Skip to content

Latest commit

 

History

History
128 lines (94 loc) · 3.63 KB

PACE.md

File metadata and controls

128 lines (94 loc) · 3.63 KB

PACE Framework

This document describes the project design using the PACE framework (Plan → Acquire → Construct → Execute), providing a concise and structured　overview of how this clinical data science project was conceived and implemented.

P — Plan

Research Objective

To evaluate whether pre-ICU exposure to RAAS inhibitors (ACE inhibitors and angiotensin receptor blockers) is associated with clinical outcomes among ICU-admitted patients with COPD.

Primary Question

Among adult ICU patients with COPD in MIMIC-IV:

  • Is pre-ICU RAAS exposure associated with differences in in-hospital mortality?

Secondary Questions

  • Do ACE inhibitors and ARBs show similar or distinct associations?
  • Are results robust to extended covariate adjustment?
  • Does the association remain stable across modeling strategies?

Study Design

  • Retrospective observational cohort study
  • Real-world evidence using MIMIC-IV v3.1
  • ICU stay–level analysis with strict temporal ordering

A — Acquire

Data Source

  • MIMIC-IV v3.1
  • Accessed via BigQuery

Core Tables Utilized

  • ICU stays: mimiciv_icu.icustays
  • Diagnoses: mimiciv_hosp.diagnoses_icd
  • Patients: mimiciv_hosp.patients
  • Admissions: mimiciv_hosp.admissions
  • Prescriptions: mimiciv_hosp.prescriptions

Cohort Acquisition Strategy

  1. Build a standardized adult ICU cohort (age ≥ 18)
  2. Restrict to first ICU stay per hospital admission
  3. Identify COPD admissions using ICD-9 and ICD-10 codes
  4. Preserve ICU-stay–level granularity throughout

All cohort extraction and exposure derivation steps are implemented entirely in SQL to ensure reproducibility.

C — Construct

Exposure Definition

RAAS exposure is defined strictly before or at ICU admission using inpatient prescription orders. This does not directly capture outpatient chronic use and is intended as a proximal pre-ICU exposure proxy.

Exposure variables include:

  • acei_pre_icu
  • arb_pre_icu
  • raas_any_pre_icu
  • Monotherapy and mutually exclusive exposure groups

Detailed exposure logic is derived from medication names in physionet-data.mimiciv_3_1_hosp.prescriptions, aggregated at the ICU-stay level.

Baseline Covariates

Constructed prior to outcome modeling:

  • Demographics (age, sex)
  • Comorbidities
  • Severity indicators (e.g., SOFA where available)
  • ICU and hospital length of stay metrics

Data Integrity Principles

  • One row per ICU stay
  • Explicit handling of missing exposure as non-exposure
  • No Python-side merging required for cohort construction
  • SQL pipelines serve as the single source of truth

E — Execute

Modeling Strategy

Primary analyses:

  • Kaplan–Meier survival curves
  • Cox proportional hazards models

Model progression:

  1. Unadjusted RAAS exposure model
  2. ACEi vs ARB subgroup comparison
  3. Extended covariate Cox model

Model Diagnostics

  • Proportional hazards assessment
  • Forest plots with confidence intervals
  • Sensitivity to covariate inclusion

Outputs

  • Hazard ratios with 95% confidence intervals
  • Survival curves
  • Interpretable, publication-ready figures
  • Analysis-ready datasets for reproducibility

Reproducibility & Governance

  • All datasets generated via versioned SQL scripts (/sql)
  • All analyses documented in modular notebooks (/notebooks)
  • No PHI stored or committed
  • Credential files excluded via .gitignore

Intended Use

This project is designed as:

  • A clinical data science portfolio project
  • A demonstration of end-to-end real-world evidence workflow
  • A reproducible template for ICU-based observational studies

No clinical or medical recommendations are implied.