This is the Simulated interVention (SimVene) software for predicting 40% eGFR decline.

• SimVene reduces the complexity for implementing and managing simulated interventions to support clinical decision making.

• This library automates preprocessing, unknown structure learning (via RAUS), model evaluation, and sensitivity analysis to include sharp regression discontinuity and do-calculus.

• We introduce an "orchestration card" to help manage the RAUS implementation across multiple virtual machines, scenarios, sites, and substructures.

• We introduce a sharp regression discontinuity design implementation to randomize patients into control and treatment groups.

• We introduce a do-calculus implementation to delete incoming edges to desired interventions from parents.

• SimVene is built on top of Ranking Approaches for Unknown Structures (RAUS) by Gordon et al. and Bayes Net Toolbox (BNT) by Murphy et al.

Cite

• Please cite the SimVene software if you use the SimVene software in your work.

INSERT ZENODO DOI HERE

The SimVene software is first implemented in the papers:

1. "Predicting ≥40% eGFR Decline Using Dynamic Bayesian Networks in Persons With or At-Risk of Chronic Kidney Disease" by Panayiotis Petousis*, David Gordon*, Susanne B. Nicholas, and Alex A.T. Bui on behalf of CURE-CKD

2. "Testing Causal Explanations: A Case Study for Understanding Chronic Kidney Disease" by Panayiotis Petousis*, David Gordon, Susanne B. Nicholas, and Alex A.T. Bui on behalf of CURE-CKD

3. "Using Bayesian Networks to Simulate eGFR Trajectories" published at the 34th meeting of the European Renal Cell Study Group (ERCSG) Abstract.

*First authors had equal contributions.

How to use SimVene

To install the requirements, run: (Note use Python 3.7.4)

$ make requirements

Example Commands

Preprocessing

To preprocess the data, run:

$ make preprocess_data

To split the data (Train: 60%, Validation: 20%, Test: 20%), run:

$ make split_data

To discretize the data for all sites (Combined, UCLA, Providence), run:

$ make discretize_data
$ make discretize_data_using_UCLA
$ make discretize_data_using_Prov

Unknown Structure Learning via RAUS

To format data for input to RAUS, run:

$ make raus_format_data

Setup RAUS environment requiremnts by following environment setup instrunctions

Note: Each scenario represents a different set of variables used for structure learning. For the above Papers scenario 2 was used, as it represented the most clinically interpretable model.

To perform unknown structure learning (via RAUS) for Scenario 1 (36 CPUs in parallel), run:

$ make -j 4 scenario_1_1
$ make -j 4 scenario_1_2
$ make -j 4 scenario_1_3

or, run:

$ make learn_UCLA_contemporals & make learn_UCLA_initial_conditions & make learn_UCLA_intra_for_full_DBN & make learn_UCLA_inter_for_full_DBN
$ make learn_PSJH_contemporals & make learn_PSJH_initial_conditions & make learn_PSJH_intra_for_full_DBN & make learn_PSJH_inter_for_full_DBN
$ make learn_Combined_contemporals & make learn_Combined_initial_conditions & make learn_Combined_intra_for_full_DBN & make learn_Combined_inter_for_full_DBN

To perform unknown structure learning (via RAUS) for Scenario 2 (36 CPUs in parallel), run:

$ make -j 4 scenario_2_1
$ make -j 4 scenario_2_2
$ make -j 4 scenario_2_3

or, run:

$ make learn_UCLA_contemporals_norace & make learn_UCLA_initial_conditions_norace & make learn_UCLA_intra_for_full_DBN_norace & make learn_UCLA_inter_for_full_DBN_norace
$ make learn_PSJH_contemporals_norace & make learn_PSJH_initial_conditions_norace & make learn_PSJH_intra_for_full_DBN_norace & make learn_PSJH_inter_for_full_DBN_norace
$ make learn_Combined_contemporals_norace & make learn_Combined_initial_conditions_norace & make learn_Combined_intra_for_full_DBN_norace & make learn_Combined_inter_for_full_DBN_norace

To perform unknown structure learning (via RAUS) for Scenario 3 (36 CPUs in parallel), run:

$ make -j 4 scenario_3_1
$ make -j 4 scenario_3_2
$ make -j 4 scenario_3_3

or, run:

$ make learn_UCLA_counts_contemporals & make learn_UCLA_counts_initial_conditions & make learn_UCLA_counts_intra_for_full_DBN & make learn_UCLA_counts_inter_for_full_DBN
$ make learn_PSJH_counts_contemporals & make learn_PSJH_counts_initial_conditions & make learn_PSJH_counts_intra_for_full_DBN & make learn_PSJH_counts_inter_for_full_DBN
$ make learn_Combined_counts_contemporals & make learn_Combined_counts_initial_conditions & make learn_Combined_counts_intra_for_full_DBN & make learn_Combined_counts_inter_for_full_DBN

To perform unknown structure learning (via RAUS) for Scenario 4 (36 CPUs in parallel), run:

$ make -j 4 scenario_4_1
$ make -j 4 scenario_4_2
$ make -j 4 scenario_4_3

or, run:

$ make learn_UCLA_counts_contemporals_norace & make learn_UCLA_counts_initial_conditions_norace & make learn_UCLA_counts_intra_for_full_DBN_norace & make learn_UCLA_counts_inter_for_full_DBN_norace
$ make learn_PSJH_counts_contemporals_norace & make learn_PSJH_counts_initial_conditions_norace & make learn_PSJH_counts_intra_for_full_DBN_norace & make learn_PSJH_counts_inter_for_full_DBN_norace
$ make learn_Combined_counts_contemporals_norace & make learn_Combined_counts_initial_conditions_norace & make learn_Combined_counts_intra_for_full_DBN_norace & make learn_Combined_counts_inter_for_full_DBN_norace

Just for Paper 3 - use the following script (scenario 4 only UCLA) - for other papers continue to next steps

python DBN_model_learning/ERSA_model/full_model_BN_vars_dataset.py

Model Parameter Learning and Evaluation

Note: to use SMILE the API for Bayesian Networks by Bayesfusion you need to obtain a free academic license and replace the existing license file DBN_model_learning/pysmile_license.py. (https://www.bayesfusion.com/downloads/)

To format data for GeNIe, run:

$ make genie_format_data

To undersample the training set, run:

$ make data_undersampling

To train the full Bayesian networks, run:

$ make full_BNs_models_eval

To train the full dynamic Bayesian networks, run:

$ make full_DBNs_models_eval

To get the results for all the full dynamic Bayesian networks, run:

$ make full_DBNs_models_results

To get the results for all the full DBNs with bootstrap (stratified to maintain outcome proportions in sampling) and Confidence Intervals (CIs):

$ make full_DBNs_models_results_with_CIs_stratified

To get the results for all the full DBNs with bootstrap (balanced to maintain 50:50 outcome proportions in sampling) and Confidence Intervals (CIs):

$ make full_DBNs_models_results_with_CIs_balanced

To evaluate the statistical significance between the full dynamic Bayesian networks, run (not necessary for papers results - can be skipped):

$ make full_DBNs_models_stat_sign

To get the results for the best full dynamic Bayesian networks, run (not necessary for papers results - can be skipped):

$ make get_best_DBNs_models_results

To visualize the performance of all models, run:

$ make models_perf_visualization

To visualize the performance of all models on stratified bootstrap and CIs, run:

$ make models_perf_visualization_stratified

To visualize the performance of all models on balanced bootstrap and CIs, run:

$ make models_perf_visualization_balanced

To visualize the performance of all models on external datasets, run:

$ make models_perf_ext_datasets_visualization

To visualize the performance of all models on stratified bootstrap and CIs on external datasets, run:

$ make models_perf_ext_datasets_visualization_stratified

To visualize the performance of all models on balanced bootstrap and CIs on external datasets, run:

$ make models_perf_ext_datasets_visualization_balanced

Paper 2

Associational Analysis and Causal analysis

To perform sensitivity/causal analysis using Do queries on every variables model site and testing site, run:

$ make sens_analysis_of_outcome

or by site, run:

$ make sens_analysis_of_outcome_UCLA_UCLA
$ make sens_analysis_of_outcome_UCLA_PSJH
$ make sens_analysis_of_outcome_PSJH_PSJH
$ make sens_analysis_of_outcome_PSJH_UCLA
$ make sens_analysis_of_outcome_Combined_Combined

To perform sensitivity analysis alone (associational) without using Do queries on every variables model site and testing site, run:

$ make sens_analysis_of_outcome_pop_profiling_all_vars

or by site, run:

$ make sens_analysis_of_outcome_pop_profiling_all_vars_UCLA_UCLA
$ make sens_analysis_of_outcome_pop_profiling_all_vars_UCLA_PSJH
$ make sens_analysis_of_outcome_pop_profiling_all_vars_PSJH_PSJH
$ make sens_analysis_of_outcome_pop_profiling_all_vars_PSJH_UCLA
$ make sens_analysis_of_outcome_pop_profiling_all_vars_Combined_Combined

To visualize the sensitivity analysis via Jupyter Notebook, run:

Note: change following variables to create combinations of sites. Example: model_site = "UCLA" testing_site = "PSJH"

$ jupyter notebook ./notebooks/Sensitivity analysis.ipynb
$ jupyter notebook ./notebooks/Sensitivity analysis association.ipynb

One notebook for causal analysis and one notebook for the associational analysis