits in the dataset were included, however, only one randomly-chosen visit per patient was included, in order to avoid the potential for a few patients who frequently experienced the outcome to overly influence patient-level categorical variable selection. Patients were excluded who had the outcome of hypotensive septic shock, as described below, during the first two hours after arrival.

Outcome Measures

Hypotensive septic shock was the outcome measure, and was defined as the presence of 1) systolic hypotension per consensus sepsis age-based definitions within 24 hours after ED arrival, AND 2) either administration of vasoactive medication or ≥30 mL/kg of isotonic crystalloids intravenously (1500 mL in patients weighing ≥50 kg), within the first 24 hours.¹⁸ This outcome included an objective physiologic finding along with resuscitative treatment within a meaningful timeframe. It represented disease severe enough and temporally proximate to the model, such that early detection and appropriate early resuscitative care could improve outcomes.^1,2

Outcomes were assessed using vital signs and medication administration records, and were classified prior to construction of the model.

Data and Analysis:

The dataset was exported to SAS 9.4 (SAS, Cary, NC) for data management and preliminary analysis. R 3.5.2 (R Foundation for Statistical Computing, Vienna, Austria) was used for additional analysis.¹⁹ We first assessed data missingness. Fewer than 5% of visits had predictor variables missing (excluding laboratory variables), and we excluded these visits (SDC 1).

Model Derivation (Training Set)

We conducted all of the following model derivation procedures in the training set only. We initially considered available predictor variables, described in detail in SDC 1. These included patient demographics, chronic conditions,²⁰ prior visit data, ED visit data such as arrival mode and triage category, vital signs, and laboratory data. Only variables available in the EHR during the first two hours of care were considered; data entered about the ED visit later were not included. Where patient information such as number of prior visits or diagnoses were considered, these were ascertained from prior visits available in the EHR at the time of hospital arrival. Initial, last, minimum and maximum vital signs from the first two hours were available.

We conducted initial exploratory analyses in the training set prior to fitting the models, in order to reduce the number of predictors considered, and to use piecewise or polynomial transformations where appropriate. To create a predictive model, we used lasso, a regularization technique that uses a penalty term to shrink the regression coefficients and perform model selection. The model derivation procedures are described in detail in the Supplemental Statistical Methods (SDC 1)

To internally validate the model, we evaluated the performance of the model in the training set, calculating the area under the receiver operating characteristic (AUROC) curve and calibration plots.²¹ Although a user could chose a classification threshold based on the needs of a specific clinical setting, we calculated two classification thresholds of particular potential usefulness: 1) We used Youden’s J statistic to identify the threshold that optimized sensitivity and specificity; and 2) chose the threshold that produced 90% sensitivity, due to the importance of prioritizing sensitivity in sepsis. At both thresholds, we evaluated sensitivity, specificity, positive and negative predictive value.

Model Validation (Test Sets)

After the final model was chosen, we then evaluated the AUROC, calibration plots, and test characteristics in the test sets. We constructed precision-recall curves, which plot sensitivity vs. positive predictive value, and are considered a useful way to assess a model in which the result will not be inflated by a large number of true negatives.²². In a secondary analysis designed to more closely mimic real-world conditions, all visits from patients with multiple visits were left in the test sets, and AUROC and calibration evaluated in these sets.

Results

Of 4686 patients, 2495 met inclusion criteria for initial analysis; after applying exclusions, 2318 patients were analyzed (Figure 1). Patient characteristics are described in Table 1. The outcome, hypotensive septic shock, occurred between two hours and 24 hours after arrival in 197 (8.5%) of patients, and the timing of hypotension onset is shown in SDC 1.

Figure 1.

Study flow diagram

Table 1.

Study population

Characteristic	Training Set (n=1594)	Temporal Test Set (n=664)	Geographic Test Set (n=60)
Primary outcome, hypotensive septic shock, No. (%)	137 (9%)	55 (8%)	5 (8%)
Male, No. (%)	883 (55%)	343 (52%)	31 (52%)
Age in years, median (IQR)	5.9 (10)	5.2 (10)	4.9 (10)
Private payer health insurance, No. (%)	686 (43%)	278 (42%)	37 (62%)
Chronic complex conditiona, No. (%)	943 (59%)	400 (60%)	13 (22%)
Central venous line present on arrival, No. (%)	353 (22%)	129 (19%)	1 (2%)
Arrived via Emergency Medical Services, No. (%)	201 (12%)	76 (11%)	10 (17%)
Hospitalized in the last year, No. (%)	824 (52%)	365 (55%)	11 (18%)
Not identified as potential sepsis in ED, No. (%)	152 (9.5%)	5 (0.8%)	14 (23%)
ED Disposition
Home, No. (%)	433 (27%)	201 (30%)	16 (27%)
Inpatient ward or Operating Room, No. (%)	816 (51%)	330 (50%)	27 (45%)
Intensive Care Unit, No. (%)	344 (22%)	133 (20%)	17 (28%)
Deceased	1 (0%)	0 (0%)	0 (0%)
Hospital course
Vasoactive agent, No. (%)	87 (7%)	31 (7%)	4 (9%)
Positive-pressure ventilation, No. (%)	124 (11%)	52 (11%)	4 (9%)
Hematologic organ dysfunction, No. (%)b	196 (17%)	91 (20%)	0 (0%)
Hepatic organ dysfunction, No. (%)b	130 (11%)	54 (12%)	1 (2%)
Renal organ dysfunction, No. (%)b	30 (3%)	15 (3%)	1 (2%)
Median (IQR), hospital length of stay among admitted patients, days	3.7 (4.9)	3.7 (4.8)	3 (4.6)
Median (IQR), Intensive Care Unit length of stay among admitted patients, days	1.6 (2.5)	1.6 (3.9)	2.0 (5.7)
30-Day In-Hospital Mortality, No. (%)	15 (0.9%)	3 (0.5%)	0 (0%)

^aChronic complex conditions were defined using Feudtner ICD-9 and ICD-10 codes²⁰

^bOrgan dysfunction defined by pediatric severe sepsis definitions from Goldstein et al¹⁸

After initial exploratory analysis of the predictor variables, 64 parameters were considered for inclusion in the model (SDC 1). The plot of cross validation error compared to the number of parameters in the model demonstrated that mean cross validation error did not substantially decline after 20 parameters (SDC 1).

We developed a model using 20 parameters (Table 2). The AUROC curve in the training set was 0.85 (0.82-0.88), in the temporal test set it was 0.83 (0.78-0.89), and in the geographic test set it was 0.83 (0.60-1.00) (Figure 2). Figure 3 shows the calibration curves with good calibration in the training set and temporal test set. In the geographic test set, the smallest dataset with 60 cases and 5 outcomes, calibration was worse with the model overestimating risk. A secondary analysis in which all patients with repeated visits were left in the test sets yielded similar AUROC and calibration curves (SDC 1).

Table 2.

Final model for the prediction of septic shock among patients in whom ED clinicians were concerned for sepsis using all data available during the first two hours after ED arrival. The model is a linear predictor that estimates the log odds of septic shock, using the sum of the intercept and the predictors multiplied by their coefficient. To transform the risk of septic shock to a probability, use the transformation e^xB/(1+e^xB).

Predictive variable	log odds increase in septic shock probability with each unit increase in predictor
Intercept	1.8819
First systolic blood pressure, mmHg	−0.0357
Last systolic blood pressure, mmHg	−0.0174
Last diastolic blood pressure <69 mmHga	−0.0356
Age, years * heart rate, beat per minute	0.0019
Age, years * shock index, beat per minute/mmHgb *	−0.0987
Age, years * respiratory rate, breaths per minute	0.0018
Oncological comorbidity	−0.7688
Hospitalized in the last year	−0.5398
Indwelling central line present on arrival	−0.4467
Venous lactate normal (<2 mmol/L)	−0.4927
Albumin low (<3 g/dL)	0.8530
ALT high (<3 years: >80 U/L, >3 years: >70 U/L)	0.5477
Bilirubin measured	0.3845
BUN*age high (>66 mg/dL*years)	0.2757
Creatinine high   <3 years: >0.4 mg/dL   3 to <6 years: >0.53 mg/dL   6 to <8 years: >0.66 mg/dL   8 to <11 years: >0.75 mg/dL   11 to <14 years: >0.85 mg/dL   Male 14-18 years: >1.1 mg/dL   Female 14-18 years: >1.03 mg/dL	0.5784
Glucose high (>120 mg/dL)	0.2849
Absolute neutrophil count low (<500/uL)	−0.3149
Bands high (>5/uL)	0.7884
Bands normal (≤5/uL)	−0.1783
Hemoglobin abnormal (<10 g/dL or >16 g/dL)	−1.2005

^aDiastolic blood pressure was modeled in piecewise splines. Values ≥69 mmHg, should be set to 0 in the model; values <69 mmHg, set to DBP-69.

^bShock index = HR/SBP

Figure 2.

Receiver operating characteristic curves in the training and test sets

Figure 3.

Calibration plots

a) Training set comparison of predicted proportion of patients with shock versus observed proportion of patients with shock, within deciles of risk. Calibration slope was 1.037, intercept was −0.003.

b) Temporal test set comparison of predicted proportion of patients with shock versus observed proportion of patients with shock, within deciles of risk. Calibration slope was 1.097, intercept was 0.004.

c) Geographic test set comparison of predicted proportion of patients with shock versus observed proportion of patients with shock, within tertiles of risk. Tertiles were used because the numbers were small in this dataset. Calibration slope was 0.699, intercept was −0.009.

Precision-recall curves (SDC 1) demonstrated that as sensitivity (recall) increased, the positive predictive value (precision), decreased, but the model retained predictive value above baseline population rate.

The threshold for classifying patients that optimized both sensitivity and specificity in the training set was a predicted probability of hypotensive shock of 9%, which yielded 79 (71-85)% sensitivity and 78 (75-80)% specificity in the training set. The threshold that yielded 90% sensitivity in the training set was a predicted probability of hypotensive shock of 5%. Test characteristics at each of these thresholds in all three datasets are shown in Table 3.

Table 3.

Test characteristics of the model in training and test sets. Two thresholds for classification of high-risk patients were derived in the training set: a threshold designed to optimize both sensitivity and specificity (Youden’s J), and a threshold designed to produce 90% sensitivity.

Dataset	Predicted Risk Threshold	Sensitivity (95% Confidence Intervals)	Specificity (95% Confidence Intervals)	Positive Predictive Value (95% Confidence Intervals)	Negative Predictive Value (95% Confidence Intervals)
Training	0.09 (Youden’s)	0.79 (0.71, 0.85)	0.78 (0.75, 0.80)	0.25 (0.21, 0.29)	0.98 (0.96, 0.98)
Temporal Test	0.09 (Youden’s)	0.69 (0.55, 0.81)	0.82 (0.79, 0.85)	0.26 (0.19, 0.34)	0.97 (0.95, 0.98)
Geographic Test	0.09 (Youden’s)	0.80 (0.28, 0.99)	0.64 (0.5, 0.76)	0.17 (0.05, 0.37)	0.97 (0.85, 1.00)
Training	0.05 (90% Sensitivity)	0.90 (0.83, 0.94)	0.62 (0.6, 0.65)	0.18 (0.15, 0.21)	0.98 (0.97, 0.99)
Temporal Test	0.05 (90% Sensitivity)	0.84 (0.71, 0.92)	0.65 (0.61, 0.69)	0.18 (0.13, 0.23)	0.98 (0.96, 0.99)
Geographic Test	0.05 (90% Sensitivity)	0.80 (0.28, 0.99)	0.40 (0.27, 0.54)	0.11 (0.03, 0.25)	0.96 (0.78, 1.00)

A secondary model was generated with missing laboratory values imputed rather than considered as categorically missing. This model had similar performance in the test sets, with a lower AUROC (SDC 1).

Discussion

This study derived and validated a model to predict septic shock in children in the first 24 hours of care, using clinical and laboratory data routinely available in the EHR two hours after arrival in emergency departments. The model discriminated well, with an AUROC of 0.83-0.85, and demonstrated excellent calibration (Figure 3).

This model was designed for potential use after a clinician had already indicated concern that a patient may have sepsis in the first hours of emergency care, in order to highlight the highest risk patients. Failure of early escalation of therapy contributes to preventable morbidity from sepsis in children.⁴ In one study, even at the time community emergency physicians had initiated transfer of a patient to a pediatric ICU, they had not recognized that the patient had shock, nor had they performed appropriate resuscitation.¹ Thus, this model is relevant to a crucial timepoint in the diagnosis and treatment of sepsis which may lead to improved outcomes.

Our previous research created a model that predicted risk of shock using only clinical data available upon ED arrival, and although it had good discrimination (AUROC of 0.75-0.87), the present model, takes advantage of the additional data obtained in the next two hours to improve on this model.¹⁵ Children with sepsis can deteriorate in unpredictable fashion, thus in suspected cases the two-hour model is ideally suited to be of benefit when used in conjunction with the clinician’s diagnosis and the arrival model to maximize all opportunities for early identification. This model was designed for implementation in an EHR. A strength is its use of real-time data obtained during clinical practice that already exists in the EHR database, rather than predictor variables obtained under research conditions which may not mimic clinical reality.

The threshold for considering patients high-risk can be customized to the clinical context. The precision-recall curves demonstrate the tradeoff between sensitivity (recall) and precision (positive predictive value) [SDC 1]. A sensitivity of at least 90% for the high-risk condition of sepsis would be desirable in emergency settings, yielding a positive predictive value of 18%, which corresponds to a Number Needed to Alert of approximately 5 [Table 3]. This means that an alert based on this model would identify 5 patients for every 1 who develops shock. There may be settings, particularly those with less available pediatric critical care expertise, where a higher sensitivity may be worth the tradeoff of a higher NNA and associated potential alert fatigue.

Dewan and Sanchez-Pinto have proposed using the NNA to inform the clinical actions taken after an alert is triggered.²³ They suggest that patients with an NNA of 5-10 should receive bedside evaluation, and patients with an NNA <5 receive confirmatory testing, escalation of care, and early therapeutic intervention if warranted. Early identification of organ dysfunction, clinical and laboratory-based, has been emphasized in recent pediatric sepsis algorithms, and bedside evaluation and laboratory testing to complete organ dysfunction evaluation would be warranted in patients identified at the NNA of 5.²⁴ Escalation of care to use local protocolized sepsis care bundles with close monitoring and examination would be additional appropriate actions at this sensitivity and NNA threshold. Treatment decisions requiring more individual nuance and carrying potentially greater risk, such as the decision to initiate transportation to a pediatric center, would likely still require case-by-case clinical decision making. This model can bring high risk patients to a physician’s attention, but itself cannot prescribe all appropriate treatments at this sensitivity threshold.

There are few direct comparisons in the literature, because of the unique timepoint and outcome studied. The majority of predictive models in pediatric sepsis have predicted mortality in hospitalized patients, while this model is designed to predict shock earlier, in a heterogeneous cohort of children with suspected sepsis. Existing models have excellent AUROC, ranging from 0.71 to 0.88 in patients already admitted to hospital or ICU, and rely on data obtained over many hours of care.^{10, 11} The Paediatric Emergency Triage (PET) score predicted the risk of death in children in Africa with serious febrile illness using only clinical variables to produce an AUROC of 0.82 in derivation and 0.77 and 0.86 in validation datasets.⁹ However prediction models are context dependent, and differences in population, resources and mortality rate in the United States compared to Africa underlines the importance of developing a model for the more proximate outcome of shock in US pediatric emergency settings.

Other important comparators are diagnostic tools for pediatric sepsis in the emergency setting. These tools are point-based scores designed by expert opinion and modified through quality improvement cycles, but are not models, and thus AUROC has not been reported.^{12-14, 25} These tools were designed using an outcome of clinical treatment on a sepsis pathway, or were an operationalization of a tool to identify patients meeting the severe sepsis definition. Thus, these existing pediatric sepsis tools have utility in diagnosis, but were not designed to be predictive.

Unmeasured variables are a concern in EHR data. Few visits (1.7%, Figure 1 and SDC 1) were missing a vital sign or categorical non-laboratory predictor. For this study, visits with a missing vital sign were excluded; in a real-life implementation, this limitation could be addressed in the design of a decision-support tool that might prompt the clinicians to obtain the missing data.

A unique aspect of our model was the approach to unmeasured laboratory variables. Two hours after ED arrival, laboratory results are not available in all patients. Typically, in mortality prediction models in sepsis, unmeasured variables have been considered normal.^{10, 11, 26} In an inpatient setting, where the majority of laboratory values included in such models are typically measured, this assumption may be a reasonable approach. However, the earlier that a prediction is made in the course of care, the more that this assumption may become suspect. In this model, we considered categorical normal or abnormal values, as well as the value of “unreported.” This allowed the model to incorporate laboratory results, as well as the predictive information that exists in the decision not to perform a laboratory test. In settings with testing patterns expected to differ from the study sites, the alternative median imputation model might generalize better. In this model, median values are substituted for missing laboratory values, and it demonstrated similar discrimination and calibration (SDC 1).

This model depended upon a clinician indicating an initial concern for sepsis, and was not designed to be a sniffer, designed to independently detect a condition among all patients. It will be important to test the generalizability of clinician suspicion for sepsis and its impact on this model. A potential advantage of this design is that it has the potential to deliver decision support to clinicians when they are unsure of the diagnosis, when early therapy may improve outcomes. In implementation studies of predictive decision support for sepsis in adults, investigators noted that alert fatigue and decision support that was redundant with clinical assessment were contributing factors when desired outcomes were not achieved.^{27, 28} This model sought to avoid those pitfalls by being applied after an initial clinical impression of possible sepsis, with the inherent limitation that it would not find a patient completely missed by a clinician. Thus, the dataset included many patients with risk factors which prompted clinicians to consider sepsis, such as an oncologic condition, which were negatively associated with the outcome of shock because the majority of such patients did not develop shock. This model was derived using observational data, and there may have been patients whose shock was averted by clinically-appropriate care received. In addition, it is subject to possible misclassification bias if patients discharged from the ED presented to a non-study hospital. Thus, this model should not be used to rule-out or de-escalate care.

There are limitations to the generalizability of these results. Although this study included pediatric emergency care settings of various sizes with multiple provider types, these sites were all within the same care network. The use of test sets with non-random splits is considered a pseudo-external validation; although more robust than internal validation, the model should in the future be tested in truly external data. The geographic test set was small, with only 60 patients and 5 outcomes, limiting the precision of results from this dataset. This smaller, community-based site should be considered only a pilot demonstration that this model is promising in a non-tertiary emergency care setting, but more study in such settings is needed. An additional limitation of this study is not identifying whether patients had culture-positive sepsis. In the hospitals studied, laboratory testing was advised in patients with suspected sepsis, 50-70% of results were available at two hours, and the least sick patients were least likely to have laboratory testing performed. Recent pediatric sepsis guidelines emphasize the importance of laboratory evaluation for organ dysfunction in suspected sepsis, however the performance of the model in settings where practice patterns differ will require study.²⁴

Conclusions

This model used routine EHR data to predict septic shock risk in children with suspected infection 2 hours after arrival, a critical time for treatment and transfer decisions. The threshold for considering patients high-risk could be customized to the clinical context. It demonstrated good discrimination and calibration in validation test sets within a regional pediatric care system, that included non-tertiary care sites. Additional external validation is needed to determine the generalizability of this model, which has the potential to improve early identification of pediatric septic shock.