A fully reproducible real-world evidence (RWE) study examining 28-day mortality outcomes in ICU sepsis patients treated with vasopressors versus fluid resuscitation alone, using MIMIC-IV electronic health record data from over 65,000 ICU admissions at Beth Israel Deaconess Medical Center.

  • Tools: Python, SQL, Snowflake, dbt, Docker, lifelines, scikit-learn, SHAP
  • Methods: Propensity score matching (PSM), Kaplan-Meier survival analysis, Cox proportional hazards, SHAP TreeExplainer
  • Data: MIMIC-IV v3.1 (PhysioNet) — hosp and ICU modules
  • Highlights: End-to-end dbt pipeline from raw EHR tables to analysis-ready cohort; 1:1 nearest-neighbor PSM with Austin (2011) caliper; Schoenfeld residuals PH assumption testing; treatment effect heterogeneity via SHAP interaction values
  • GitHub: View Repository

Overview

Sepsis kills approximately 270,000 Americans annually, and treatment decisions in the first hours of ICU admission are among the highest-stakes in medicine. This project asks a question clinicians actively debate: among sepsis patients in the ICU, does early vasopressor use lead to different mortality outcomes than fluid resuscitation alone? Using MIMIC-IV — a de-identified EHR dataset from over 65,000 ICU admissions at Beth Israel Deaconess Medical Center — I built a fully reproducible RWE pipeline from raw data to clinical findings. Pipeline: Raw MIMIC-IV tables are ingested into Snowflake and transformed through a documented dbt project into an analysis-ready cohort. Staging models clean and standardize the source tables. Mart models encode the study protocol: sepsis ICD-10 inclusion criteria, 24-hour treatment windows, index date logic, and organ dysfunction covariates. Causal inference: Because vasopressor patients are sicker at baseline — higher lactate, worse renal function, more hemodynamic instability — a naive comparison is meaningless. I applied 1:1 nearest-neighbor propensity score matching using a logistic regression on 11 covariates, following Austin (2011) caliper recommendations. Post-matching standardized mean differences fell below 0.1 for all covariates. Survival analysis: Kaplan-Meier curves and a Cox proportional hazards model on the matched cohort estimate the adjusted treatment effect. The proportional hazards assumption was verified via Schoenfeld residuals. A sensitivity analysis comparing crude vs adjusted hazard ratios quantifies the confounding removed by matching.

Figure: Kaplan-Meier survival curves for vasopressor vs fluid resuscitation cohorts (28-day follow-up, propensity score matched)

Interpretability: A gradient boosted classifier with SHAP TreeExplainer characterizes non-linear feature effects and treatment effect heterogeneity — revealing which patient subgroups show the strongest association between vasopressor use and mortality outcomes.

Figure: SHAP summary plot — feature contributions to vasopressor treatment effect by patient subgroup