Skip to main content

A Python library for healthcare AI

Project description

Deployment & Documentation & Stats

PyPI version Documentation status MyBinder GitHub stars GitHub forks Downloads Downloads

Build Status & Coverage & Maintainability & License

Build Status Circle CI Build status Maintainability License
PyHealth Logo

Development Status: As of 01/04/2021, PyHealth is under active development and in its alpha stage. Please follow, star, and fork to get the latest functions!

PyHealth is a comprehensive Python package for healthcare AI, designed for both ML researchers and healthcare and medical practitioners. PyHealth accepts diverse healthcare data such as longitudinal electronic health records (EHRs), continuous signials (ECG, EEG), and clinical notes (to be added), and supports various predictive modeling methods using deep learning and other advanced machine learning algorithms published in the literature.

The library is proudly developed and maintained by researchers from Carnegie Mellon University, IQVIA, and University of Illinois at Urbana-Champaign. PyHealth makes many important healthcare tasks become accessible, such as phenotyping prediction, mortality prediction, and ICU length stay forecasting, etc. Running these prediction tasks with deep learning models can be as short as 10 lines of code in PyHealth.

PyHealth comes with three major modules: (i) data preprocessing module; (ii) learning module and (iii) evaluation module. Typically, one can run the data prep module to prepare the data, then feed to the learning module for model training and prediction, and finally assess the results with the evaluation module. Users can use the full system as mentioned or just selected modules based on their own needs:

  • Deep learning researchers may directly use the processed data along with the proposed new models.
  • Healthcare and Medical personnel, may leverage our data preprocessing module to convert the medical data to the format that machine learning models could digest, and then perform the inference tasks to get insights from the data. This package can support them in various health analytics tasks including disease detection, risk prediction, patient subtyping, health monitoring, etc.

PyHealth is featured for:

  • Unified APIs, detailed documentation, and interactive examples across various types of datasets and algorithms.
  • Advanced models, including latest deep learning models and classical machine learning models.
  • Wide coverage, supporting sequence data, image data, series data and text data like clinical notes.
  • Optimized performance with JIT and parallelization when possible, using numba and joblib.
  • Customizable modules and flexible design: each module may be turned on/off or totally replaced by custom functions. The trained models can be easily exported and reloaded for fast execution and deployment.

API Demo for LSTM on Phenotyping Prediction:

# load pre-processed CMS dataset
from import sequencedata as expdata_generator

expdata_id = ''
cur_dataset = expdata_generator(exp_id=exp_id)
cur_dataset.get_exp_data(sel_task='mortality', )

# initialize the model for training
from pyhealth.models.sequence.lstm import LSTM
# enable GPU
expmodel_id = 'test.model.lstm.0001'
clf = LSTM(expmodel_id=expmodel_id, n_batchsize=20, use_gpu=True, n_epoch=100), cur_dataset.valid)

# load the best model for inference
pred_results = clf.get_results()

# evaluate the model
from pyhealth.evaluation.evaluator import func
r = func(pred_results['hat_y'], pred_results['y'])

Citing PyHealth:

PyHealth paper is under review at JMLR (machine learning open-source software track). If you use PyHealth in a scientific publication, we would appreciate citations to the following paper:

  author  = {Zhao, Yue and Qiao, Zhi and Xiao, Cao and Glass, Lucas and Sun, Jimeng},
  title   = {PyHealth: A Python Library for Health Predictive Models},
  year    = {2021},


Zhao, Y., Qiao, Z., Xiao, C., Glass, L. and Sun, J., 2021. PyHealth: A Python Library for Health Predictive Models.

Key Links and Resources:

Table of Contents:


It is recommended to use pip for installation. Please make sure the latest version is installed, as PyHealth is updated frequently:

pip install pyhealth            # normal install
pip install --upgrade pyhealth  # or update if needed
pip install --pre pyhealth      # or include pre-release version for new features

Alternatively, you could clone and run file:

git clone
cd pyhealth
pip install .

Required Dependencies:

  • Python 3.5, 3.6, or 3.7
  • combo>=0.0.8
  • joblib
  • numpy>=1.13
  • numba>=0.35
  • pandas>=0.25
  • scipy>=0.20
  • scikit_learn>=0.20
  • tqdm
  • torch (this should be installed manually)
  • xgboost (this should be installed manually)
  • xlrd >= 1.0.0
  • zipfile36
  • PyWavelets
  • torch
  • torchvision
  • xgboost

Warning 1: PyHealth has multiple neural network based models, e.g., LSTM, which are implemented in PyTorch. However, PyHealth does NOT install these DL libraries for you. This reduces the risk of interfering with your local copies. If you want to use neural-net based models, please make sure PyTorch is installed. Similarly, models depending on xgboost, would NOT enforce xgboost installation by default.

API Cheatsheet & Reference

Full API Reference: ( API cheatsheet for most learning models:

  • fit(X_train, X_valida): Fit a learning model.
  • inference(X): Predict on X using the fitted estimator.
  • evaluator(y, y^hat): Model evaluation.

Model load and reload:

  • load_model(): Load the best model so far.

Preprocessed Datasets & Implemented Algorithms

(i) Preprocessed Datasets (customized data preprocessing function is provided in the example folders):

Type Abbr Description Processed Function Link
Sequence: EHR-ICU MIMIC III A relational database containing tables of data relating to patients who stayed within ICU. \examples\data_generation\dataloader_mimic
Sequence: EHR-ICU MIMIC_demo The MIMIC-III demo database is limited to 100 patients and excludes the noteevents table. \examples\data_generation\dataloader_mimic_demo
Sequence: EHU-Claim CMS DE-SynPUF: CMS 2008-2010 Data Entrepreneurs Synthetic Public Use File \examples\data_generation\dataloader_cms
Image: Chest X-ray Pediatric Pediatric Chest X-ray Pneumonia (Bacterial vs Viral vs Normal) Dataset N/A
Series: ECG PhysioNet AF Classification from a short single lead ECG recording Dataset. N/A

You may download the above datasets at the links. The structure of the generated datasets can be found in datasets folder:

  • \datasets\cms\x_data...csv
  • \datasets\cms\y_data\phenotyping.csv
  • \datasets\cms\y_data\mortality.csv

The processed datasets (X,y) should be put in x_data, y_data correspondingly, to be appropriately digested by deep learning models. We include some sample datasets under \datasets folder.

(ii) Machine Learning and Deep Learning Models :

For sequence data:

Type Abbr Class Algorithm Year Ref
Classical Models RandomForest pyhealth.models.sequence.rf Random Forests 2000 [2]
Classical Models XGBoost pyhealth.models.sequence.xgboost XGBoost: A scalable tree boosting system 2016 [3]
Neural Networks LSTM pyhealth.models.sequence.lstm Long short-term memory 1997 [7]
Neural Networks GRU pyhealth.models.sequence.gru Gated recurrent unit 2014 [4]
Neural Networks RETAIN pyhealth.models.sequence.retain RETAIN: An Interpretable Predictive Model for Healthcare using Reverse Time Attention Mechanism 2016 [5]
Neural Networks Dipole pyhealth.models.sequence.dipole Dipole: Diagnosis Prediction in Healthcare via Attention-based Bidirectional Recurrent Neural Networks 2017 [9]
Neural Networks tLSTM pyhealth.models.sequence.tlstm Patient Subtyping via Time-Aware LSTM Networks 2017 [1]
Neural Networks RAIM pyhealth.models.sequence.raim RAIM: Recurrent Attentive and Intensive Model of Multimodal Patient Monitoring Data 2018 [10]
Neural Networks StageNet pyhealth.models.sequence.stagenet StageNet: Stage-Aware Neural Networks for Health Risk Prediction 2020 [6]

For image data:

Type Abbr Class Algorithm Year Ref
Neural Networks CNN pyhealth.models.sequence.basiccnn Face recognition: A convolutional neural-network approach 1997 [8]
Neural Networks Vggnet pyhealth.models.sequence.typicalcnn Very deep convolutional networks for large-scale image recognition 2014  
Neural Networks Inception pyhealth.models.sequence.typicalcnn Rethinking the Inception Architecture for Computer Vision    
Neural Networks Resnet pyhealth.models.sequence.typicalcnn Deep Residual Learning for Image Recognition    
Neural Networks Resnext pyhealth.models.sequence.typicalcnn Aggregated Residual Transformations for Deep Neural Networks    
Neural Networks Densenet pyhealth.models.sequence.typicalcnn Densely Connected Convolutional Networks    
Neural Networks Mobilenet pyhealth.models.sequence.typicalcnn MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications    

For ecg/egg data:

Type Abbr Class Algorithm Year Ref
Classical Models RandomForest pyhealth.models.ecg.rf Random Forests 2000 [2]
Classical Models XGBoost pyhealth.models.ecg.xgboost XGBoost: A scalable tree boosting system 2016 [3]
Neural Networks BasicCNN1D pyhealth.models.ecg.conv1d Face recognition: A convolutional neural-network approach 1997 [8]
Neural Networks DBLSTM-WS pyhealth.models.ecg.dblstm_ws A novel wavelet sequence based on deep bidirectional LSTM network model for ECG signal classification 2018  
Neural Networks DeepRes1D pyhealth.models.ecg.deepres1d Heartbeat classification using deep residual convolutional neural network from 2-lead electrocardiogram 2019  
Neural Networks AE+BiLSTM pyhealth.models.ecg.sdaelstm Automatic Classification of CAD ECG Signals With SDAE and Bidirectional Long Short-Term Network 2019  
Neural Networks KRCRnet pyhealth.models.ecg.rcrnet K-margin-based Residual-Convolution-Recurrent Neural Network for Atrial Fibrillation Detection 2019  
Neural Networks MINA pyhealth.models.ecg.mina MINA: Multilevel Knowledge-Guided Attention for Modeling Electrocardiography Signals 2019  

Examples of running ML and DL models can be found below, or directly at \examples\learning_examples\

(iii) Evaluation Metrics :

Type Abbr Metric Method
Binary Classification average_precision_score Compute micro/macro average precision (AP) from prediction scores
Binary Classification roc_auc_score Compute micro/macro ROC AUC score from prediction scores
Binary Classification recall, precision, f1 Get recall, precision, and f1 values
Multi Classification To be done here    

(iv) Supported Tasks:

Type Abbr Description Method
Multi-classification phenotyping Predict the diagnosis code of a patient based on other information, e.g., procedures \examples\data_generation\
Binary Classification mortality prediction Predict whether a patient may pass away during the hospital \examples\data_generation\
Regression ICU stay length pred Forecast the length of an ICU stay \examples\data_generation\

Quick Start for Data Processing

We propose the idea of standard template, a formalized schema for healthcare datasets. Ideally, as long as the data is scanned as the template we defined, the downstream task processing and the use of ML models will be easy and standard. In short, it has the following structure: add a figure here. The dataloader for different datasets can be found in examples/data_generation. Using “examples/data_generation/” as an exmaple:

  1. First read in patient, admission, and event tables.

    from pyhealth.utils.utility import read_csv_to_df
    patient_df = read_csv_to_df(os.path.join('data', 'mimic-iii-clinical-database-demo-1.4', 'PATIENTS.csv'))
    admission_df = read_csv_to_df(os.path.join('data', 'mimic-iii-clinical-database-demo-1.4', 'ADMISSIONS.csv'))
  2. Then invoke the parallel program to parse the tables in n_jobs cores.

    from import parallel_parse_tables
    all_results = Parallel(n_jobs=n_jobs, max_nbytes=None, verbose=True)(
     for i in range(n_jobs))
  3. The processed sequential data will be saved in the prespecified directory.

    with open(patient_data_loc, 'w') as outfile:
        json.dump(patient_data_list, outfile)

The provided examples in PyHealth mainly focus on scanning the data tables in the schema we have, and generate episode datasets. For instance, “examples/data_generation/” demonstrates the basic procedure of processing MIMIC III demo datasets.

  1. The next step is to generate episode/sequence data for mortality prediction. See “examples/data_generation/”

    with open(patient_data_loc, 'w') as outfile:
        json.dump(patient_data_list, outfile)

By this step, the dataset has been processed for generating X, y for phenotyping prediction. It is noted that the API across most datasets are similar. One may easily replicate this procedure by calling the data generation scripts in \examples\data_generation. You may also modify the parameters in the scripts to generate the customized datasets.

Preprocessed datasets are also available at \datasets\cms and \datasets\mimic.

Quick Start for Running Predictive Models

Note: Before running examples, you need the datasets. Please download from the GitHub repository “datasets”. You can either unzip them manually or running our script “”

Note: “examples/learning_models/” demonstrates the basic API of using GRU for mortality prediction. It is noted that the API across all other algorithms are consistent/similar.

Note: If you do not have the preprocessed datasets yet, download the \datasets folder ( and from PyHealth repository, and run \examples\learning_models\ to prepare/unzip the datasets.

Note: For “certain examples”, pretrained bert models are needed. You will need to download these pretrained models at:

Please download, unzip, and save to ./auxiliary folder.

  1. Setup the datasets. X and y should be in x_data and y_data, respectively.

    # load pre-processed CMS dataset
    from import sequencedata as expdata_generator
    expdata_id = ''
    cur_dataset = expdata_generator(exp_id=exp_id)
    cur_dataset.get_exp_data(sel_task='mortality', )
  2. Initialize a LSTM model, you may set up the parameters of the LSTM, e.g., n_epoch, learning_rate, etc,.

    # initialize the model for training
    from pyhealth.models.sequence.lstm import LSTM
    # enable GPU
    expmodel_id = 'test.model.lstm.0001'
    clf = LSTM(expmodel_id=expmodel_id, n_batchsize=20, use_gpu=True, n_epoch=100)
  3. Model loading, Load the saved model, default for ‘best’, maybe can personally set via ‘0’, ‘latest’, etc.

  4. Model training, parameters are learnt on the train datasets and verified on valid datasets, cur_dataset.valid)
  5. Model inferring, make prediction on the test datasets

    pred_results = clf.get_results()
  6. Evaluation on the model. Multiple metrics are supported.

    # evaluate the model
    from pyhealth.evaluation.evaluator import func
    r = func(pred_results['hat_y'], pred_results['y'])

Algorithm Benchmark

The comparison among of implemented models will be made available later with a benchmark paper. TBA soon :)

Blueprint & Development Plan

The long term goal of PyHealth is to become a comprehensive healthcare AI toolkit that supports all sorts of data types and predictive tasks.

  • The compatibility and the support of OMOP format datasets
  • Model persistence (save, load, and portability)
  • The release of a benchmark paper with PyHealth


[1]Baytas, I.M., Xiao, C., Zhang, X., Wang, F., Jain, A.K. and Zhou, J., 2017, August. Patient subtyping via time-aware lstm networks. In KDD.
[2](1, 2) Breiman, L., 2001. Random forests. Machine learning, 45(1), pp.5-32.
[3](1, 2) Chen, T. and Guestrin, C., 2016, August. Xgboost: A scalable tree boosting system. In KDD.
[4]Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H. and Bengio, Y., 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078.
[5]Choi, E., Bahadori, M.T., Sun, J., Kulas, J., Schuetz, A. and Stewart, W., 2016. Retain: An interpretable predictive model for healthcare using reverse time attention mechanism. In Advances in Neural Information Processing Systems (pp. 3504-3512).
[6]Gao, J., Xiao, C., Wang, Y., Tang, W., Glass, L.M. and Sun, J., 2020, April. StageNet: Stage-Aware Neural Networks for Health Risk Prediction. In Proceedings of The Web Conference 2020 (pp. 530-540).
[7]Hochreiter, S. and Schmidhuber, J., 1997. Long short-term memory. Neural computation, 9(8), pp.1735-1780.
[8](1, 2) Lawrence, S., Giles, C.L., Tsoi, A.C. and Back, A.D., 1997. Face recognition: A convolutional neural-network approach. IEEE transactions on neural networks, 8(1), pp.98-113.
[9]Ma, F., Chitta, R., Zhou, J., You, Q., Sun, T. and Gao, J., 2017, August. Dipole: Diagnosis prediction in healthcare via attention-based bidirectional recurrent neural networks. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining (pp. 1903-1911).
[10]Xu, Y., Biswal, S., Deshpande, S.R., Maher, K.O. and Sun, J., 2018, July. Raim: Recurrent attentive and intensive model of multimodal patient monitoring data. In Proceedings of the 24th ACM SIGKDD international conference on Knowledge Discovery & Data Mining (pp. 2565-2573).

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pyhealth-0.0.6.tar.gz (118.0 kB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page