Medical time series datasets

This module contains the implementation of multiple medical time series datasets following the tensorflow dataset API.

Currently implemented datasets are:

• physionet2012 (mortality prediction)
• mimic3_mortality (mortality prediction)
• mimic3_phenotyping (mortality prediction)
• physionet2019 (online sepsis early prediction)

Example usage

In order to get a tensorflow dataset representation of one of the datasets simply import tensorflow_datasets and this module. The datasets can then be accessed like any other tensorflow dataset.

import tensorflow_datasets as tfds
import medical_ts_datasets

physionet_dataset = tfds.load(name='physionet2012', split='train')

Instance structure

Each instance in the dataset is represented as a nested directory of the following structure:

• statics: Static variables such as demographics or the unit the patient was admitted to
• time: Scalar time variable containing the time since admission in hours
• values: Observation values of time series, these by default contain NaN for modalities which were not observed for the given timepoint.
• targets: Directory of potential target values, the available endpoints are dataset specific.
• metadata: Directory of metadata on an individual patient, such as the identifier using in the dataset.

Supervised dataset

If the load method is called with the flag as_supervised=True, it will return a dataset which can readily be used together with keras. Here each instance is represented by a (X, y) tuple and the X tuple contains the following 4 elements: time, values, measurements (indicators if a value was measured or not) and length.

Citing

When using datasets from this repository, please do not forget to cite the following references besides a reference to the original dataset:

@InProceedings{horn2020set,
    title = {{S}et {F}unctions for {T}ime {S}eries},
    author = {Max Horn and Michael Moor and Christian Bock and Bastian Rieck and Karsten Borgwardt},
    booktitle = {Proceedings of the 37th International Conference on Machine Learning},
    year = {2020},
    volume = {119},
    series = {Proceedings of Machine Learning Research},
    publisher = {PMLR}
}

@misc{TFDS,
  title = {{TensorFlow Datasets}, A collection of ready-to-use datasets},
  howpublished = {\url{https://www.tensorflow.org/datasets}},
}