rvandewater/YAIB has 112 GitHub stars and 36 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-15 15:57 UTC. Last push 2026-11-08.
Measured star growth
- in 24h
- 0
- in 7 days
- 0
- in 15 days
- Update delayed
- in 1 month
- Update delayed
Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.
About
🧪Yet Another ICU Benchmark: a holistic framework for the standardization of clinical prediction model experiments. Provide custom datasets, cohorts, prediction tasks, endpoints, preprocessing, and models. Paper: https://arxiv.org/abs/2306.05109
Machine LearningREADME.md
Original
🧪 Yet Another ICU Benchmark
Yet another ICU benchmark (YAIB) provides a framework for doing clinical machine learning experiments on Intensive Care Unit (ICU) and other Electronic Health Record (EHR) data.
We support the following datasets out of the box:
| Dataset | MIMIC-III / IV | eICU-CRD | HiRID | AUMCdb |
|---|---|---|---|---|
| Admissions | 40k / 73k | 200k | 33k | 23k |
| Version | v1.4 / v2.2 | v2.0 | v1.1.1 | v1.0.2 |
| Frequency (time-series) | 1 hour | 5 minutes | 2 / 5 minutes | up to 1 minute |
| Originally published | 2015 / 2020 | 2017 | 2020 | 2019 |
| Origin | USA | USA | Switzerland | Netherlands |
New datasets can also be added. We are currently working on a package to make this process as smooth as possible. The benchmark is designed for operating on preprocessed parquet files.
We provide five common tasks for clinical prediction by default:
| No | Task | Frequency | Type |
|---|---|---|---|
| 1 | ICU Mortality | Once per Stay (after 24H) | Binary Classification |
| 2 | Acute Kidney Injury (AKI) | Hourly (within 6H) | Binary Classification |
| 3 | Sepsis | Hourly (within 6H) | Binary Classification |
| 4 | Kidney Function (KF) | Once per stay | Regression |
| 5 | Length of Stay (LoS) | Hourly (within 7D) | Regression |
New tasks can be easily added. To get started right away, we include the eICU and MIMIC-III demo datasets in our repository.
The following repositories may be relevant as well:
- YAIB-cohorts: Cohort generation for YAIB.
- YAIB-models: Pretrained models for YAIB.
- ReciPys: Preprocessing package for YAIB pipelines.
- For all YAIB-related repositories, please see: https://github.com/stars/rvandewater/lists/yaib.
📄 Paper
To reproduce the benchmarks in our paper, see the ML reproducibility document. If you use this code in your research, please cite the following publication:
@inproceedings{vandewaterYetAnotherICUBenchmark2024,
title = {Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML},
shorttitle = {Yet Another ICU Benchmark},
booktitle = {The Twelfth International Conference on Learning Representations},
author = {van de Water, Robin and Schmidt, Hendrik Nils Aurel and Elbers, Paul and Thoral, Patrick and Arnrich, Bert and Rockenschaub, Patrick},
year = {2024},
month = oct,
urldate = {2024-02-19},
langid = {english},
}
This paper can also be found on arXiv: 2306.05109.
💿 Installation
YAIB is installed from source using uv and the repository's pyproject.toml.
Installation from source
First, clone this repository:
git clone https://github.com/rvandewater/YAIB.git
cd YAIB
Please note the branch. The newest features and fixes are available on the development branch:
git checkout development
Install uv if needed:
curl -LsSf https://astral.sh/uv/install.sh | sh
Pin Python, create the virtual environment, and install all project dependencies:
uv python install 3.12
uv sync
To install development dependencies as well:
uv sync --dev
The hyperimpute and PyPOTS imputation baselines and the SSSD imputation models need the optional imputation extra. It adds about 100 packages, including TensorFlow:
uv sync --extra imputation
Platform notes
- On macOS, YAIB should use the default PyPI
torchwheel. - On Linux and Windows, the project is configured to resolve the CUDA-enabled
torchwheel from the PyTorch index. - On Macs with Metal Performance Shaders, run YAIB with
PYTORCH_ENABLE_MPS_FALLBACK=1if needed.
After installation, check that the CLI is available:
uv run icu-benchmarks train -h
YAIB will automatically list available processors at initialization in its log files.
👩💻 Usage
Please refer to the documentation for detailed information on how to use and extend YAIB, including the API reference.
Quickstart 🚀 (demo data)
The authors of MIMIC-III and eICU have made a small demo dataset available to demonstrate their use. They can be found on PhysioNet: MIMIC-III Clinical Database Demo and eICU Collaborative Research Database Demo. These datasets are published under the Open Data Commons Open Database License v1.0 and can be used without credentialing procedure. We have created demo cohorts processed solely from these datasets for each of our currently supported task endpoints. To the best of our knowledge, this complies with the license and the respective dataset author's instructions. Usage of the task cohorts and the dataset is only permitted with the above license.
We strongly recommend completing a human subject research training to ensure you properly handle human subject research data.
In the folder demo_data we provide processed publicly available demo datasets from eICU and MIMIC with the necessary labels for Mortality at 24h, Sepsis, Acute Kidney Injury, Kidney Function, and Length of Stay.
If you do not yet have access to the ICU datasets, you can run the following command to train models for the included demo cohorts:
uv run icu-benchmarks train \
-d demo_data/mortality24/mimic_demo \
-n mimic_demo \
-t BinaryClassification \
-tn Mortality24 \
-m XGBClassifier \
--seed 2222 \
-l ../yaib_logs/ \
--tune
If you want to reproduce the LightGBM example with cache generation:
uv run icu-benchmarks train \
-d demo_data/mortality24/mimic_demo \
-n mimic_demo \
-t BinaryClassification \
-tn Mortality24 \
-m LGBMClassifier \
-hp LGBMClassifier.min_child_samples=10 \
--generate_cache \
--load_cache \
--seed 2222 \
-l ../yaib_logs/ \
--tune
For a list of available flags, run
uv run icu-benchmarks train -h.
On Macs with Metal Performance Shaders, run with
PYTORCH_ENABLE_MPS_FALLBACK=1if needed.
For Windows based systems, the next line character (
\) needs to be replaced by^(Command Prompt) or`(PowerShell).
Alternatively, the easiest method to train all the models in the paper is to run these commands from the repository root:
uv run wandb sweep --verbose experiments/benchmark_classification.yml
uv run wandb sweep --verbose experiments/benchmark_regression.yml
This will create two hyperparameter sweeps for Weights & Biases for the classification and regression tasks. You can then run the following command to train the models:
uv run wandb agent <sweep_id>
Tip: You can choose to run each of the configurations on a SLURM cluster instance with
uv run wandb agent --count 1 <sweep_id>.
Note: You will need to have a Weights & Biases account and be logged in to run the above commands.
Getting the datasets
HiRID, eICU, and MIMIC-IV can be accessed through PhysioNet. A guide to this process can be found here. AUMCdb can be accessed through a separate procedure. We do not have involvement in the access procedure and cannot answer requests for data access.
Cohort creation
Since the datasets were created independently of each other, they do not share the same data structure or data identifiers. In order to make them interoperable, use the preprocessing utilities provided by the ricu package. Ricu pre-defines a large number of clinical concepts and how to load them from a given dataset, providing a common interface to the data that is used in this benchmark. Please refer to our cohort definition code for generating the cohorts using our Python interface for ricu. After this, you can run the benchmark once you have gained access to the datasets.
👟 Running YAIB
Preprocessing and Training
The following command will run training and evaluation on the MIMIC demo dataset for binary mortality prediction at 24h with LGBMClassifier:
uv run icu-benchmarks train \
-d demo_data/mortality24/mimic_demo \
-n mimic_demo \
-t BinaryClassification \
-tn Mortality24 \
-m LGBMClassifier \
-hp LGBMClassifier.min_child_samples=10 \
--generate_cache \
--load_cache \
--seed 2222 \
-l ../yaib_logs/ \
--tune
For a list of available flags, run
uv run icu-benchmarks train -h.
Run with
PYTORCH_ENABLE_MPS_FALLBACK=1on Macs with Metal Performance Shaders.
For Windows based systems, the next line character (
\) needs to be replaced by^(Command Prompt) or`(PowerShell), respectively.
Alternatively, the easiest method to train all the models in the paper is to run these commands from the directory root:
uv run wandb sweep --verbose experiments/benchmark_classification.yml
uv run wandb sweep --verbose experiments/benchmark_regression.yml
This will create two hyperparameter sweeps for WandB for the classification and regression tasks. This configuration will train all the models in the paper. You can then run the following command to train the models:
uv run wandb agent <sweep_id>
Tip: You can choose to run each of the configurations on a SLURM cluster instance by
uv run wandb agent --count 1 <sweep_id>.
Note: You will need to have a WandB account and be logged in to run the above commands.
Regression target scaling
Use -t Regression for both length of stay and kidney function. Targets are min/max scaled using only the training partition of each fold; validation/test reuse that transform without clipping. Feature scaling is independent. Targets must use consistent units across datasets.
Regression metrics MAE_native, MSE_native, and RMSE_native (ML) are in original target units (squared units for MSE). R2 is unchanged. train/loss, val/loss, and test/loss remain in model space for optimization/early stopping; regression tuning returns native-unit MSE instead. Do not compare the new native metrics to older scaled metrics or average errors across different clinical units.
Override the default with Gin bindings, for example:
# No target scaling
-hp 'base_regression_preprocessor.target_scaling="none"'
# Explicit fixed bounds, for reproducing prescribed normalization
-hp 'base_regression_preprocessor.target_scaling="fixed"' \
'base_regression_preprocessor.outcome_min=0' \
'base_regression_preprocessor.outcome_max=168'
Every regression fold saves target_transform.json alongside its model. Keep these together: evaluation and fine-tuning reuse the source transform, and reject missing/mismatched state. Regression ML .joblib exports now contain a dictionary with model, target_transform, and trained_columns; DL checkpoints also embed the transform. Only load trusted model/cache artifacts. Legacy regression models must be retrained; old preprocessing caches are invalidated. A complete legacy bounds pair still selects fixed scaling with a deprecation warning.
For custom integrations, preprocess_data() now returns (data, target_transform); pass both to train_common(). Non-regression preprocessing returns None as its transform. Cache hits restore data and fitted state together. Input files are still assumed immutable within a cached run. Saved regression predictions and SHAP contributions are in native units; SHAP base values are saved separately as *_shap_base_values.npy.
Evaluate or Finetune
It is possible to evaluate a model trained on another dataset without additional training. In this case, the source dataset is the demo data from MIMIC and the target is the eICU demo:
uv run icu-benchmarks \
--eval \
-d demo_data/mortality24/eicu_demo \
-n eicu_demo \
-t BinaryClassification \
-tn Mortality24 \
-m LGBMClassifier \
--generate_cache \
--load_cache \
-s 2222 \
-l ../yaib_logs \
-sn mimic \
--source-dir ../yaib_logs/mimic_demo/Mortality24/LGBMClassifier/2022-12-12T15-24-46/repetition_0/fold_0
A similar syntax is used for finetuning, where a model is loaded and then retrained. To run finetuning, replace
--evalwith-ft.
Models
We provide several existing machine learning models that are commonly used for multivariate time-series data.
pytorch is used for the deep learning models, lightgbm for the boosted tree approaches, and sklearn for other classical machine learning models.
The benchmark provides (among others) the following built-in models:
- Logistic Regression: Standard regression approach.
- Elastic Net: Linear regression with combined L1 and L2 priors as regularizer.
- LightGBM: Efficient gradient boosting trees.
- Long Short-term Memory (LSTM): The most commonly used type of recurrent neural networks for long sequences.
- Gated Recurrent Unit (GRU): An extension to LSTM which showed improvements (paper).
- GRU-D: A GRU that decays its inputs and hidden state with the time since each variable was last measured, using the missingness indicators of the preprocessing. Ported from the YAIB fork of Katja Jagd and Rachael DeVries.
- Temporal Convolutional Networks (TCN): 1D convolution approach to sequence data. By using dilated convolution to extend the receptive field of the network it has shown great performance on long-term dependencies.
- Transformers: The most common attention-based approach.
- Bi-Axial Transformer (BAT): Attends across the variables within each hour and causally across the hours of each variable, with the missingness indicators as observation flags (pretraining on YAIB). Ported from the YAIB fork of Katja Jagd and Rachael DeVries.
- Mamba: A selective state-space model that scales linearly with sequence length. EHRMamba showed that it works well on EHR data; YAIB includes the Mamba backbone, trained from scratch on the hourly features rather than pretrained on tokenized events, so results are not comparable to EHRMamba's (see
mamba.pyfor the differences). Runs in plain PyTorch, or with the fused kernels ofmamba-ssmif installed on an NVIDIA GPU.
🛠️ Development
To adapt YAIB to your own use case, see Extending YAIB and the development guide in the documentation. We appreciate contributions to the project. Please read the contribution guidelines before submitting a pull request.
Acknowledgements
This project has been developed partially under the funding of “Gemeinsamer Bundesausschuss (G-BA) Innovationsausschuss” in the framework of “CASSANDRA - Clinical ASSist AND aleRt Algorithms” (project number 01VSF20015). We would like to acknowledge the work of Alisher Turubayev, Anna Shopova, Fabian Lange, Mahmut Kamalak, Paul Mattes, and Victoria Ayvasky for adding PyTorch Lightning, Weights and Biases compatibility, and several optional imputation methods to a later version of the benchmark repository.
We do not own any of the datasets used in this benchmark. This project uses heavily adapted components of the HiRID benchmark. We thank the authors for providing this codebase and encourage further development to benefit the scientific community. The demo datasets have been released under an Open Data Commons Open Database License (ODbL).
License
This source code is released under the MIT license, included here. We do not own any of the datasets used or included in this repository.