Publication

• Title: Electronic Sepsis Screening Among Patients Admitted to Hospital Wards: A Stepped-Wedge Cluster Randomized Trial
• Acronym: SCREEN
• Year: 2025
• Journal published in: JAMA
• Citation: Arabi YM, Vishwakarma RK, Al-Dorzi HM, Al Mofleh A, Murthy S, Bihari S, et al; SCREEN Trial Group. Electronic sepsis screening among patients admitted to hospital wards: a stepped-wedge cluster randomized trial. JAMA. 2025 Mar 4;333(9):763-773.

Context & Rationale

• Background

  • Sepsis is common, time-critical, and frequently first recognised outside ICUs; delays in recognition and early treatment are associated with worse outcomes.
  • Ward-based clinical deterioration can be subtle, documentation may be intermittent, and escalation pathways vary; missed or late recognition is a persistent patient-safety problem.
  • Electronic health record (EHR) screening and automated alerts are widely implemented, but the evidence base for patient-centred benefit (especially mortality) has been inconsistent and often non-randomised.
  • Guidelines and quality initiatives have promoted systematic sepsis recognition/response programmes despite low-to-moderate certainty evidence for specific digital alert strategies.¹
  • Potential downsides include alert fatigue, overdiagnosis, increased broad-spectrum antibiotic exposure, iatrogenic harms, and resource diversion.
• Research Question/Hypothesis

  • Whether real-time electronic screening using qSOFA-derived alerts on hospital wards, coupled to a structured response workflow, reduces in-hospital mortality by day 90 compared with masked (silent) background alerts (usual care).
• Why This Matters

  • Electronic sepsis alerting is a high-visibility digital safety intervention with major opportunity cost; large pragmatic randomised evidence is essential before widespread adoption.
  • Even small absolute mortality effects can be clinically important at scale, but must be weighed against unintended harms (antibiotic-related complications, escalation cascades, false positives).
  • Stepped-wedge implementation designs are common for hospital-wide digital interventions; SCREEN offers a methodologically informative test case for this class of trials.

Design & Methods

• Research Question: Among patients admitted to hospital wards, does a real-time electronic sepsis screening and alert system (qSOFA-based) reduce in-hospital mortality by day 90 compared with masked alerts/usual care?
• Study Type: Pragmatic stepped-wedge cluster randomized trial; multicentre; ward-level clusters; 5 hospitals; 9 sequences; 10 study periods (baseline period with all clusters in control, then sequential crossover to screening).
• Population:
  • Setting: Adult/adolescent hospital wards (medical, surgical, mixed, oncology) in 5 hospitals in Saudi Arabia; October 2019 to July 2021.
  • Inclusion criteria: Patients aged 14 years or older admitted to participating eligible wards during the study periods.
  • Key exclusions (cluster and patient level): Non-eligible wards (paediatric/neonatal, obstetric, cardiac); patients without commitment to full life-support at ward admission; data from wards during periods of conversion to ICU beds for COVID-19 surges were excluded from analysis.
  • Analysis assignment: Patients were analysed according to the allocation status of the ward at first ward admission (even if subsequently transferred).
• Intervention:
  • Electronic screening trigger: qSOFA-derived alert generated when 2 or more qSOFA components were present within a 12-hour window (based on recorded respiratory rate, systolic blood pressure, and mental status documentation).
  • Delivery: Real-time EHR alerts displayed (revealed) to nurses and physicians on intervention wards, with escalation prompts and a structured sepsis assessment workflow embedded in the EHR.
  • Response pathway: Bedside assessment and recommended diagnostic/therapeutic actions (eg, lactate testing, cultures, fluids, antibiotics, escalation as appropriate) guided by the alert workflow; implementation and procedures described in the published protocol and prespecified statistical analysis plan.²³
• Comparison:
  • Masked alert: The same electronic alert algorithm ran in the background but remained silent (not displayed) on control wards.
  • Usual care: Clinicians managed suspected infection/sepsis per routine practice and local pathways without the intervention’s visible alert prompts.
• Blinding: Unblinded at the clinical level (visible alerts could not be masked in intervention periods); outcomes were prespecified and largely EHR-derived, limiting subjective outcome assessment bias for the primary endpoint.
• Statistics: A total of 65,250 patients were required to detect a 0.79% absolute reduction in in-hospital mortality by day 90 (from 3.13% to 2.46%) with 80% power at the 5% two-sided significance level (intracluster correlation coefficient 0.22); primary analysis used mixed-effects generalised linear modelling with log link to estimate adjusted relative risk accounting for period effects and clustering; patients analysed according to ward allocation at admission; secondary outcomes controlled using false discovery rate procedures.²³
• Follow-Up Period: In-hospital follow-up to day 90 after ward admission (discharge before day 90 counted as alive for the primary endpoint); last follow-up reported in late October 2021.

Key Results

This trial was not stopped early. The stepped-wedge schedule was modified during the COVID-19 pandemic (two consecutive periods extended from 2 to 3 months, and some ward-period data were excluded when wards were converted to ICUs).

Outcome	Sepsis screening	No screening	Effect	p value / 95% CI	Notes
In-hospital mortality by day 90 (primary)	937/29,442 (3.2%)	961/30,613 (3.1%)	aRR 0.85	95% CI 0.77 to 0.93; P<0.001	Unadjusted risk difference 0.0% (95% CI −0.2% to 0.3%); reported number needed to screen 206 (95% CI 140 to 425).
ICU admission	1,753/29,442 (6.0%)	1,550/30,613 (5.1%)	aOR 1.03	95% CI 0.94 to 1.13; P=0.50	FDR-adjusted P=0.50.
Rapid response team activation	1,402/29,442 (4.8%)	1,330/30,613 (4.3%)	aOR 0.96	95% CI 0.89 to 1.04; P=0.27	FDR-adjusted P=0.30.
Code blue activation	296/29,442 (1.0%)	188/30,613 (0.6%)	aOR 1.24	95% CI 1.02 to 1.50; P=0.03	FDR-adjusted P=0.04.
Vasopressor therapy	862/29,442 (2.9%)	844/30,613 (2.8%)	aRR 0.86	95% CI 0.78 to 0.94; P=0.002	FDR-adjusted P=0.004.
Mechanical ventilation	814/29,442 (2.8%)	700/30,613 (2.3%)	aOR 0.93	95% CI 0.84 to 1.03; P=0.18	FDR-adjusted P=0.22.
Incident kidney replacement therapy	1,294/28,842 (4.5%)	1,088/30,168 (3.6%)	aOR 1.20	95% CI 1.11 to 1.31; P<0.001	Denominators exclude pre-existing dialysis and those started within 48 hours of ward admission; FDR-adjusted P<0.001.
New multidrug-resistant organisms	526/29,442 (1.8%)	598/30,613 (2.0%)	aOR 0.88	95% CI 0.78 to 0.99; P=0.03	FDR-adjusted P=0.04.
New Clostridioides difficile infection	186/29,442 (0.6%)	147/30,613 (0.5%)	aOR 1.30	95% CI 1.03 to 1.65; P=0.03	FDR-adjusted P=0.04.
Hospital length of stay (days)	Median 3.7 (IQR 1.8–7.8)	Median 3.9 (IQR 2.0–8.1)	β −0.10	95% CI −0.12 to −0.07; P<0.001	Negative binomial model; FDR-adjusted P<0.001.
ICU-free days up to day 90	Mean 86.9 (SD 15.7)	Mean 87.0 (SD 15.6)	β 0.005	95% CI 0.002 to 0.008; P<0.001	Negative binomial model; missing data: 56 vs 45; FDR-adjusted P=0.004.
Antibiotic-free days up to day 90	Mean 83.8 (SD 16.1)	Mean 83.9 (SD 16.0)	β 0.004	95% CI 0.001 to 0.006; P=0.001	Negative binomial model; patients who died counted as 0; FDR-adjusted P=0.004.

• Across 60,055 ward admissions, electronic screening was associated with lower adjusted in-hospital mortality by day 90 (aRR 0.85; 95% CI 0.77 to 0.93), despite a near-zero unadjusted absolute risk difference (0.0%).
• Process separation favoured more diagnostic and supportive care actions after alerts (eg, lactate testing and intravenous fluids), but secondary outcomes showed a mixed profile including higher code blue, higher incident kidney replacement therapy, and higher C difficile infection.
• Absolute effects were small (low baseline mortality), which increases sensitivity to secular trends and modelling assumptions in stepped-wedge implementation trials.

Internal Validity

• Randomisation and allocation
  • Ward clusters were randomised into 9 sequences using a concealed computer-generated list, with sequential crossover from masked to revealed alerts.
  • Stepped-wedge designs mitigate between-cluster confounding but are vulnerable to time-varying secular trends and implementation effects; reporting standards emphasise explicit handling of period effects and contamination.⁴
• Dropout, exclusions, and pandemic-related disruptions
  • Two wards were excluded after randomisation, and additional ward-period data were excluded when wards were converted to ICUs for COVID-19 surges (reducing analysable cluster-periods and potentially altering case-mix).
  • 2,323/62,378 ward admissions (3.7%) were excluded because there was no commitment for full life-support at admission, which may limit inference to full-treatment-eligible populations.
• Performance and detection bias
  • Blinding was not feasible; clinician behaviour (diagnostics, fluids, antibiotics, escalation) was an intended mediator and could be influenced by co-interventions and evolving practice over time.
  • Primary outcome (in-hospital mortality by day 90) is objective, but is sensitive to discharge practices and competing risks of discharge alive vs death in-hospital.
• Protocol adherence and separation of the intervention
  • Alert burden in patients who triggered alerts: total alerts 9,447 vs 8,418; median alerts per patient 3 (IQR 1–5) vs 2 (IQR 1–4); time from ward arrival to first alert 23.1 hours (IQR 8.7–53.5) vs 19.4 hours (IQR 8.0–45.4).
  • Alert acknowledgement: nurse 3,154/4,299 (73.4%) vs 658/5,394 (12.2%); physician 3,109/4,299 (72.3%) vs 687/5,394 (12.7%).
  • Key process differences after alert (screening vs no screening): lactate within 12 hours 612/4,299 (14.2%) vs 556/5,394 (10.3%); intravenous fluids within 12 hours 617/4,299 (14.4%) vs 364/5,394 (6.8%); blood cultures during ward stay 2,592/4,299 (60.3%) vs 2,691/5,394 (49.9%); new antibiotics during ward stay 3,473/4,299 (80.8%) vs 3,892/5,394 (72.2%).
  • Control wards were not “pure” no-alert environments at the patient level (acknowledgement and revealed-alert exposure occurred in a minority), consistent with patient transfers and transition-period contamination, which would be expected to dilute contrasts.
• Baseline comparability and risk profile
  • Age and comorbidity burden were broadly similar (median age 59 years in both; Charlson comorbidity index median 2 in both), but ward mix and infection profiles differed over time (eg, COVID-19 diagnosis at admission 11.8% vs 5.2%; immunocompromised state 22.7% vs 15.8%).
  • Overall mortality was low (3.1%–3.2%), which amplifies the influence of small absolute differences and model specification choices.
• Statistical rigour
  • Analyses were prespecified with modelling approaches to account for clustering and period effects, with a contingency plan for model non-convergence (switching to adjusted odds ratios).³
  • Multiple secondary outcomes were addressed using false discovery rate control, reducing (but not eliminating) the risk of false positive secondary signals.
  • Target sample size was not fully reached (60,055 analysed vs 65,250 planned), but the primary outcome remained statistically significant.

Conclusion on Internal Validity: Overall, internal validity is moderate: the trial used a prespecified stepped-wedge cluster design with objective primary outcome and demonstrable process separation, but unblinded implementation during a pandemic with time-varying case-mix and evidence of contamination means the mortality signal remains sensitive to modelling and secular-trend assumptions.

External Validity

• Population representativeness
  • Broad, pragmatic inclusion of ward admissions across medical, surgical, mixed, and oncology wards supports applicability to routine inpatient care.
  • Key exclusions (paediatric/neonatal, obstetric, cardiac wards; patients without full life-support commitment) limit applicability to those settings and goals-of-care contexts.
  • Conducted during COVID-19 surges with ward-to-ICU conversions; the effect may differ in non-pandemic conditions.
• Applicability
  • Generalisation depends on EHR capability (real-time data capture and alert delivery), staffing models, escalation pathways, and local sepsis treatment practices.
  • qSOFA-based triggering is simple and scalable but has known sensitivity limitations for early sepsis detection; different trigger thresholds or algorithms may yield different benefit–harm balances.
  • Hospitals with high baseline adherence to sepsis bundles may see less incremental effect; conversely, settings with delayed recognition might see larger process gains but also more overtreatment.

Conclusion on External Validity: External validity is moderate: findings are most applicable to similarly resourced hospitals with mature EHR infrastructure and ward escalation pathways; transferability to resource-limited systems or substantially different ward case-mix and discharge practices is uncertain.

Strengths & Limitations

• Strengths:
  • Large pragmatic ward population (n=60,055) across multiple hospitals and ward types.
  • Stepped-wedge cluster randomisation suited to system-wide digital intervention rollout.
  • Objective primary outcome and prespecified protocol/statistical analysis plan.²³
  • Demonstrated process separation after alerts (diagnostics/supportive care) consistent with biological plausibility for benefit.
• Limitations:
  • Unblinded implementation with potential co-interventions and behavioural effects; stepped-wedge susceptibility to secular trends and time-varying confounding, accentuated by the COVID-19 pandemic.
  • Mortality signal depended on adjusted modelling with minimal unadjusted absolute difference, increasing sensitivity to model specification and period adjustment.
  • Evidence of contamination (acknowledgement and revealed-alert exposure in a minority of controls) likely reduced treatment contrast.
  • Secondary outcomes suggested potential harms (higher code blue, higher incident kidney replacement therapy, higher C difficile infection), emphasising the need for benefit–harm evaluation.

Interpretation & Why It Matters

• Clinical implications

  • SCREEN provides large-scale randomised evidence that ward-based electronic screening can be associated with lower adjusted in-hospital mortality, but the absolute risk difference was small and the signal relied on adjustment for clustering and time.
  • Implementation should be coupled to a coherent clinical response pathway and antimicrobial stewardship, given signals of increased C difficile infection and incident kidney replacement therapy.
  • Hospitals considering sepsis alerts should evaluate local baseline performance, alert specificity, staffing capacity, and downstream consequences (escalation cascades and iatrogenic harms), not only process metrics.
• Mechanistic plausibility

  • Process outcomes in alert-triggered patients showed higher rates of lactate testing and fluid ordering and increased culture and antibiotic activity across the ward stay, consistent with a pathway by which earlier recognition could improve outcomes.
  • The simultaneous increase in certain adverse outcomes suggests that more aggressive investigation/treatment and escalation may carry measurable trade-offs.

Controversies & Subsequent Evidence

• Adjusted vs unadjusted treatment effect in a stepped-wedge trial
  • The primary outcome showed a near-zero unadjusted absolute mortality difference but a statistically significant adjusted relative risk reduction; this makes interpretation particularly dependent on correct modelling of period effects, clustering, and time-varying case-mix.
  • Stepped-wedge trials are explicitly recognised as vulnerable to secular trends and implementation effects; rigorous reporting and transparent modelling choices are essential, especially when external shocks (eg, a pandemic) occur mid-trial.⁴
• Benefit–harm trade-offs
  • Secondary outcomes suggested potential unintended consequences (higher code blue activation, higher incident kidney replacement therapy, and higher C difficile infection), plausibly mediated by increased diagnostic/treatment intensity and antibiotic exposure.
  • This reinforces that “improved process metrics” are not inherently synonymous with net clinical benefit when interventions alter escalation and prescribing behaviour.
• How SCREEN fits with the broader digital alert literature
  • Systematic reviews of digital sepsis alerts report heterogeneous effects, frequent improvements in process measures, and uncertain or variable effects on mortality, reflecting differences in alert algorithms, response pathways, and study design quality.⁵
  • Related evidence on predictive analytics for sepsis (including machine learning approaches) highlights persistent issues with external validity, calibration drift, and the need for impact evaluations rather than accuracy-only studies.⁶
• Guideline alignment and knowledge translation
  • International sepsis guidelines have supported structured recognition and early management programmes, but specific recommendations for electronic alerting have historically rested on low certainty evidence; SCREEN materially strengthens the randomised evidence base in ward populations.¹

Summary

• Pragmatic stepped-wedge cluster randomised trial across 5 hospitals and 43 analysable wards (n=60,055 ward admissions) testing a qSOFA-triggered electronic sepsis alert with a structured response pathway.
• Primary outcome: lower adjusted in-hospital mortality by day 90 with screening (aRR 0.85; 95% CI 0.77 to 0.93), but minimal unadjusted absolute difference (0.0%).
• Clear process separation in alert-triggered patients (higher nurse/physician acknowledgement; higher lactate testing and fluid ordering; higher culture/antibiotic activity over the ward stay).
• Mixed secondary outcomes, including signals consistent with potential harms (higher code blue, higher incident kidney replacement therapy, higher C difficile infection) alongside lower vasopressor use and shorter hospital length of stay.
• Interpretation requires careful attention to stepped-wedge time confounding and pandemic-era secular trends, given the reliance on adjusted modelling for the mortality signal.

Further Reading

Other Trials

• 2012Hooper MH, Weavind L, Wheeler AP, et al. Randomized trial of automated, electronic monitoring to facilitate early detection of sepsis in the intensive care unit. Crit Care Med. 2012;40(7):2096-2101.
• 2015Semler MW, Weavind L, Hooper MH, et al. An electronic tool for the evaluation and management of sepsis in the ICU: a randomized controlled trial. Crit Care Med. 2015;43(8):1595-1602.
• 2017Shimabukuro DW, Barton CW, Feldman MD, Mataraso SJ, Das R. Effect of a machine learning-based severe sepsis prediction algorithm on patient survival and hospital length of stay: a randomised clinical trial. BMJ Open Respir Res. 2017;4(1):e000234.
• 2019Downing NL, Rolnick J, Poole SF, et al. Electronic health record-based clinical decision support alert for severe sepsis: a randomised evaluation. BMJ Qual Saf. 2019;28(9):762-768.
• 2022Taranbichi Y, Cheng A, Bar-Shain D, et al. Improving timeliness of antibiotic administration using a provider and pharmacist facing sepsis early warning system in the emergency department setting: a randomized controlled quality improvement initiative. Crit Care Med. 2022;50(3):418-427.

Systematic Review & Meta Analysis

• 2019Joshi M, Ashrafian H, Arora S, et al. Digital alerting and outcomes in patients with sepsis: systematic review and meta-analysis. J Med Internet Res. 2019;21(12):e15166.
• 2020Fleuren LM, Klausch TLT, Zwager CL, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis. Intensive Care Med. 2020;46(3):383-400.
• 2022Wang JY, Chen YX, Guo SB. Usefulness of qSOFA, SIRS, and NEWS for diagnosing severe sepsis and predicting mortality: a meta-analysis. PLoS One. 2022;17(4):e0266755.
• 2018Luther MK, Timbrook TT, Caffrey AR, et al. Vancomycin plus piperacillin-tazobactam and acute kidney injury in adults: a systematic review and meta-analysis. Crit Care Med. 2018;46(1):12-20.

Observational Studies

• 2011Sawyer AM, Deal EN, Labelle AJ, et al. Implementation of a real-time computerized sepsis alert in nonintensive care unit patients. Crit Care Med. 2011;39(3):469-473.
• 2017Manaktala S, Claypool SR. Evaluating the impact of a computerized surveillance algorithm and decision support system on sepsis mortality. J Am Med Inform Assoc. 2017;24(1):88-95.
• 2017Hiensch R, Poeran J, Saunders-Hao P, et al. Impact of an electronic sepsis initiative on antibiotic use and healthcare facility-onset Clostridium difficile infection rates. Am J Infect Control. 2017;45(10):1091-1100.
• 2019Raines K, Sevilla Berrios RA, Guttendorf J. Sepsis education initiative targeting qSOFA screening for non-ICU patients to improve sepsis recognition and time to treatment. J Nurs Care Qual. 2019;34(4):318-324.
• 2016Schorr C, Odden A, Evans L, et al. Implementation of a multicenter performance improvement program for early detection and treatment of severe sepsis in general medical-surgical wards. J Hosp Med. 2016;11(suppl 1):S32-S39.

Guidelines

• 2025National Institute for Health and Care Excellence. Suspected sepsis: recognition, diagnosis and early management for adults. NICE guideline NG253. 2025.
• 2025National Institute for Health and Care Excellence. Suspected sepsis: recognition, diagnosis and early management for children and young people. NICE guideline NG254. 2025.
• 2025National Institute for Health and Care Excellence. Suspected sepsis: recognition, diagnosis and early management for pregnant women. NICE guideline NG255. 2025.
• 2021Evans L, Rhodes A, Alhazzani W, et al. Surviving Sepsis Campaign: international guidelines for management of sepsis and septic shock 2021. Intensive Care Med. 2021;47(11):1181-1247.
• 2016Singer M, Deutschman CS, Seymour CW, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 2016;315(8):801-810.

Notes

• Primary outcome was in-hospital mortality by day 90; discharge alive before day 90 was treated as survival for the primary endpoint.
• Secondary outcome P values were adjusted using false discovery rate procedures; interpret secondary signals (including harms) within this multiplicity framework and the stepped-wedge temporal structure.

Overall Takeaway

SCREEN is a landmark large pragmatic stepped-wedge trial testing real-time ward-based electronic sepsis screening at scale. It suggests that embedding a qSOFA-triggered alert within an EHR-based response workflow can be associated with lower adjusted in-hospital mortality, but the small absolute effect, reliance on adjusted modelling, and concurrent signals of potential harms mean implementation should be cautious, monitored, and paired with stewardship and robust evaluation.

Bibliography

• 1Evans L, Rhodes A, Alhazzani W, et al. Surviving Sepsis Campaign: international guidelines for management of sepsis and septic shock 2021. Intensive Care Med. 2021;47(11):1181-1247.
• 2Arabi YM, Alsaawi A, AlZahrani M, et al. Electronic early notification of sepsis in hospitalized ward patients: a study protocol for a stepped-wedge cluster randomized controlled trial. Trials. 2021;22(1):695.
• 3Arabi YM, Vishwakarma RK, Al-Dorzi HM, et al; SCREEN Trial Group. Statistical analysis plan for the stepped wedge cluster randomized trial of electronic early notification of sepsis in hospitalized ward patients (SCREEN). Trials. 2021;22(1):828.
• 4Hemming K, Taljaard M, McKenzie JE, et al. Reporting of stepped wedge cluster randomised trials: extension of the CONSORT 2010 statement with explanation and elaboration. BMJ. 2018;363:k1614.
• 5Joshi M, Ashrafian H, Arora S, et al. Digital alerting and outcomes in patients with sepsis: systematic review and meta-analysis. J Med Internet Res. 2019;21(12):e15166.
• 6Fleuren LM, Klausch TLT, Zwager CL, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis. Intensive Care Med. 2020;46(3):383-400.