Code repository for the MLHC 2022 paper "Learning Optimal Summaries of Clinical Time-series with Concept Bottleneck Models"

Installation and Requirements

• conda env create -f environment.yml
• conda activate myenv

Data Source

The MIMIC-III data source used in these files was created by running this notebook.

Taxonomy

Before running any other files in this directory, you first must run the vasopressor_preprocess_data.ipynb notebook. This notebook stores standardized data files to a data/ subdirectory.

Models and helpers

• models.py contains all of the model parameters and training code for the custom model classes (logreg and LSTM).
• custom_losses.py contains all of the custom loss functions used during model training.
• preprocess_helpers.py contains helper functions to load and pre-process patient data.
• param_initializations.py contains helper functions used to initialize model parameters.
• weights_parser.py contains the WeightsParser class, which is used to easily match neural network weights to their corresponding feature names and summary functions.

Scripts

• gpu_*

  • gpu_greedy_top_concepts.py -- script for greedy optimization method for feature selection within concepts
  • gpu_greedy_top_concepts_baseline.py -- script for greedy optimization for feature selection in Johnson et al. model
  • gpu_lstm_baseline_compound.py -- script for training LSTM model using hyperparameters parsed from command-line arguments
  • gpu_vasopressor_baseline.py -- script for training Johnson et al. logistic regression model using hyperparameters parsed from command-line arguments
  • gpu_vasopressor_bottleneck.py -- script for training concept bottleneck models using hyperparameters parsed from command-line arguments
  • gpu_vasopressor_bottleneck_completeness.py -- script for training concept bottleneck models with concept scores in order to calculate completeness score (per Yeh et al.)
  • gpu_vasopressor_bottleneck_finetune.py -- script for finetuning a trained concept bottleneck models but only using the top k features
  • gpu_weights_top_concepts_baseline.py -- script for feature selection optimization based on coefficient magnitude for Johnson et al. model

• launch_*

  • launch_greedy_baseline_jobs.py -- launches GPU jobs for gpu_greedy_top_concepts.py
  • launch_greedy_bottleneck_jobs.py -- launches GPU jobs for gpu_greedy_top_concepts_baseline.py
  • launch_lstm_jobs.py -- launches GPU jobs for gpu_lstm_baseline_compound.py
  • launch_vasopressor_baseline_jobs.py -- launches GPU jobs for gpu_vasopressor_baseline.py
  • launch_vasopressor_bottleneck_jobs.py -- launches GPU jobs for gpu_vasopressor_bottleneck.py
  • launch_weights_baseline_jobs.py -- launches GPU jobs for gpu_weights_top_concepts_baseline.py

Jupyter Notebooks

• model-verification.ipynb contains the code which aggregates data from experiments and plots figures/graphs seen in paper