Chemotherapy Side Effect Prediction Pipeline

This repository provides a reproducible pipeline to:

• Extract cohorts from MIMIC-IV and a Code dataset from a MII node (Erlangen University Hospital, UKEr) for patients with cancer diagnoses undergoing chemotherapy.
• Train classical machine learning and temporal deep learning models to predict chemotherapy-induced side effects using longitudinal lab data and demographic features.

Models

Classical models

• Tree-based: Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGB), CatBoost (CatB)
• Other: Logistic Regression (LR), Multilayer Perceptron (MLP)

Temporal models

• Regular time series: GRU, LSTM, Temporal Convolutional Network (TCN), SAnD
• Irregular time series: GRU-D, InterpNet, STraTS

Cohorts

• Patients with a cancer diagnosis and at least one chemotherapy procedure.
• Aplasia: defined by transfusion procedures or low absolute neutrophil count.
• Neutropenic fever: concurrent neutropenia and fever diagnoses.
• Prediction target: onset of the condition within the prediction window (45 days after discharge for aplasia, 30 days for neutropenic fever) and before the next chemotherapy cycle.
• Observation window: 14 days prior to discharge.

Structure

• config/ – Configuration files

  • constants.py – Shared constants and paths
  • model_params.yaml – Grid search parameters for classical models
  • ml_config_params_best.yaml – Best parameters/cohort for classical models
  • ts_config_params.yaml – Grid search parameters for neural network models
  • ts_config_params_best.yaml – Best parameters/cohort for neural network models

• ml_model_training/ – Training scripts for classical models

• ts_model_training/ – Training scripts for deep learning models

Usage

1. Create the environment:

conda env create -f environment.yml
conda activate flabnet_ml_pipeline_env

2. Train models

Classical models

Run the training script with:

bash train_ml_models.sh [mimic|uker] [prepare|train|full_run]

• mimic – for MIMIC-IV cohorts
• uker – for Erlangen University Hospital cohorts

Workflow options:

• prepare – prepare inputs
• train – train models on prepared inputs
• full_run – prepare inputs and train models

Example:

bash train_ml_models.sh mimic full_run

Temporal deep models

To run temporal deep models on already prepared inputs:

bash train_ts_models.sh