Prioritizing Sepsis Anticoagulation with the SMART Score

• SepsisFormer (transformer) reported strong prognostic discrimination (AUC 0.9301; sensitivity 0.9346) using 36 predictors.

• The derived SMART score—using only age plus seven routine coagulation/inflammation labs—stratified patients into four risk tiers (Mild to Dangerous) with stepwise observed mortality.

• SMART achieved AUC 0.7360 in the local ICU (outperforming SOFA) but performance was lower (~0.65–0.68) in larger public cohorts like MIMIC-IV.

• In an observational analysis, patients in SMART moderate and severe strata receiving ≥3 days of heparin showed reduced 28-day mortality (HR 0.60 and 0.57, respectively). This is a hypothesis-generating signal requiring safety monitoring, not a directive for automatic treatment.

• Pooled four retrospective ICU cohorts (MIMIC‑III + eICU‑CRD, MIMIC‑IV, and a local ICU) totaling 12,408 adult Sepsis‑3 patients for model development and validation.

• Trained SepsisFormer on 36 routine predictors using domain-adaptive generation (MMID‑SMOTE) to address cross-site heterogeneity.

• Derived the simplified SMART score by discretizing inputs and fitting LASSO logistic regression. Defined four risk tiers based on points: Mild [4–9), Moderate [9–13), Severe [13–17), and Dangerous [17–20].

• Identified two subphenotypes (CIS1 and CIS2) using unsupervised clustering and evaluated heterogeneous effects of heparin exposure (defined as ≥3 consecutive days) using Cox models.

• SepsisFormer (36 predictors) showed strong discrimination (AUC 0.9301), outperforming standard ML baselines.

• The simplified SMART scorecard retained prognostic signal but with reduced discrimination in external public datasets (AUCs ~0.65–0.67). However, in the local ICU, it achieved AUC 0.7360 and outperformed SOFA, qSOFA, APACHE II, and SIRS.

• Unsupervised clustering identified two reproducible subphenotypes. CIS2 was characterized by a specific coagulation–inflammation signature (higher APTT/INR, higher WBC, lower platelets) and higher mortality rates.

• Transcriptomic analyses provided orthogonal biological support, highlighting genes related to coagulation and inflammation (e.g., STAT5B, MTHFR, HPSE).

• In the observational analysis (MIMIC-IV), heparin use was associated with reduced 28-day mortality in SMART Moderate (HR 0.60, 95% CI 0.48–0.75) and Severe (HR 0.57, 95% CI 0.46–0.70) groups. No statistically significant benefit was seen in the Mild or Dangerous strata.

• Short term: Implement SMART as an EHR-calculated pharmacist flag to standardize anticoagulation review for Moderate/Severe patients. Do not auto-initiate anticoagulation based solely on this study.

• Informatics: Confirm LOINC mapping and lab turnaround times for the seven required labs. Define rules for missing values and run a silent-mode validation to measure alert volume and local calibration.

• Safety: Require pharmacist triage for any pilot, incorporating bleeding-risk checks, DDI screening, reversal plans, and predefined stop rules. Track bleeding events and mortality weekly.

• Business/Governance: Use this tool to triage stewardship resources. Before operational changes, request vendor calibration data and estimate pharmacist FTE requirements for monitoring.

Strengths:

• Multi-center development and external validation across public and local cohorts, utilizing domain adaptation to handle data heterogeneity.

• Creation of a clinically interpretable SMART score that uses only low-cost, routine laboratory measures, lowering barriers to local replication and real-world testing.

• Multi-view explainability (SHAP, clustering, transcriptomics) supports the biological plausibility of the identified subphenotypes.

Limitations:

• Anticoagulant findings are observational and retrospective. Exposure was defined as ≥3 consecutive days without granular detail on dose, product type, or indication, leaving room for confounding by indication and selection bias.

• SMART showed modest discrimination in large public cohorts (AUC < 0.70), suggesting local calibration is essential.

• The study lacks systematic reporting of bleeding and safety outcomes, which is critical for any anticoagulation intervention.

• Generalizability to non-ICU populations or earlier care pathways is unproven.

SMART is a well-constructed, EHR-ready sepsis risk flag with an observational anticoagulation signal. It justifies local validation and a safety-monitored, pharmacist-first pilot, but protocolized heparin practice should not change without prospective trials.

• Sepsis is biologically and clinically heterogeneous; existing prognostic scores often lack fine-grained stratification.

• SMART relies on a small panel of routinely available labs (APTT, INR, WBC, neutrophils, lymphocytes, monocytes, platelets) plus age. This allows for reproducible phenotyping using data already present in the EHR.

• For pharmacy teams, this score provides a transparent method to prioritize anticoagulation reviews and stewardship, focusing limited pharmacist time on patients where treatment effects may differ.