This is the repository accompanying the Modular Clinical Decision Support Networks (MoDN) Updatable, Interpretable, and Portable Predictions for Evolving Clinical Environments paper.
Modular Clinical Decision Support Networks (MoDN) is a novel decision tree composed of feature-specific neural network modules. It creates dynamic personalised representations of patients, and can make multiple predictions of diagnoses and features, updatable at each step of a consultation. The model is validated on a real-world Clinical Decision Support System (CDSS) derived dataset, comprising 3,192 pediatric outpatients in Tanzania.
MoDN significantly outperforms 'monolithic' baseline models (which take all features at once at the end of a consultation) with a mean macro
To test collaborative learning between imperfectly interoperable (IIO) datasets, we create subsets with various percentages of feature overlap and port a MoDN model trained on one subset to another. Even with only 60% common features, fine-tuning a MoDN model on the new dataset or just making a composite model with MoDN modules matched the ideal scenario of sharing data in a perfectly interoperable setting.
MoDN integrates into consultation logic by providing interpretable continuous feedback on the predictive potential of each question in a CDSS questionnaire. The modular design allows it to compartmentalise training updates to specific features and collaboratively learn between IIO datasets without sharing any data.
The experiments were run on python version 3.8.10. The data folder must contain the data (link to anonymized data https://zenodo.org/record/400380#.Yug5kuzP00Q). The models folder contains the scripts used to run the different experiments and produce the plots.
The script main.py calls the preprocessing pipeline on the data and trains the model either performing 5 times 2-fold CV (saving the different metrics) or just training a single model.
The script iio_training.py performs the IIO experiments (i.e. compartmentalization and fine tuning). The different models and performance scores are saved to the updated_centralized folder.
After having run both these files, one can run statistical_tests.py to produce the plots (uses saved metrics by the two previous scripts).
Run unsupervised_learning.py to produce the plot with the clustering of the state. Metrics and performance scores are saved to the saved_objects folder and plots to the saved_plots folder.
baselines.py contains the functions to compute the KNN and logistic regression baselines.
dataset_generation.py puts the data in shape to be used by the models.
distributed_training_parameters.py contains the parameters used by the distributed_training.py file.
graph_functions.py contains many functions to produce some plots.
modules.py contains the module and state definitions.
training_procedures_epoct.py contains the training and testing processes for the model.
utils_distributed.py contains some utlitary functions for the compartmentalisation and fine-tuning experiments.
utils_epoct.py contins utilitary functions specific to the epoct data and utils.py general utilitary functions.
To reproduce the results reported in Modular Clinical Decision Support Networks (MoDN) Updatable, Interpretable, and Portable Predictions for Evolving Clinical Environments, install the necessary dependencies using:
sudo apt install texlive texlive-latex-extra texlive-fonts-recommended dvipng cm-super
pip install -r requirements.txt (from the root directory)
Create a data folder and download the anonymized data from https://zenodo.org/record/400380#.Yug5kuzP00Q. Then run the different scripts as described in the Plots paragraph.