ACCURACY OF ARTIFICIAL INTELLIGENCE IN PREDICTING SEPSIS FROM EMERGENCY LABORATORY PARAMETERS : A SYSTEMATIC REVIEW OF RANDOMISED CONTROLLED TRIALS AND PRIMARY STUDIES

Abstract

Introduction: Sepsis remains among the most frequently missed diagnoses in the emergency department (ED), and the diagnostic pathway is critically dependent on the emergency laboratory. Conventional tools such as qSOFA, SIRS, and MEWS/NEWS discriminate poorly at triage, while single biomarkers (lactate, CRP, PCT, MDW) have a ceiling of diagnostic performance when interpreted in isolation. Artificial intelligence (AI) and machine learning (ML) can integrate the multidimensional output of the emergency laboratory in real time, potentially converting routine analyte data into an actionable probability of sepsis. Methods: Systematic review following PRISMA 2020 and PRISMA-DTA. Primary studies (RCTs, prospective/retrospective cohorts, case-control, cross-sectional) of adult patients evaluated for suspected sepsis in the ED, ward, or ICU were included; systematic reviews/meta-analyses were excluded from characteristics tables. Risk of bias assessed with QUADAS-2 extended with PROBAST domains, and RoB 2 for randomised trials. Fourteen outcome domains were extracted; certainty of evidence rated with GRADE. Results: Of 4,191 records, 23 studies (2 RCTs, 4 prospective cohorts, 15 retrospective, 1 case-control, 1 cross-sectional; >3.4 million encounters) met criteria. Pooled AUROC for AI-based sepsis prediction was 0.874 (95% CI 0.856-0.891; I²=91.4%), with summary sensitivity 0.834 and specificity 0.812. Complete-blood-count-only models achieved AUROC 0.872 (95% CI 0.857-0.887), falling only modestly on external validation (0.805-0.845). AI significantly outperformed qSOFA (+0.191 AUROC), SIRS (+0.128), MEWS/NEWS (+0.171), and physician judgement (+0.110) (all p<0.001). Discrimination declined significantly on external validation (0.895 vs 0.842; Δ=-0.053; p=0.004), with one model failing catastrophically (0.82→0.47) in an out-of-distribution population. Lead time before clinical recognition ranged from 2.1 hours in the ED to 17.6 hours in the ICU; one model detected sepsis up to 7 days before admission. Lactate, creatinine, white blood cell count, platelets, bilirubin, and procalcitonin were the dominant laboratory contributors. In implementation studies, AI-assisted care was associated with reductions in in-hospital mortality (17-58%), length of stay (11-32%), and time to anti-infective administration, but randomised evidence remained sparse and one meta-analysis of randomised screening trials found no significant mortality benefit (OR 0.84; 95% CI 0.60-1.18). Risk of bias was low in 9 studies, unclear in 8, and high in 6. Discussion: The evidence consistently supports the hypothesis that AI models applied to emergency laboratory parameters discriminate sepsis substantially better than conventional scores or single biomarkers. The discriminatory signal resides largely in routine, inexpensive analytes; the dominant limitations are not discrimination but calibration, positive predictive value, and generalisability—the low prevalence of sepsis in an unselected ED population drives PPV as low as 0.18-0.29 even at excellent AUROC, and distribution shift between institutions erodes performance. Conclusion: AI applied to emergency laboratory parameters predicts sepsis with good-to-excellent discrimination (pooled AUROC 0.874), significantly and consistently exceeding qSOFA, SIRS, MEWS/NEWS, and physician judgement, and providing a clinically actionable lead time. Complete-blood-count-only models retain most of this accuracy and are immediately deployable in any laboratory performing automated haematology. Nonetheless, certainty of evidence is moderate for discrimination and low for patient-important outcomes because of substantial heterogeneity, incomplete external validation, and a scarcity of randomised implementation trials.