Additive Predictive Value of the Sepsis ImmunoScore ® : A Multicenter, Prospective Evaluation of AI-Enabled Prediction of Sepsis Risk and Patient Outcomes 

Technical Summary: Additive Predictive Value of the Sepsis ImmunoScore®

Problem Statement
Early identification of sepsis in emergency and acute care settings remains a significant clinical challenge due to the nonspecific nature of initial symptoms, which often complicates risk assessment and decision-making regarding antimicrobial initiation, monitoring levels, and intensive care unit (ICU) disposition. While numerous electronic health record (EHR)-based decision-support tools exist, many rely on rule-based triggers derived from vital signs or laboratory abnormalities. These systems frequently generate high rates of false alerts, leading to clinician override, alert fatigue, and limited impact on actual patient outcomes. Although artificial intelligence (AI) and machine learning (ML) approaches are emerging, few have demonstrated real-world clinical utility in meaningfully enhancing decision-making beyond simply increasing alert volume. There is a need for tools that can provide additive predictive value to provider judgment, particularly in identifying occult risk and stratifying patients for critical care utilization.

Methodology
The authors conducted a multicenter, prospective observational study across three U.S. health systems (Carle Foundation Hospital, Corewell Health, and Mercy Hospitals East Community) between April 2024 and August 2025. The study enrolled 445 adults (≥18 years) with suspected infection, defined by a clinical decision to order a first blood culture in the ED or ICU.

• Intervention/Tool: The study evaluated the Sepsis ImmunoScore®, an FDA-authorized, pathogen-agnostic AI/ML tool. The algorithm synthesizes up to 22 clinical variables and biomarkers (including C-reactive protein [CRP] and procalcitonin [PCT]) to generate a risk score (1–100) and corresponding risk categories (Low, Medium, High, Very High). The model utilizes SHAP (Shapley Additive exPlanations) values for transparency.
• Data Collection:
  • Provider Assessment: Clinicians completed a standardized six-question survey within 3.5 hours of the blood culture order, blinded to the Sepsis ImmunoScore results. Surveys assessed probabilities for infection, Sepsis-3 risk (within 24 hours), and risk of organ support (within 24 hours).
  • Sepsis ImmunoScore Calculation: Remnant lithium-heparin plasma samples were retrieved for retrospective calculation of the score if collected within a specific window relative to the blood culture order. CRP and PCT were measured centrally.
  • Endpoints: The primary endpoint was the presence of Sepsis-3 within 24 hours of the blood culture order. Secondary endpoints included ICU transfer, vasopressor initiation, assisted ventilation, critical care interventions, and 28-day mortality.
• Statistical Analysis: Discriminative performance was assessed using the Area Under the Receiver Operating Characteristic Curve (AUROC). Additive value was tested using Likelihood Ratio Tests (LRT) and Continuous Net Reclassification Improvement (NRI). Risk stratification was analyzed via cross-classification of provider and Sepsis ImmunoScore categories, utilizing Cochran–Mantel–Haenszel (CMH) tests.

Key Contributions and Results
The study analyzed 401 patients with valid Sepsis ImmunoScore results, of whom 103 (25.7%) met Sepsis-3 criteria.

1. Improved Discrimination:

   • The combined Provider + Sepsis ImmunoScore model achieved an AUROC of 0.83 (95% CI, 0.78–0.88) for predicting Sepsis-3 within 24 hours.
   • This represented a significant improvement over Provider Alone assessment (AUROC 0.68; ΔAUC = +0.15; p < 0.001).
   • The Sepsis ImmunoScore alone achieved an AUROC of 0.81.
   • Similar additive improvements were observed for all secondary endpoints (ICU transfer, vasopressor use, critical care, and mortality), with Delta AUCs ranging from +0.03 to +0.06.

2. Statistical Significance of Additive Value:

   • Likelihood Ratio Tests confirmed that including the Sepsis ImmunoScore significantly improved model fit beyond provider assessment alone (χ² = 86.44; p < 0.001 for Sepsis-3).
   • Continuous Net Reclassification Improvement (NRI) was positive and statistically significant for all endpoints (NRI = 0.98 for Sepsis-3), indicating the combined model correctly reclassified events upward and non-events downward relative to provider assessment alone.

3. Risk Stratification and Occult Risk:

   • Cross-classification analysis revealed a monotonic gradient of risk within every provider-assessed stratum (Low, Medium, High/Very High).
   • Notably, among patients assessed as Low Risk by providers, those classified as High/Very High by the Sepsis ImmunoScore had a Sepsis-3 event rate of 25.7%, compared to 3.8% for those classified as Low by both.
   • Within the provider High/Very High risk group, patients with a High/Very High Sepsis ImmunoScore had a 55.7% sepsis rate, compared to 13.3% for those with a Low Sepsis ImmunoScore (a 4.2-fold gradient).
   • A High/Very High Sepsis ImmunoScore was associated with significantly higher odds of Sepsis-3 across all provider strata (Adjusted OR 6.45; p < 0.001).

4. Consistency:

   • The tool demonstrated consistent performance across the three sites, whereas unaided provider discrimination varied widely (AUROC 0.57 to 0.83).
   • The Sepsis ImmunoScore's performance remained stable (AUROC 0.80) even in the broader ED population where clinician surveys were not completed.

Significance and Claims
The paper claims that the Sepsis ImmunoScore consistently enhances clinician assessment by improving discrimination and uncovering "occult risk" within provider-defined categories. The authors posit that the tool captures complementary, biology-based risk factors (via biomarkers and multivariate signals) that may not be readily apparent to clinicians in time-pressured environments, particularly when early signs are subtle.

The study supports the Sepsis ImmunoScore's potential as an EHR-integrated decision-support tool that aids in early sepsis identification and risk stratification. The authors emphasize that the tool is designed to enhance, rather than replace, clinical judgment. It functions as a rapid, quantitative, and explainable signal that can reinforce a clinician's decision or prompt re-evaluation when the biomarker-based risk diverges from the clinical appearance.

The authors explicitly state that this study establishes predictive value but does not prove that access to the score changes provider behavior or patient outcomes. They conclude that prospective interventional studies are warranted to determine the impact of real-time, Sepsis ImmunoScore-guided decision support on clinical care and patient outcomes. The study acknowledges limitations, including its observational design, potential selection bias regarding missing data (where the sickest patients were more likely to lack valid scores), and the fact that clinician surveys were optional.