Beyond Composite Indices: Comprehensive Social Determinants Improve Heart Failure Readmission Prediction

This repo contains the code for the article Beyond Composite Indices: Comprehensive Social Determinants Improve Heart Failure Readmission Prediction.

Requirements

Heart failure 30-day hospital readmission prediction:

python 3.9
imblearn==0.0
joblib==1.2.0
numpy==1.24.4
pandas==2.0.0
pymongo==4.7.0
scikit_learn==1.4.2
shap==0.45.0
tqdm==4.65.0
xgboost==1.7.6

Datasets

The social determinants of health (SDOH) datasets used in this study can be found below:

Dataset	Number of SDOH variables Used
AHRQ SDOHD	760

Reproducibility

1. Prepare patient data, gather SDOH data and merge /data/

Run

/data/patient_inclusion.Rmd

to apply study inclusion, exclusion criteria.

Then run

/data/merge_SDOH.Rmd

to merge SDOH data with patients.

2. Heart Failure (HF) Readmission Prediction

The patient dataset is unavailable due to privacy reasons --- however the following commands demonstrate the steps we used to train and evaluate binary classification models (using clinical and public SDOH data):

To train binary classification models on HF 30-day hospital readmission prediction (in file, choose classification algorithm, features):

python classification_driver_nestKfold.py

To analyze results of the HF models:

python analyze_classification_perf_results.py

To analyze fairness of the HF models:

python fairness_analyze_results.py

To get feature importance information (from trained XGBoost models):

python analyze_XGB_SHAP.py

3. Code to generate all tables and plots

First, pull readmission prediction model results from your local MongoDB collections:

• To gather prediction performance values:

    python /scripts/analyze_classification_perf_results.py

• To gather prediction fairness values:

    python /scripts/fairness_analyze_results.py 

• To gather feature importance values for XGBoost models:

    python /scripts/analyze_XGB_SHAP.py

• Then, to generate the patient characteristics table, use:

    /data/table1_generate.Rmd

• To tabulate model performance and fairness run:

    /data/calculate_HF_performance.Rmd

• To tabulate feature importance run:

    /data/plot_HF_SHAP.Rmd

Note that all SDOH features used from AHRQ SDOHD can be found below in Related Documents.

All Related Documents:

• Expanded SDOH variables used from AHRQ SDOHD

Rates of missingness for all expanded SDOH variables (i.e., from AHRQ SDOHD) can be found in summary_statistics/

• TotalCohort_missing_rates_by_race.csv
• TotalCohort_missing_rates_by_readmission.csv
• TotalCohort_missing_rates_by_readmission_black.csv
• TotalCohort_missing_rates_by_readmission_white.csv

Folder Structure

data/
|-- adi-download-2020-tract/
|-- sdi-download-2019-tract/
|-- feat_base.json
|-- feat_column.json
|-- subgroup_cols_fast.json
|-- 2010-18-all-granularities-AHRQ-dict.xlsx
|-- count_features.py
|-- num_unique_SDOH_features.py
|-- calculate_HF_performance.Rmd
|-- merge_SDOH.Rmd
|-- patient_inclusion.Rmd
|-- plot_HF_SHAP.Rmd
|-- table1_generate.Rmd
|-- data_cleaners.R
|-- gen_chars.R

scripts/
|-- analyze_classification_perf_results.py
|-- analyze_XGB_SHAP.py
|-- classification_driver_nestKfold.py
|-- evalHelper.py
|-- fairness_analyze_results.py
|-- fake_patient_data.csv

summary_statistics/
|-- AHRQ_used_county_metadata.csv
|-- AHRQ_used_tract_metadata.csv
|-- missing_rates_final_allstates_modelinput.csv
|-- TotalCohort_missing_rates_by_race.csv
|-- TotalCohort_missing_rates_by_readmission_black.csv
|-- TotalCohort_missing_rates_by_readmission_white.csv
|-- TotalCohort_missing_rates_by_readmission.csv
|-- domains_lists.json