Companion code for "Bayesian logistic regression for online recalibration and revision of risk prediction models with performance guarantees"

Installation

We use pip to install things into a python virtual environment. Refer to requirements.txt for package requirements. We use nestly + SCons to run simulations.

File descriptions

generate_data_single_pop.py -- Simulate a data stream from a single population following a logistic regression model.

• Inputs:
  • --simulation: string for selecting the type of distribution shift. Options for this argument are the keys in SIM_SETTINGS in constants.py.
• Outputs:
  • --out-file: pickle file containing the data stream

generate_data_two_pop.py -- Simulate a data stream from two subpopulations, where each are generated using logistic regression models. Similar arguments as generate_data_single_pop.py. The percentage split beween the two subpopulations is controlled by the --subpopulations argument.

• Outputs:
  • --out-file: pickle file containing the data stream

create_modeler.py -- Creates a model developer who fits the original prediction model and may propose a continually refitted model at each time point.

• Inputs:
  • --data-file: pickle file with the entire data stream
  • --simulation: string for selecting the model refitting strategy by the model developer. Options are to keep the model locked (locked), refit on all accumulated data (cumulative_refit), and refit on the latest observations within some window length (boxed, window length specified by --max-box). The last two options is to train an ensemble with the original and the cumulative_refit models (combo_refit) and train an ensemble with the original and the boxed models (combo_boxed).
• Outputs:
  • --out-file: pickle file containing the modeler

main.py -- Given the data and the model developer, run online model recalibration/revision using MarBLR and BLR.

• Inputs:
  • --data-file: pickle file with the entire data stream
  • --model-file: pickle file with the model developer
  • --type-i-regret-factor: Type I regret will be controlled at the rate of args.type_i_regret_factor * (Initial loss of the original model)
  • --reference-recalibs: comma-separated string to select which other online model revisers to run. Options are no updating at all locked, ADAM adam, cumulative logistic regression cumulativeLR.
• Outputs:
  • --obs-scores-file: csv file containing predicted probabilities and observed outcomes on the data stream
  • --history-file: csv file containing the predicted and actual probabilities on a held-out test data stream (only available if the data stream was simulated)
  • --scores-file: csv file containing performance measures on a held-out test data stream (only available if the data stream was simulated)
  • --recalibrators-file: pickle file containing the history of the online model revisers

Reproducing simulation results

The simulation_recalib folder contains the first set of simulations for online model recalibration. The simulation_revise folder contains the second set of simulations where we perform online logistic revision. The simulation_revise folder contains the third set of simulations where we perform online ensembling of the original model with a continually refitted model. The copd_analysis folder contains code for online model recalibration and revision for the COPD dataset. To reproduce the simulations, run scons .