Variation in Antimicrobial Consumption in Critically Ill Adults

Project Overview

Antibiotics are potentially life-saving medications, but their inappropriate use is leading to alarming emergence of resistant organisms and deaths from antibiotic resistance. Although administration of antibiotics in the ICU is pivotal in the management of severe infections and sepsis, antibiotic guidelines differ between ICUs across institutions depending on patient factors, resistance patterns, accessibility, and clinical practice patterns.

This multi-site project aims to characterize the heterogeneity in antibiotic consumption across a diverse set of US medical healthcare centers using the CLIF (Critical Care Learning from Intensive Care and Feeding) consortium data.

Aim

To evaluate the between-hospital variation in prescription of antibiotics in critically ill adults.

Data Sources

Database: CLIF

CLIF Sites: No restriction

Study Period: 2018-2024

Setup

1. Install UV Package Manager

Mac/Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

2. Install Dependencies

uv sync

3. Run the Pipeline

Step 1: Generate ICU Cohort

This step creates the ICU cohort with all clinical features and SOFA scores.

uv run marimo edit code/01_cohort.py

Step 2: Calculate DOT and Antibiotic Metrics

This step performs all DOT, ASC, DASC, and AFD calculations, generates visualizations, and creates Table 1.

uv run marimo edit code/02_DOT.py

4. Configure Site

Rename clif_config_template.json to clif_config.json and update it with your site-specific configuration:

{
    "site": "your_site_name",
    "data_directory": "/path/to/your/clif/data",
    "filetype": "parquet",
    "timezone": "US/Central"
}

5. Prepare Reference Files

This file contains: - Antibiotic: Antibiotic names matching CLIF med_category values - Score: Spectrum coverage scores for each antibiotic

Required CLIF Tables

Table	Columns	Categories/Filters
adt	hospitalization_id, location_category, location_type, in_dttm, out_dttm	location_category == 'icu'
hospitalization	patient_id, hospitalization_id, age_at_admission, admission_dttm, discharge_dttm, discharge_category	-
patient	patient_id, sex_category, ethnicity_category, race_category, language_category, death_dttm	-
vitals	hospitalization_id, recorded_dttm, vital_category, vital_value	temp_c, map, spo2, weight_kg
labs	hospitalization_id, lab_result_dttm, lab_category, lab_value, lab_value_numeric	wbc, creatinine, platelet_count, po2_arterial, bilirubin_total
patient_assessments	hospitalization_id, recorded_dttm, assessment_category, numerical_value	gcs_total
medication_admin_continuous	All columns (including med_dose, med_dose_unit)	norepinephrine, epinephrine, phenylephrine, angiotensin, vasopressin, dopamine, dobutamine, milrinone, isoproterenol
medication_admin_intermittent	hospitalization_id, admin_dttm, med_category, med_route_category, med_dose, med_dose_unit	All antibiotics from antibiotic_spectrum_scoring.csv
respiratory_support	hospitalization_id, recorded_dttm, device_category, fio2_set	All device categories (NIPPV, High Flow NC, etc.)

Study Population

Inclusion Criteria

Adults (≥18 years old)
First (index) ICU admission within hospitalization_id (location_category == 'icu')
Medical ICU only (location_type == 'medical_icu')
ICU admission between 2018-2024
ICU discharge ≤2024 (excludes partial 2025 data)
ICU length of stay > 6 hours (0.25 days)

Exclusion Criteria

ICU stays ≤6 hours (0.25 days)
ICU discharge dates in 2025

Clinical Features Computed

All features are filtered to the ICU stay window (start_dttm to end_dttm):

Vital Signs

highest_temperature (°C) - Maximum temperature during ICU stay
lowest_temperature (°C) - Minimum temperature during ICU stay
lowest_map (mmHg) - Minimum mean arterial pressure during ICU stay

Laboratory Values

highest_wbc (10³/μL) - Maximum white blood cell count during ICU stay
highest_creatinine (mg/dL) - Maximum creatinine during ICU stay

Respiratory Support

NIPPV_ever (binary) - Non-invasive positive pressure ventilation usage
HFNO_ever (binary) - High-flow nasal cannula usage

Medications

vasopressor_ever (binary) - Any vasopressor usage during ICU stay
no_of_vasopressor (count) - Number of unique vasopressor categories used

Severity Scores

SOFA scores - Computed using clifpy ClifOrchestrator
- Includes respiratory, cardiovascular, hepatic, coagulation, renal, and neurological components

Outcomes

hospital_los_days - Hospital length of stay (days)
icu_los_days - ICU length of stay (days)
inpatient_mortality (binary) - Death during hospitalization
icu_mortality (binary) - Death during ICU stay

Methodology

Step 0: Create 24-Hour Windows

ICU "days" are defined as 24-hour windows starting from ICU admission time, not calendar days:

Window 0: First 24 hours after ICU admission (ICU Day 0)
Window 1: Second 24 hours (ICU Day 1)
Window N: Nth 24-hour period

Each window represents 1 patient-day.

Example: Patient admitted to ICU on 2024-01-15 at 08:00, discharged 2024-01-17 at 14:00 - Window 0: 2024-01-15 08:00 → 2024-01-16 08:00 (24 hours) - Window 1: 2024-01-16 08:00 → 2024-01-17 08:00 (24 hours) - Window 2: 2024-01-17 08:00 → 2024-01-17 14:00 (6 hours, capped at discharge) - Total: 3 windows = 3 patient-days

Step 1: Calculate DOT per Antibiotic

For each hospitalization_id and each antibiotic:

Count the number of 24-hour windows containing ≥1 dose of that antibiotic
If antibiotic is administered at any time during a window → count as 1 DOT for that antibiotic
Track antibiotic-free days (windows with zero antibiotic doses)

Formula: DOT = Number of windows with ≥1 dose of antibiotic

Output: Patient-level table with DOT per antibiotic per hospitalization

Step 2: Calculate Patient-Days (PD)

Calculate the total number of patient-days across all hospitalizations.

Formula: PD = Σ (number of 24-hour windows per hospitalization)

Example: If cohort has 10,000 hospitalizations with average ICU LOS of 5 days → PD = 50,000

Step 3: Calculate DOT per 1000 PD by Antibiotic

For each individual antibiotic, calculate standardized consumption rate.

Formula: DOT per 1000 PD = (Total DOT for antibiotic / Total PD) × 1000

Example: - Vancomycin total DOT = 5,000 days (across all hospitalizations) - Total PD = 50,000 patient-days - Vancomycin DOT per 1000 PD = (5,000 / 50,000) × 1000 = 100

Interpretation: For every 1000 patient-days in the ICU, Vancomycin is administered on 100 days.

Output: Antibiotic-level metrics table (one row per antibiotic)

Step 4: Calculate Overall DOT per 1000 PD (Cohort-Level)

Calculate overall antibiotic consumption across ALL antibiotics for the MICU cohort.

Formula: Overall DOT per 1000 PD = (Total DOT all antibiotics / Total PD) × 1000

Difference from Step 3: Step 3 calculates per individual antibiotic; Step 4 aggregates across all antibiotics.

Note: Since the cohort is restricted to Medical ICU only, location-type stratification is not performed.

Output: Cohort-level overall metrics

Step 5: Calculate Daily ASC per Window

For each 24-hour window (ICU Days 0-10), calculate Antibiotic Spectrum Coverage (ASC).

Formula: Daily ASC = Σ (spectrum_score for each antibiotic used in window)

IMPORTANT: This is NOT DOT × spectrum_score. The formula sums spectrum scores of all antibiotics used in a window.

Example: - Window 0 (ICU Day 0): Vancomycin (score=3.5) + Piperacillin-Tazobactam (score=4.2) - Daily ASC = 3.5 + 4.2 = 7.7 - Window 1 (ICU Day 1): Vancomycin only (score=3.5) - Daily ASC = 3.5 - Window 2 (ICU Day 2): No antibiotics - Daily ASC = 0

Output: Mean and standard deviation of daily ASC for windows 0-10 (for multi-site comparison)

Step 6: Calculate DASC per 1000 PD

Calculate Days of Antibiotic Spectrum Coverage (DASC) per 1000 patient-days.

Process:

For each hospitalization, sum daily ASC across ALL ICU windows → DASC
Calculate DASC per 1000 PD

Formula: DASC per 1000 PD = (Total DASC across all hospitalizations / Total PD) × 1000

Example: - Hospitalization with 5 ICU days: - Window 0: Daily ASC = 7.5 - Window 1: Daily ASC = 7.5 - Window 2: Daily ASC = 3.5 - Window 3: Daily ASC = 3.5 - Window 4: Daily ASC = 0 - DASC = 7.5 + 7.5 + 3.5 + 3.5 + 0 = 22.0

Output: - Overall: DASC per 1000 PD for entire cohort (all years combined) - By Year: DASC per 1000 PD for each year (2018-2024) for temporal trend analysis

Step 7: Calculate Antibiotic-Free Days (AFD)

Count the number of 24-hour windows without any antibiotic administration.

Formula: AFD Rate = (Antibiotic-free windows / Total windows) per hospitalization

Example: - Hospitalization with 5 ICU days: - Day 0: Antibiotics given - Day 1: Antibiotics given - Day 2: No antibiotics (antibiotic-free) - Day 3: No antibiotics (antibiotic-free) - Day 4: Antibiotics given - AFD Rate = 2 / 5 = 0.40 (40%)

Output: Mean and standard deviation of AFD rate across all hospitalizations

Step 8: Calculate Year-Based ASC Summary

Group daily ASC data by year (based on window start timestamp) to enable time series analysis.

Process: 1. Join daily ASC data with window timestamps 2. Extract year from window_start 3. Calculate summary statistics per year: mean, SD, median, min, max

Output: Year-based ASC summary table (2018-2024)

Step 9: Generate Visualizations

Create publication-ready plots for multi-site comparison:

ASC by Year Plot: Smoothed spline with ±1 SD band showing temporal trends
ASC by ICU Day Plot: Smoothed spline with ±1 SD band for Days 0-10 post-admission

Output: PNG files saved to RESULTS_UPLOAD_ME/

Step 10: Generate Table 1 Summary Statistics

Generate comprehensive summary statistics table including:

Demographics: Age, sex, race/ethnicity, BMI, ICU/hospital LOS
Clinical Characteristics: Vitals, SOFA scores, interventions (vasopressors, NIPPV, HFNO), labs
Antibiotic Metrics: Top 15 antibiotics by DOT per 1000 PD, overall DOT, daily ASC scores (Days 0-10), DASC per 1000 PD, AFD rate
Outcomes: Total patient-days, mortality rates

Output: - table1_summary.json (machine-readable for aggregation) - table1_summary.csv (human-readable for review)

Key Metrics Definitions

Metric	Abbreviation	Definition
Days of Therapy	DOT	Number of 24-hour windows where patient receives ≥1 dose of antibiotic
Patient-Days	PD	Total number of 24-hour windows across all hospitalizations
Antibiotic Spectrum Coverage	ASC	Sum of spectrum scores for all antibiotics used in a 24-hour window
Days of Antibiotic Spectrum Coverage	DASC	Sum of daily ASC values across all ICU windows for a hospitalization
Antibiotic-Free Days	AFD	Number of 24-hour windows without any antibiotic administration
SOFA Score	SOFA	Sequential Organ Failure Assessment score (computed via clifpy ClifOrchestrator)

Output Files

PHI_DATA/ (Patient-Level - DO NOT SHARE)

These files contain patient-level data and should NOT be shared outside your institution:

cohort_icu_first_stay.parquet - Complete cohort with all features and SOFA scores
dot_hospital_level.parquet - DOT per antibiotic per hospitalization (wide format)
daily_asc_patient_level.parquet - Daily ASC for each window per hospitalization
afd_patient_level.parquet - Antibiotic-free days per hospitalization

RESULTS_UPLOAD_ME/ (Safe to Share - Summary Statistics)

These files contain only aggregate summary statistics and are safe to share for multi-site comparison:

dot_antibiotic_level.csv - DOT per 1000 PD by antibiotic (cohort-level)
dot_cohort_level.csv - Overall DOT metrics (all antibiotics combined)
daily_asc_summary.csv - Mean/SD of daily ASC for Days 0-10 (for plotting)
dasc_overall.csv - DASC per 1000 PD (all years combined)
dasc_by_year.csv - DASC per 1000 PD by year (2018-2024)
asc_by_year_summary.csv - Year-based ASC trends for time series analysis
afd_summary.csv - Antibiotic-free days summary statistics
table1_summary.json - Complete Table 1 in JSON format (machine-readable)
table1_summary.csv - Complete Table 1 in CSV format (human-readable)
asc_by_year_plot.png - Visualization of ASC trends over years
asc_by_window_plot.png - Visualization of ASC by ICU day (Days 0-10)

Data Sharing Instructions

After Successful Pipeline Completion

Once both scripts (01_cohort.py and 02_DOT.py) have completed successfully and all output files are generated in RESULTS_UPLOAD_ME/, contact the RUSH coordinating site to obtain the BOX folder link for uploading your results.

Contact: Vaishvik_Chaudhari@rush.edu

What to Upload

IMPORTANT: Upload ONLY files from the RESULTS_UPLOAD_ME/ folder. DO NOT upload files from PHI_DATA/ folder as they contain patient-level data.

Upload Checklist: - [ ] dot_antibiotic_level.csv - [ ] dot_cohort_level.csv - [ ] daily_asc_summary.csv - [ ] dasc_overall.csv - [ ] dasc_by_year.csv - [ ] asc_by_year_summary.csv - [ ] afd_summary.csv - [ ] table1_summary.json - [ ] table1_summary.csv - [ ] asc_by_year_plot.png - [ ] asc_by_window_plot.png

Name		Name	Last commit message	Last commit date
Latest commit History 24 Commits
VACCIA		VACCIA
code		code
.gitignore		.gitignore
.python-version		.python-version
LICENSE		LICENSE
README.md		README.md
antibiotic_spectrum_scoring.csv		antibiotic_spectrum_scoring.csv
clif_config_template.json		clif_config_template.json
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Folders and files

Latest commit

History

Repository files navigation

Variation in Antimicrobial Consumption in Critically Ill Adults

Project Overview

Aim

Data Sources

Setup

1. Install UV Package Manager

2. Install Dependencies

3. Run the Pipeline

Step 1: Generate ICU Cohort

Step 2: Calculate DOT and Antibiotic Metrics

4. Configure Site

5. Prepare Reference Files

Required CLIF Tables

Study Population

Inclusion Criteria

Exclusion Criteria

Clinical Features Computed

Vital Signs

Laboratory Values

Respiratory Support

Medications

Severity Scores

Outcomes

Methodology

Step 0: Create 24-Hour Windows

Step 1: Calculate DOT per Antibiotic

Step 2: Calculate Patient-Days (PD)

Step 3: Calculate DOT per 1000 PD by Antibiotic

Step 4: Calculate Overall DOT per 1000 PD (Cohort-Level)

Step 5: Calculate Daily ASC per Window

Step 6: Calculate DASC per 1000 PD

Process:

Step 7: Calculate Antibiotic-Free Days (AFD)

Step 8: Calculate Year-Based ASC Summary

Step 9: Generate Visualizations

Step 10: Generate Table 1 Summary Statistics

Key Metrics Definitions

Output Files

PHI_DATA/ (Patient-Level - DO NOT SHARE)

RESULTS_UPLOAD_ME/ (Safe to Share - Summary Statistics)

Data Sharing Instructions

After Successful Pipeline Completion

What to Upload

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 1

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages