==============================
This repository is the official code of "Out-of-distribution Reject Option Method for Dataset Shift Problem in Early Disease Onset Prediction" (Tosaki et al. Scientific Reports, 2025).
- Our proposed method "Out-of-distribution Reject Option (ODROP)" enables reliable and accurate predictions across health and medical data affected by dataset shift without updating the prediction model.
- This repository contains toy examples using ODROP method.
- The detailed usage is explained in the notebooks.
- The OOD detection methods including this repository are followings:
- VAE Reconstruction Loss
- NN Classifier Ensemble Std
- NN Classifier Ensemble Epistemic
- NN Classifier Energy Score
- NN Classifier Gaussian mixture-based Energy Measurement (GEM)
├── data
│ ├── external <- Toy data is here.
│
├── docs
│
├── models
│
├── notebooks <- Jupyter notebooks for example.
│
├── references
│
├── reports
│ └── figures <- Generated figures
│
├── requirements.txt <- The requirements file for reproducing the analysis environment.
│
├── setup.py <- makes project pip installable (pip install -e .) so src can be imported
├── src <- Source code for use in this project.
│ ├── __init__.py
│ │
│ ├── data
│ │ └── make_dataset.py <- Scripts for generating toy dataset.
│ │
│ │
│ ├── models <- Scripts to train models and then use trained models.
│ │ │
│ │ ├── weight <- model weight
│ │ ├── adversarial_train.py <- adversarial training (ensemble) models
│ │ ├── NN.py <- Simple Binary Classification NN model
│ │ └── train_model.py <- training (ensemble) Simple NN models
│ │
│ └── OODmetrics
│ │ ├── ensemble.py <- epistemic, aleatoric uncertainty
│ │ └── nn_posthoc.py <- post hoc method with Simple NN Classifier for OOD score. ex: Energy-based, GEM score
│ └── utils
│ │ └── util.py <- for Rejection curve plot
│ │
│ └── visualization <- Scripts to create exploratory and results oriented visualizations
│ └── visualize.py
Project based on the cookiecutter data science project template. #cookiecutterdatascience