Procurement Biopsy Pathomics for Kidney Allograft Outcome Prediction

Multimodal ComPRePS: Integrating High-dimensional Procurement Biopsy Pathomics and Clinical Data for Prediction of Allograft Outcomes

This repository contains code and models for the ComPRePS (Comprehensive Prediction System) framework, which predicts kidney transplant outcomes using machine learning methods integrating procurement biopsy pathomics and clinical data. The analysis focuses on two primary outcomes:

1. 1-year estimated Glomerular Filtration Rate (eGFR) - continuous outcome
2. Delayed Graft Function (DGF) - binary outcome

Repository Contents

procurement-biopsy-pathomics-ml/
├── R/                          # R scripts for analysis
│   ├── Create_Train_Test_Exclusion_Data.R
│   ├── Binary_Internal_Validation_Metrics.R
│   ├── Continuous_Internal_Validation_Metrics.R
│   ├── Model_Saving_and_Test_Performance_Metrics_with_Exclusion.R
│   ├── Visualize_MSE_or_C_training.R
│   ├── Helper_CV_Functions.R
│   └── CKD_Helper_Functions.R
├── models/                     # Saved trained models
│   ├── random_forest_model_optimal_eGFR.rds
│   ├── random_forest_model_optimal_DGF.rds
│   └── KDPI_model_eGFR.rds
├── data/                       # Dataset
│   └── Renal_Data.csv
└── README.md

Data and Models

Dataset

• File: data/Renal_Data.csv
• Download: Google Drive Link
• Contains donor clinical factors, pathomic features, and transplant outcomes
• Features include: donor demographics, clinical variables, KDPI scores, and image-derived features

Pre-trained Models

Three models are available for download:

1. random_forest_model_optimal_eGFR.rds (734 KB)

   • Download: Google Drive Link
   • Random forest model for predicting 1-year eGFR
   • Trained on MRMR-selected features with optimized hyperparameters

2. random_forest_model_optimal_DGF.rds (75 KB)

   • Download: Google Drive Link
   • Random forest model for predicting delayed graft function
   • Trained on MRMR-selected features with optimized hyperparameters

3. KDPI_model_eGFR.rds (7 KB)

   • Download: Google Drive Link
   • Baseline linear regression model using KDPI alone
   • Used for comparison to machine learning models

Requirements

R Packages

install.packages(c(
  "randomForest",
  "permimp",
  "ggplot2",
  "pROC",
  "dplyr",
  "tidyr",
  "gridExtra",
  "stringr",
  "caret",
  "mRMRe",
  "foreach",
  "doParallel"
))

R Version

• Tested on R version 4.0 or higher

Related Publication

This repository contains code and models for:

Rodrigues, L., Paul, A.S., Rubin, J., Magdy, H., Gupta, A., Pardinhas, C., Pimenta, C., Fernandes, B., Simões, I., Sousa, V., Figueirerdo, A., Alves, R., Zee, J., Sarder, P. (in preparation). Multimodal ComPRePS: Integrating High-dimensional Procurement Biopsy Pathomics and Clinical Data for Prediction of Allograft Outcomes.

Repository: https://github.com/jeremysrubin/procurement-biopsy-pathomics-ml

Workflow Overview

1. Data Preparation and Feature Selection

Script: R/Create_Train_Test_Exclusion_Data.R

This script:

• Performs train/test split (80/20) stratified by outcome
• Removes features with missing values
• Applies MRMR (minimum Redundancy Maximum Relevance) feature selection
• Handles multicategorical variables
• Saves processed datasets for downstream analysis

Key function:

generate_training_testing_exclusion_data(
  input.file.name = "data/Renal_Data.csv",
  num.MRMR.features = 30,  # Number of features to select
  outcome = "C",            # "C" for eGFR, "B" for DGF
  train.sub.name = "train_indices.csv",
  test.sub.name = "test_indices.csv",
  make.split = TRUE         # TRUE for first run with DGF outcome
)

Important: Run this function first with DGF outcome (outcome = "B", make.split = TRUE) to create the train/test split, then use the same split for eGFR prediction.

2. Model Training and Internal Validation

Scripts:

• R/Binary_Internal_Validation_Metrics.R - for DGF prediction
• R/Continuous_Internal_Validation_Metrics.R - for eGFR prediction

These scripts:

• Loop over 1–100 MRMR-selected features automatically
• Perform 5-fold cross-validation with 100 bootstrap resamples (parallelized across available CPU cores)
• Tune random forest hyperparameters (nodesize: 1, 5, 9)
• Compare random forest against KDPI baseline
• Generate internal validation metrics and save one results file per MRMR feature count

Helper functions:

• R/Helper_CV_Functions.R - Cross-validation utilities
• R/CKD_Helper_Functions.R - CKD staging and visualization

3. Visualization of Internal Validation Results

Script: R/Visualize_MSE_or_C_training.R

Visualizes internal validation performance, showing separate lines for each machine learning model and KDPI baseline. Highlights the optimal number of MRMR-selected features across all algorithms (the point on the graph with the lowest MSE or highest AUC).

4. Final Model Training and Evaluation

Script: R/Model_Saving_and_Test_Performance_Metrics_with_Exclusion.R

This is the main analysis script that:

• Trains optimal random forest model on full training set
• Saves trained models (the .rds files in models/)
• Evaluates performance on held-out test set
• Generates predictions for excluded kidneys
• Creates visualizations:
  • Feature importance plots
  • CKD stage distributions
  • ROC curves (for DGF)
  • Predicted vs. true eGFR scatterplots

Usage:

# Set outcome type at line 28
outcome <- "C"  # "C" for eGFR, "B" for DGF

# Source required functions
source("R/Helper_CV_Functions.R")
source("R/CKD_Helper_Functions.R")

# Run the script
set.seed(382025)
# Script will automatically load data and generate all outputs

Reproducibility Notes

Random Seeds

All random processes use seed 382025 for reproducibility:

• Train/test splitting
• MRMR feature selection
• Cross-validation folds
• Random forest training
• Feature importance calculation (permutation importance)

Results Variability

Results may vary slightly from the published analysis due to:

• Differences in training/testing cohorts
• Bootstrap resampling procedures
• Optimization of machine learning models over hyperparameters
• Cross-validation procedures
• Randomness in MRMR feature selection
• Permutation-based feature importance calculations

This variability is expected and inherent to machine learning methods involving stochastic processes.

Evaluation Metrics

• eGFR (continuous): Mean Squared Error (MSE)
• DGF (binary): AUC, sensitivity, specificity, Youden's index

CKD Staging

The analysis includes visualization of Chronic Kidney Disease (CKD) staging:

• Stage 1: eGFR ≥ 90
• Stage 2: eGFR 60-89
• Stage 3a: eGFR 45-59
• Stage 3b: eGFR 30-44
• Stage 4: eGFR 15-29
• Stage 5: eGFR < 15

File Descriptions

File	Purpose
Create_Train_Test_Exclusion_Data.R	Data preprocessing, feature selection, train/test split
Binary_Internal_Validation_Metrics.R	Internal validation for DGF prediction
Continuous_Internal_Validation_Metrics.R	Internal validation for eGFR prediction
Model_Saving_and_Test_Performance_Metrics_with_Exclusion.R	Final model training and evaluation
Visualize_MSE_or_C_training.R	Visualization of internal validation performance
Helper_CV_Functions.R	Cross-validation utility functions
CKD_Helper_Functions.R	CKD classification and plotting functions

Citation

If you use this code or models in your research, please cite:

Rodrigues, L., Paul, A.S., Rubin, J., Magdy, H., Gupta, A., Pardinhas, C., Pimenta, C., Fernandes, B., Simões, I., Sousa, V., Figueirerdo, A., Alves, R., Zee, J., Sarder, P. (in preparation). Multimodal ComPRePS: Integrating High-dimensional Procurement Biopsy Pathomics and Clinical Data for Prediction of Allograft Outcomes.

@article{rodrigues2026compreps,
  title={Multimodal ComPRePS: Integrating High-dimensional Procurement Biopsy Pathomics and Clinical Data for Prediction of Allograft Outcomes},
  author={Rodrigues, Luis and Paul, Anindya S. and Rubin, Jeremy and Magdy, Haitham and Gupta, Akshita and Pardinhas, Clara and Pimenta, Carolina and Fernandes, Beatriz and Simões, Ilda and Sousa, Vitor and Figueirerdo, Arnaldo and Alves, Rui and Zee, Jarcy and Sarder, Pinaki},
  journal={In preparation},
  year={2026},
  note={Code and models available at: https://github.com/jeremysrubin/procurement-biopsy-pathomics-ml}
}

Author

Jeremy Rubin

Contact

For questions or issues, please open an issue on GitHub or contact jrub@umd.edu.