Alerts and Alarms—Can Electronic Medical Records Help Children and Adolescents Survive Sepsis? Authors: Ben Driver, Franz E Babl, Daryl Cheng, Ben Gelbart, Elliot Long, Sarah McNab, Joshua Osowicki, Pauline Salib, Amanda Williams, Gabrielle M Haeusler Published: 2026 ABSTRACT ABSTRACTEarly recognition and treatment of paediatric sepsis has been shown to improve survival. Automated sepsis alerts embedded in electronic medical records (EMRs) offer real‐time identification of high‐risk patients and may improve timeliness of sepsis care, with possible limitations including overdiagnosis, overtreatment and alert fatigue. It is currently unclear whether automated alerts improve clinical outcomes in paediatric sepsis. We conducted a systematic review and narrative synthesis of studies evaluating clinical outcomes (mortality, PICU admission rate, length of stay), process measures (time to fluids/antibiotics, bundle adherence), and balancing measures (hospitalisation rate, antibiotic use) after implementation of EMR‐based automated sepsis alerts in patients under 18 years of age. Twelve studies met inclusion criteria. Automated alerts were associated with improvements in process measures, notably reduced time to antibiotics (4/6 studies) and fluids (2/5 studies). Mortality was reduced in one study of PICU patients (29.9% vs. 17.4%, p = 0.011), and a trend towards reduced PICU admissions was observed. No significant impact on hospital or PICU length of stay was noted. Alerts did not increase hospital admission rates or antibiotic usage. Automated sepsis alerts may improve time to fluids and antibiotics in paediatric sepsis. While inconsistently reported, alerts were not found to increase hospital admission rates or antibiotic use. The impact on clinical outcomes, particularly mortality, remains uncertain. High quality randomised studies are needed to define their role in improving paediatric sepsis outcomes.Keywords: alert, children, electronic medical record, paediatric, sepsis Introduction Sepsis is a leading cause of paediatric mortality globally [1]. Most deaths occur due to shock and organ dysfunction within 48–72 h of initial treatment [2]. Early recognition and treatment of children with sepsis have been shown to reduce mortality [2, 3, 4]. To improve early recognition of sepsis, international guidelines recommend systematic screening for sepsis in children [2, 5]. However, the most effective method of sepsis screening remains unclear. The first sepsis screening tools relied on manual input of patient data into forms or electronic medical records (EMRs). More recently, there have been efforts to embed automated screening tools into EMRs, which prospectively monitor patient data to generate interruptive or non‐interruptive alerts that prompt bedside evaluation and treatment. Automated screening tools have the potential to identify children at high risk of deterioration early in the clinical course but have limitations including overdiagnosis, overtreatment, and alert fatigue [6]. Alerts must be adequately sensitive to detect most cases of severe sepsis and septic shock while being specific enough to minimise false positive cases. Implementation of adult sepsis alert systems in the emergency department (ED) has been associated with improved patient outcomes such as reduced mortality, improved adherence to sepsis bundle elements, and reduced hospital length of stay [7]. However, detection of sepsis in paediatric patients is especially difficult: there is no gold standard for diagnosis, children with critical and benign illnesses have overlapping clinical features, the period of physiological compensation can be prolonged (e.g., hypotension is a late sign of sepsis in children), and establishing thresholds for alerts is complicated by developmental changes in normal ranges for most vital signs. There is emerging evidence that automated sepsis screening in the paediatric setting may identify patients more rapidly and have greater sensitivity and specificity than manual screening tools [6, 8]. However, it is not known whether automated screening is associated with improved outcomes. To address this evidence gap, we reviewed studies of automated sepsis alerts in paediatric patients that published clinical outcomes such as mortality, paediatric intensive care unit (PICU) admission, and length of stay (LOS); or process measures such as time to implementation of bundle elements and bundle adherence. To assess overdiagnosis and overtreatment, we integrated balancing measures such as antibiotic use and hospital admission rate. We aim to assess the quality of evidence for paediatric sepsis alert systems and help to direct future research efforts into evaluating sepsis alerts. Materials and Methods Search Strategy The online databases Medline, Embase, and PubMed were searched on 8th September 2024 using a structured search that contained sepsis terms, paediatric terms, and terms relating to screening/alerts (Supporting Information—Search Strategy). The search was completed with the assistance of the hospital librarian. Study Selection and Outcomes Automated sepsis screening was defined as an EMR‐based system that collects real‐time clinical data to generate interruptive or non‐interruptive alerts to facilitate sepsis diagnosis and treatment. Titles and abstracts were reviewed using Covidence software to identify articles for full‐text review. The inclusion criteria were peer reviewed studies written in English that described the target population (paediatric patients less than 18 years of age), intervention (automated sepsis alert), comparator (standard of care, including historic controls) and outcomes (listed below) and were published after 2010. Outcomes were grouped into (i) clinical outcomes (mortality, PICU admission, LOS), (ii) process measures (time to bundle elements, bundle adherence), and (iii) balancing measures (admission rate, antibiotic use). Neonatal studies and EMR simulation studies were excluded. Studies describing adult and paediatric patients were excluded if paediatric data were not presented separately. Data Extraction Data was extracted from the original manuscripts by a single reviewer. Data extracted included location, study duration, study design, alert criteria, clinical outcomes, process measures, and balancing measures. Risk of bias was assessed using the Newcastle‐Ottawa Scale for cohort studies [9]. Where statistical tests were conducted by the original authors on study data, these were displayed alongside the outcomes. Results Search Results The initial search yielded 903 studies. After duplicate removal and screening titles and abstracts, the full text of 50 articles was reviewed by a single reviewer (Figure 1) [10]. Twelve studies met our inclusion criteria and were included in this review (Tables 1, 2, 3). PRISMA flowsheet. TABLE 1Summary of included studies.Author, year (country)Study duration and settingDesignControlInterventionHuddle (Y/N)Time to fluids (median unless stated)Time to antibiotics (median unless stated)Bundle adherencePICU admissionMortality n (%)Length of stay (median unless stated)Time zero Balamuth, 2017 (USA) 2013–2015 ED Prospective cohort, single siteNo screening, sepsis protocol and order set available (n = 86 037)Automated screening (n = 96 472, inc 326 severe sepsis)Yes – after alert firing consultant/fellow, bedside nurse to determine whether sepsis protocol activatedNRNRNRPre: 1.5% (n = 1291), post 1.4% (n = 1351)In‐hospital: pre 0.02% (n = 13), post 0.02% (n = 17)ED LOS: pre 203 m, post 196 mNAEmerson, 2019 (USA) 2015–2018 ED Prospective cohort, single siteNo screening (n = 439)Automated screening (n = 1271)NoNR All: pre 154 m, post 114 m (mean) FN: pre 116 m, post 55 m (mean) NRNRNRNRTriageStinson, 2019 (USA) 2016–2018 Inpatients (excluding cardiology and PICU) Retrospective cohort, single siteNo screening (n = 9955)Automated screening (n = 10 465, inc n = 109 shock huddle positive)Yes – ‘shock huddle’ composed of shock RN and PICU providerNRNRNRNRIn‐hospital: pre 0.52% (n = 54), post 0.54% (n = 54) of all inpatients Hospital LOS: pre 5.4d, post 5d ICU LOS: pre 4.4d, post 4.6d NADewan, 2020 (USA) 2017–2019 PICU (excluding cardiac surgery ICU and NICU) Prospective cohort, single siteManual screening (n = 70 severe sepsis)Automated screening (n = 80 severe sepsis)Yes – huddles also in place pre‐intervention. Bedside assessment with nurse and provider to determine whether sepsis protocol activated.NRNRFluid bolus < 20 min, lactate, abx < 1 h: pre 18.6% (n = 13), post 34% (n = 27)*NA – PICU patientsNRNRNRVidrine, 2020 (USA) 2016–2019 PICU Mixed prospective/retrospective cohort, single siteNo screening (n = NR ‘on average 65 cases of IPSO‐defined severe sepsis per year’)Automated screening (n = NR ‘on average 65 cases of IPSO‐defined severe sepsis per year’)Yes – huddles also in place pre‐interventionWithin 20 m: pre 54.7%, post 61.8% (p = 0.39) Pre 1.53 h, post 1.05 h* Within 1 h: pre 33.9%, post 45.5% (p = 0.20) Lactate within 1 h: pre 59.4%, post 71.1% (p = 0.15)NA – PICU patientsNRNRBlood cultureGibbs, 2021 (USA) 2017, 2019 Inpatients (excluding PICU) NR – likely mixed prospective/retrospective cohort, single siteNR (n = 79)Automated screening (n = 154)Yes – nurse, provider, charge nurse, patient care assistant to determine whether sepsis order set utilisedPre 135 m, post 73.5 m (p = 0.11)Pre 147 m, post 103 m (p = 0.44)NRPre 72.2% (n = 57), post 31% (n = 48)** 3‐day sepsis‐attributable: pre 2.5% (n = 2), post 0% (n = 0) (p = 0.11) 30‐day: pre 3.8% (n = 4), post 1.3% (n = 2) (p = 0.34) All‐cause: pre 6.3% (n = 5), post 1.9% (n = 3) (p = 0.07) Hospital LOS: pre 20.8d, post 15.2d (p = 0.08)First of alert firing, sepsis order set, positive sepsis huddle, fluid bolus, abxEisenberg, 2021 (USA) 2017–2019 ED Retrospective cohort, single siteManual screening (n = 137)Automated screening (n = 136)Yes – members of ‘care team’ at bedside to determine whether order set utilisedPre 67 m, post 79 m (NS)Pre 96 m, post 95 m (NS)NRPre 91% (n = 124), post 88% (n = 119)30‐day (severe sepsis group): pre 2.9% (n = 4), post 2.2% (n = 3) ICU free days: pre 26.8d, post 26d Hospital free days: pre 20.4d, post 20.6d TriageAlturki, 2022 (Saudi Arabia) 2014–2018 PICU Retrospective cohort, single siteNR (n = 147)Automated screening (n = 161)NoPre 107 m, post 30 m***Pre 170.5 m, post 131 m*NRNA – PICU patientsIn‐hospital: pre 29.9% (n = 44), post 17.4% (n = 28)*PICU LOS: pre 7.9d, post 6.8d (mean, p = 0.442)Alert firing (or retrospective application of alert for pre‐implementation patients)Toews, 2022 (USA) 2012, 2018 ED Retrospective cohort, single siteNo screening (n = 1416)Automated screening (n = 1416)Yes – provider determines to continue routine care, putting patient on ‘sepsis watcher’ pathway or activating ‘sepsis alert’ pathwayNRNRNRPre 23.7% (n = 335), post 23.2% (n = 328)In‐hospital: pre 1.27% (n = 18), post 1.06% (n = 15) of patients with positive sepsis alert Hospital LOS: pre 2d, post 2d PICU LOS: pre 7d, post 6d Patients with ICD sepsis code: pre 13d, post 7d PICU patients with ICD sepsis code: pre 14d, post 8d NAMartinez, 2023 (USA) 2016–2020 ED Mixed prospective/retrospective cohort, single siteNo screening (n = 1105)Automated screening (n = 1513), automated screening plus electronic huddle (n = 2117)Yes – MD/NP and nurse assess patient within 10 min of alert to determine care plan and length of time for reassessmentWithin 3 h: pre‐alert 49.0%, electronic huddle 55.2%**Within 3 h: pre‐alert 37.9%, electronic huddle 50.7%***NRNRIn‐hospital: pre‐alert 0.18% (n = 2), alert 0.13% (n = 2), electronic huddle 0.09% (n = 2)Hospital LOS: pre‐alert 5.5d, alert 5.5d, electronic huddle 4.7d (mean)Alert firingOddiri, 2023 (USA) 2014–2020 Inpatients (excluding NICU) Mixed prospective/retrospective cohort, single siteNo screening (n = 41)Automated screening (n = 68)NoNRNRAbx + fluid bolus < 1 h: pre 37%, post 69%**NRNRNRNAFeinstein, 2023 (Israel) 2016–2021 Inpatients (excluding PICU) Mixed prospective/retrospective cohort, single siteNo screening (pre‐intervention n = 1670, control ward with no alert during intervention period n = 283)Automated screening (n = 347)NoNRNRNRPre 5.6% (n = 93), post 4.6% (n = 16), control 6% (n = 17) (p = 0.43 intervention vs. control) 30‐day: pre 0.7% (n = 12), post 1.7% (n = 6), control 1.8% (n = 5) (p = 0.97 intervention vs. control) Hospital LOS: pre 4d, post 4d, control 4d PICU LOS: pre 2d, post 2d, control 5.5d, (p = 0.1 intervention vs. control) NA Note: * p < 0.05, **p < 0.01, ***p < 0.001.Abbreviations: abx, antibiotics; d, day; FN, febrile neutropenia; h, hour; inc, including; LOS, length of stay; m, minute; MD, doctor of medicine; NA, not applicable; NICU, neonatal intensive care unit; NP, nurse practitioner; NR, not reported; NS, not significant; RN, registered nurse. Summary of included studies. Balamuth, 2017 (USA) 2013–2015 ED 2015–2018 ED All: pre 154 m, post 114 m (mean) FN: pre 116 m, post 55 m (mean) 2016–2018 Inpatients (excluding cardiology and PICU) Hospital LOS: pre 5.4d, post 5d ICU LOS: pre 4.4d, post 4.6d 2017–2019 PICU (excluding cardiac surgery ICU and NICU) 2016–2019 PICU Pre 1.53 h, post 1.05 h* Within 1 h: pre 33.9%, post 45.5% (p = 0.20) 2017, 2019 Inpatients (excluding PICU) 3‐day sepsis‐attributable: pre 2.5% (n = 2), post 0% (n = 0) (p = 0.11) 30‐day: pre 3.8% (n = 4), post 1.3% (n = 2) (p = 0.34) All‐cause: pre 6.3% (n = 5), post 1.9% (n = 3) (p = 0.07) 2017–2019 ED ICU free days: pre 26.8d, post 26d Hospital free days: pre 20.4d, post 20.6d 2014–2018 PICU 2012, 2018 ED Hospital LOS: pre 2d, post 2d PICU LOS: pre 7d, post 6d Patients with ICD sepsis code: pre 13d, post 7d PICU patients with ICD sepsis code: pre 14d, post 8d 2016–2020 ED 2014–2020 Inpatients (excluding NICU) 2016–2021 Inpatients (excluding PICU) 30‐day: pre 0.7% (n = 12), post 1.7% (n = 6), control 1.8% (n = 5) (p = 0.97 intervention vs. control) Hospital LOS: pre 4d, post 4d, control 4d PICU LOS: pre 2d, post 2d, control 5.5d, (p = 0.1 intervention vs. control) Note: * p < 0.05, **p < 0.01, ***p < 0.001. Abbreviations: abx, antibiotics; d, day; FN, febrile neutropenia; h, hour; inc, including; LOS, length of stay; m, minute; MD, doctor of medicine; NA, not applicable; NICU, neonatal intensive care unit; NP, nurse practitioner; NR, not reported; NS, not significant; RN, registered nurse. TABLE 2Balancing measures.Author, yearAntibiotic useHospital admissionOther balancing measuresBalamuth, 2017NRPre 16.1%, post 15.4%Proportion of patients with severe sepsis who were not treated with ED sepsis protocol/order set (missed patients): pre 17.3%, post 3.8%Emerson, 2019NRNRNRStinson, 2019NRNA – patients already hospitalisedNRDewan, 2020NRNA – patients already hospitalisedNRVidrine, 2020NRNA – patients already hospitalisedDays between delayed sepsis recognition: pre 9d, post 28d (mean)Gibbs, 2021Antibiotic days: pre 19d, post 9d (median)**NA – patients already hospitalisedNREisenberg, 2021NRPre 8.8%, post 12.5%NRAlturki, 2022NRNA – patients already hospitalisedNRToews, 2022NRPre 42.2% (n = 597), post 40.0% (n = 567) Median cost for encounters affected by sepsis screening: pre $6454, post $5084 (21.2% reduction) Median cost for encounters with ICD sepsis code: pre $58 685, post $28 672 (51.1% reduction) Martinez, 2023NRPre‐alert 58.6%, alert 63.5%, electronic huddle 63.4%NROddiri, 2023NRNRNRFeinstein, 2023NRNA – patients already hospitalisedNR Note: *p < 0.05, **p < 0.01, ***p < 0.001.Abbreviations: NA, not applicable; NR, not reported. Balancing measures. Median cost for encounters affected by sepsis screening: pre $6454, post $5084 (21.2% reduction) Median cost for encounters with ICD sepsis code: pre $58 685, post $28 672 (51.1% reduction) Note: *p < 0.05, **p < 0.01, ***p < 0.001. Abbreviations: NA, not applicable; NR, not reported. TABLE 3Alert criteria.Author, yearAlert setting and typeSepsis alert trigger rateAlert sensitivityAlert specificityHRTempBPRRHypoxiaMental statusSurrogates for peripheral perfusion (cap refill, cool peripheries, pulses)Lab valuesHigh risk conditionBlood cultureBalamuth, 2017ED—interruptive 1112 alerts per 96 472 ED encounters (1.2%) 326 of 1112 alerts treated for severe sepsis (29.3%) Severe sepsis 86.2%Severe sepsis 99.1%xxxxxEmerson, 2019ED – interruptive, Alert criteria not providedNR—included patients included those with IV abx + blood culture orders rather than alert positive patientsNRNRStinson, 2019Inpatient—unclear if interruptive9241 eligible admissions, shock huddle in 206 cases (2.23%) comprising 109 unique patientsNRNRxxxxxxxDewan, 2020PICU—interruptiveFire rate 7.9% of daily encounters (852/10805 patient days), 1.4% requiring huddleSepsis 92.5%, severe sepsis 73.3%Sepsis 95.6%, severe sepsis 92.5%xxxxxVidrine, 2020PICU – interruptiveNRNRNRxxxxxGibbs, 2021Inpatient—interruptiveNRNRNRxxxxxxxEisenberg, 2021 ED—interruptive Full alert criteria not provided 3136 alerts per 61 195 encounters (5.1%)Severe sepsis or septic shock 84.6%Severe sepsis or septic shock 95.1%xxxxxxx (leukocyte count)xAlturki, 2022Inpatient—interruptiveNRNRNRxxxx (WCC, bands %, BGL, lactate, bilirubin, creatinine)Toews, 2022ED—interruptiveNRNRNRxxxxxxMartinez, 2023ED—unclear if interruptive2% per ED encountersNRNRxxxx x (SpO2) xxx (WCC)xxOddiri, 2023Inpatient—interruptiveNRNRNRxxxxx (PaO2/FiO2)xx (lactate, base excess, WCC, platelets, creatinine, bilirubin, ALT, PT/INR)Feinstein, 2023Inpatient347 of 2993 of non‐PICU inpatients (11.5%)NRNRxxxxxxAbbreviations: abx, antibiotics; ALT, alanine aminotransferase; BGL, blood glucose level; BP, blood pressure; HR; heart rate; INR, international normalised ratio; NR, not reported; PaO2/FiO2, ratio of arterial oxygen partial pressure to fractional inspired oxygen; PT, prothrombin time; RR, respiratory rate; SpO2, oxygen saturation; WCC, white cell count. Alert criteria. 1112 alerts per 96 472 ED encounters (1.2%) 326 of 1112 alerts treated for severe sepsis (29.3%) ED—interruptive Full alert criteria not provided x (SpO2) Abbreviations: abx, antibiotics; ALT, alanine aminotransferase; BGL, blood glucose level; BP, blood pressure; HR; heart rate; INR, international normalised ratio; NR, not reported; PaO2/FiO2, ratio of arterial oxygen partial pressure to fractional inspired oxygen; PT, prothrombin time; RR, respiratory rate; SpO2, oxygen saturation; WCC, white cell count. Risk of Bias Assessment Using the Newcastle‐Ottawa Scale, studies were assigned a quality score with ≥ 7 points indicating high quality and < 7 points indicating low quality [9]. Nine studies scored 7 and three studies scored 8, giving a mean score of 7.25 (Table 4). TABLE 4Risk of bias assessment using the Newcastle‐Ottawa Scale (NOS).StudySelectionComparabilityOutcomeAuthor, yearRepresentativeness of the exposed cohortSelection of the non‐exposed cohortAscertainment of exposureDemonstration that the current outcome of interest was not present at start of studyComparability of cohorts on the basis of the design or analysisAssessment of outcomeWas follow‐up long enough for outcomes to occurAdequacy of follow up of cohortsQuality scoreBalamuth, 2017111NA2—attempt to control for overtreatment1118Emerson, 2019111NA11117Stinson, 2019111NA11117Dewan, 2020111NA11117Vidrine, 2020111NA11117Gibbs, 2021111NA11117Eisenberg, 2021111NA2—attempts to control for clinician bias to treat based on sepsis alert by using sepsis definition with organ dysfunction criteria rather than intention to treat.1118Alturki, 2022111NA11117Toews, 2022111NA2—attempts to correct for lack of exclusions in 2012 cohort. Also, 2012 financial cost data adjusted for inflation using consumer price index.1117Martinez, 2023111NA11117Oddiri, 2023111NA11117Feinsten, 2023111NA2—control ward1118Abbreviation: NA, not applicable. Risk of bias assessment using the Newcastle‐Ottawa Scale (NOS). Abbreviation: NA, not applicable. Study Characteristics Ten (83%) studies were conducted in the USA with the remaining two studies based in Israel and Saudi Arabia. Seven (58%) studies compared automated screening to no screening, three (25%) studies compared automated to manual (paper‐based) screening, and two (17%) studies did not report on the prior standard of care. In eight studies (67%), multidisciplinary huddles were triggered by the automated sepsis alert to facilitate clinical assessment prior to sepsis pathway activation. Five (42%) studies were based in the ED, four (33%) in inpatient wards and three (25%) in PICU. Sepsis definitions varied for each study (Table 1), with the most common consensus definition being the Improving Paediatric Sepsis Outcomes (IPSO) criteria which were used in three studies (25%) [11, 12, 13]. The IPSO criteria were specifically designed for quality improvement initiatives using the EMR for sepsis case ascertainment and incorporate International Classification of Diseases (ICD‐10) codes, investigations and resuscitation measures related to sepsis management [14]. Alert trigger criteria varied between studies (Table 3). All studies included a combination of vital signs, with six also incorporating pathology or microbiological results. The trigger rate for the alerts ranged from 1.2% to 11.5% of the screened population. In the studies that reported alert sensitivity and specificity, sensitivity ranged from 73.3% to 92.5% and specificity ranged from 92.5% to 99.1%. Process Measures One or more process measures were reported in six studies [12, 13, 15, 16, 17, 18]. In four out of six studies reporting antibiotic metrics, the median time to antibiotic was either reduced or the percentage of patients receiving antibiotics within three hours was increased in the post sepsis alert cohort (statistically significant in three) [12, 15, 17, 18]. There was no change in time to antibiotic in the remaining two studies [13, 16]. In two out of five studies reporting fluid administration, the median time to fluid therapy was significantly reduced, or the percentage of patients receiving fluid therapy within three hours was significantly increased [17, 18]. There was no change in median time to fluids in three studies [12, 13, 16]. In the two studies that reported on overall bundle adherence, there was a significant improvement in the proportion of patients meeting targets for time to bundle elements in the post sepsis alert cohort [11, 19]. There was variability in the definition of time zero (Table 1). Common time zero definitions included alert firing, blood culture collection, or triage time for ED based alerts. Clinical Outcomes Eight studies evaluated mortality as an outcome. Overall mortality varied widely depending on the target patients and clinical setting, ranging from 0.13% hospital all‐cause mortality in the USA to 29.9% in PICU patients in Saudi Arabia [17, 18]. Only one study demonstrated a significant decrease in mortality. Set in PICU, in‐hospital mortality reduced from 29.9% to 17.4% post implementation [17]. Five studies evaluated PICU admission rates. A significant reduction in PICU admission rate in patients with severe sepsis was noted in one study from 72.2% to 31% [13]. In this study, the pre‐intervention cohort had a lower rate of severe sepsis per 1000 admissions (3.26 vs. 5.86) and a lower median age at time zero (5.0 vs. 7.6 years), which may affect the comparability of the groups. In the remaining four studies, a non‐significant trend towards reduced PICU admission rates post alert implementation was reported. Eight studies evaluated length of stay (LOS) as an outcome. No significant differences were reported on either hospital‐wide or PICU LOS post implementation of automated screening with the exception of reduced hospital and PICU LOS in a subgroup analysis of patients with ICD‐code confirmed sepsis in one study [20]. Balancing Measures Four studies reported on admission rates from ED and found mixed results, with no studies reporting a statistically significant change in overall hospitalisation rate in the post alert cohort (Table 2). Only one study reported on antibiotic use, reporting a significant reduction in median antibiotic duration from 19 to 9 days post sepsis alert implementation, although a concurrent antimicrobial stewardship initiative is the most likely explanation [13]. Additionally, one study showed a decrease in overall median cost of 21.1% for hospital encounters affected by sepsis screening and a decrease of 51.1% for encounters with ICD‐code confirmed sepsis compared to the pre‐alert cohort [20]. Discussion We identified 12 studies, published since 2010, that assessed the impact of automated EMR alerts on the identification of sepsis in paediatric patients in a range of settings including emergency department, inpatient wards and paediatric ICU. The studies reviewed were highly heterogenous in terms of setting, alert criteria and outcomes. Studies included were mostly pre‐ and post‐intervention cohort studies that are highly prone to confounding, making it difficult to associate differences in groups to sepsis alerts alone. More than half of the interventions incorporated huddles to inform sepsis pathway activation. Overall, the alerts were more consistently associated with improvements in process measures than with clinical outcomes such as mortality, PICU admission, or length of stay. These included improvements in bundle adherence and time to fluid and antibiotic therapy. In‐hospital mortality was significantly reduced in one study, with no change in seven studies. One study reported a decrease in PICU admission rates, with a trend towards reduced PICU admissions in the remaining four studies. Balancing measures of hospital admission rate and antibiotic use did not identify a signal for increased burden on the health care system. Given the limitation of pre‐ and post‐implementation cohort studies being prone to confounding, it is important to consider factors that may have contributed to care timeliness other than introduction of sepsis alerts and associated bundles of care. For example, Alturki reported significant reductions in time to fluids and antibiotics in PICU patients [17]. However, because the alert was applied retrospectively to past data, the reported time differences may reflect the study methods rather than true impact. Vidrine et al. reported a decrease in median time to antibiotics in PICU patients with no change in time to fluids post alert implementation [12]. In this study, the automated alert was part of a broad quality improvement initiative which included sepsis education, simulation, a new PICU sepsis algorithm, and discussion of delayed sepsis recognition as a key metric in monthly morbidity and mortality meetings. Therefore, it is difficult to ascertain the effect of automated screening alone on decreased time to process measures. An ideal study design to address the effect of confounding would be a stepped‐wedge cluster randomised controlled trial. At the time of our search, we identified no such studies evaluating automated sepsis alerts in children. However, a multi‐center stepped‐wedge cluster randomised trial in adults in 5 hospitals in Saudi Arabia found that electronic alert implementation was associated with reduced overall 90‐day mortality (adjusted risk ratio 0.85) compared to no screening [21]. One exception in study design was a study by Feinstein et al., in which the post intervention cohort was compared to both the pre‐intervention group and concurrently with a second control ward with no screening [22]. This study design allows for side‐by‐side comparison of wards with and without automated sepsis screening, but relies on different inpatient wards being comparable. There were no significant differences in PICU admission, mortality, or LOS found when comparing the post intervention and control ward groups. Overall, there was a less pronounced effect of automated alerts on outcome measures such as mortality, PICU admission rates, and length of stay. No studies demonstrated a reduction in sepsis attributable mortality. Only one study by Alturki et al. demonstrated a significant decrease in in‐hospital mortality post alert implementation [17]. However, this study was conducted in PICU patients with severe sepsis or septic shock with a high mortality rate, making the results of this study less applicable to other inpatient settings. Given the low prevalence of non‐PICU paediatric sepsis deaths, studies were underpowered to detect changes in mortality post alert implementation. In addition, there was a wide range of mortality reported by studies, which reflected both the differences in setting but also inclusion criteria, with more sensitive alerts and ward/ED settings being associated with lower mortality than more specific alerts for sepsis/septic shock and PICU. Aside from mortality, another useful indicator of impact of automated sepsis screening is PICU admission rates. One study by Gibbs et al. demonstrated a significant reduction in admissions to PICU post alert implementation [13], with the four remaining studies showing a trend towards reduced PICU admission rate. The large reduction in PICU admission rate in Gibbs et al. in patients with severe sepsis (72.2% vs. 31%) may be attributable to differences in baseline characteristics between the pre‐ and post‐intervention groups and institutional factors such as the opening of a new critical care tower during the study period. In general, automated screening was not associated with changes in overall length of stay, except for one study by Toews et al. in which automated screening was associated with reduced hospital and PICU LOS for patients with ICD‐code sepsis and an associated reduction in costs of care [20]. A concern regarding automated sepsis alerts is that they might lead to overtreatment of patients which may manifest as increased antibiotic use and increased rates of admission to hospital. While our review did not identify a signal for this, these metrics were inconsistently reported. One study reported a decrease in median antibiotic days; however, the sepsis alert was implemented concurrently with an antimicrobial stewardship initiative that likely impacted the median antibiotic days measure [13]. Reassuringly, hospital admission rates did not appear to be significantly impacted, with two studies reporting a slight decrease and two reporting a slight increase in numbers admitted. In the only study that included economic analysis, median cost of encounters affected by sepsis screening was decreased by in the post‐alert cohort overall, with cost savings being greater in episodes of ICD‐code confirmed sepsis [20]. This was primarily attributed to reduced hospital and PICU LOS in the subgroup of patients with ICD‐code confirmed sepsis. There are several important limitations to this review, most notably the heterogeneous mix of studies in different settings using diverse alert criteria making direct comparison difficult. While the included studies scored as having a low risk of bias on the Newcastle‐Ottawa Scale for cohort studies, the pre‐ and post‐implementation design of the studies included is highly prone to bias from confounding, which limits our ability to examine the influence of automated sepsis alerts in isolation. For studies that reported on process measures, there was significant variation in time zero from alert firing to blood culture collection and ED arrival. Alert firing as time zero introduces complexity when retrospectively applying alert criteria to pre‐implementation data and introduces systematic bias towards improved process measures. Finally, we limited our search to 2010 and later; however this is unlikely to have had a significant impact as most hospitals have adopted EMR in the last decade, and automated alerts are even more recent. This is reflected by the fact that our earliest included paper was published in 2017 [23]. Conclusion In the studies included, automated sepsis alerts were more associated with improvements in process measures compared to clinical outcomes such as mortality, PICU admission rate and length of stay. Automated alerts were not associated with increased admissions to hospital or more antibiotic days and were shown to reduce healthcare costs in one study. Interpretation of these results is complex as studies were highly heterogenous, and pre‐ and post‐intervention cohort studies are prone to confounding. Future randomised studies are required to determine whether automated sepsis screening translates into improved outcomes. However, the importance of sepsis as a major driver of paediatric mortality globally warrants further investigation of the utility of automated alerts in the treatment of sepsis. Funding The authors have nothing to report. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Data S1: Supporting Information. Acknowledgements We acknowledge the contribution of Poh Chua, hospital librarian at the Royal Children's Hospital, for their valuable assistance with the literature search. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References Associated Data Supplementary Materials Data S1: Supporting Information. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.