VAP Onset Prediction - Complete Analysis Pipeline

Overview

This repository contains a comprehensive analysis pipeline for predicting Ventilator-Associated Pneumonia (VAP) onset in mechanically ventilated ICU patients. The project is organized into three main cohorts:

1. Script - Main cohort where models are trained (foundation)
2. MIMIC - External cohort validation using MIMIC-IV data
3. Amikacin - External cohort validation using French Amikacinhal data

Project Structure

PredictingVAPExternal/
├── script/              # Main cohort (Script data)
│   ├── data/           # Script cohort datasets
│   ├── src/            # Source code (modeling, figures, tables)
│   ├── scripts/        # Bash scripts to run pipelines
│   ├── final_models/   # Final trained models (Random CV)
│   ├── figures/        # Generated figures
│   └── Tables/         # Generated tables
├── mimic/               # MIMIC-IV external cohort
│   ├── data/           # MIMIC data (raw, processed, labeled)
│   ├── src/            # MIMIC-specific analysis code
│   ├── scripts/        # Bash scripts for MIMIC analysis
│   └── results/        # MIMIC results (figures, tables, modeling)
├── amikacin/            # Amikacin external cohort
│   ├── data/           # Amikacin data (raw, processed, labeled)
│   ├── src/            # Amikacin-specific analysis code
│   ├── scripts/        # Bash scripts for Amikacin analysis
│   └── results/        # Amikacin results (figures, tables, modeling)
└── requirements.txt     # Python dependencies

Execution Order

Important: Run analyses in this order:

1. Script → Train models on main cohort
2. MIMIC → Validate on MIMIC-IV external cohort (uses Script models)
3. Amikacin → Validate on French Amikacinhal cohort (uses Script models)

Quick Start

# 1. Script (Main Cohort)
cd script/scripts
./run_all.sh

# 2. MIMIC (External Cohort)
cd ../../mimic/scripts
./run_figures.sh
./run_tables.sh
./run_modeling.sh

# 3. Amikacin (External Cohort)
cd ../../amikacin/scripts
./run_france_all.sh

Or run all cohorts sequentially:

./run_all_cohorts.sh

Detailed Documentation

• Script Cohort - Main cohort modeling, figures, and tables
• MIMIC Cohort - MIMIC-IV external validation
• Amikacin Cohort - French Amikacinhal external validation

Environment Setup

Prerequisites

• Python 3.8+
• Virtual environment (recommended)

Installation

# Create virtual environment
python3 -m venv vap_onset
source vap_onset/bin/activate  # macOS/Linux
# or
vap_onset\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

Key Features

• Multiple Model Types: Random Forest, XGBoost, Logistic Regression, LSTM
• Multiple Prediction Windows: 3, 5, and 7-day VAP prediction windows
• Cross-Validation Strategies: Random CV (final), Temporal CV (archived)
• Comprehensive Analysis: Figures, tables, feature importance, timeline analysis
• External Validation: Validated on MIMIC-IV and French Amikacinhal cohorts

Output Locations

Script Cohort

• Models: script/final_models/results_random_cv_20251031_115429/
• Figures: script/figures/
• Tables: script/Tables/

MIMIC Cohort

• Figures: mimic/results/figures/
• Tables: mimic/results/tables/
• Modeling: mimic/results/modeling/

Amikacin Cohort

• Figures: amikacin/results/figures/
• Tables: amikacin/results/tables/
• Modeling: amikacin/results/modeling/

Notes

• All scripts include error handling and continue even if some steps fail
• Output directories are timestamped to prevent overwriting
• Each cohort has its own configuration file (scripts/config.sh)
• See individual cohort READMEs for detailed usage instructions