Phase 1 β Data & Cohort. Extract MIMIC-IV v2.2 via PostgreSQL (pgAdmin) for adult ICU sepsis patients; replicate the published cohort structure of 4,686 patients with 874 deaths (18.7% mortality) for development, and reserve eICU-CRD v2.0 (3,718 patients, 709 deaths) for external validation . Define sepsis using Sepsis-3 criteria; handle missing data with MICE random-forest imputation and cap outliers by IQR.
Phase 2 β Feature Engineering & Selection. Engineer 82 candidate features (min/max/mean of vitals and labs over first 24h): SpO2, SBP, MBP, HR, lactate, sodium, phosphate, WBC, platelets, creatinine, age, comorbidities. Apply Boruta (R, random-forest shadow features) to retain the ~20 confirmed predictorsβtop features were SpO2_min, SBP_min, Lactate_max, Age, and Sodium_max .
Phase 3 β Model Training. Train eight algorithms with 10-fold CV + GridSearchCV hyperparameter tuning: XGBoost, GBDT, SVM, Logistic Regression, Random Forest, Neural Network, Decision Tree, KNN. XGBoost is the anchor model based on the best evidence; use SMOTE-Tomek for class imbalance if needed .
Phase 4 β External Validation. Evaluate on eICU-CRD: target AUC β0.855 (vs internal 0.911, a ~6-point drop indicating acceptable transportability). Assess calibration curves, decision curve analysis (DCA), sensitivity/specificity at clinically chosen thresholds, and compare against SOFA, NEWS, MEWS baselines .
Phase 5 β Interpretability & Deployment. Compute SHAP values for global + per-patient explanation; expect SpO2_min, SBP_min, Age, Lactate_max as dominant drivers. Use SHAP dependence plots to detect nonlinear effects (e.g., SBP via restricted cubic splines). Package as a REST API with real-time EHR integration.
Confidence: moderate-to-high that this pipeline mirrors the best available published evidence; specific AUC values are study-dependent and may not reproduce exactly on new cohorts.
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.