Variational Recurrent Neural Networks (VRNNs) for Irregular, Asynchronous Clinical Time Series Forecasting

Contains the code to forecast irregular, asynchronous, clinical time series based on VRNNs, which are variational autoencoders extended to model the sequential data [1]. In the study, Medical Information Mart for Intensive Care (Version III) (MIMIC III), is utilized as a benchmark data set. The study validates the concept of utilizing external domain information to improve the generalization capability (forecasting) of VRNNs.

Introduction

This code is based on our paper, titled Exploring Clinical Time Series Forecasting with Meta-Features in Variational Recurrent Models (Ullah, Xu, Wang, Menzel, Sendhoff & Bäck, 2020), and can be used to reproduce the experimental setup and results mentioned in the paper. The code is produced in Python 3.0. The main packages utilized in this code are presented in the next section which deals with technical requirements.

As stated earlier, we want to improve the generalization capability of VRNNs for time series forecasting. For validating our thesis, we choose a widely-accepted clinical benchmark data set, which is referred to as MIMIC III. Our aim is to incorporate external domain information, e.g., in this case disease information, to improve the forecasting ability of such models. Based on this external domain information, e.g., disease information about the patients, we can link the time series to each other, i.e., time series of a particular patient in the hospital should be similar to the time series of other patients with similar disease information. Then, when learning the VRNNs, we can also incorporate the linked time series to each other, which gives rise to better forecasting than vanila VRNN.

To perform the study, we first have to pre-process the MIMIC III data set based on the code provided by [2]. After this, we have to traverse to the directory, titled Preprocess 2.0. This directory contains three jupyter notebooks, named prepare_data.ipynb, read_data_ihm.ipynb, and resample_ts.ipynb respectively. The purpose of each of these notebooks will be explained later. Here, it is important to know that this directory contains the necessary code to preprocess the data. After that, we can utilize the transformed data set to construct the VRNNs models (vanila as well as extended models based on our idea). In our study, we evaluate the performance of VRNNs models based on 1-10 step ahead forecasting (hence a total of 10 different learning tasks). The code for this is provided in three directories, which are titled 1 Step Ahead Forecasting, 2-5 Step Ahead, and 6-10 Step Ahead respectively. Each of these directories contains a total of four notebooks, titled model.ipynb, model - sim.ipynb, model - prior.ipynb, and model - sim - prior.ipynb respectively. Where model.ipynb and model - prior.ipynb implement the vanilla VRNNs describe by [2], the other two notebooks contain their extensions based on our idea of similarity, i.e., relatedness.

In the following, we describe the technical requirements as well the instructions to run the code in a sequential manner.

Technical Reuirements

In this code, we make use of four python packages (among others), which are presented below in the table. In particular, PyTorch can be utilized to implement the VRNNs. The package SciPy is utilized to sample from the probability distributions. Apart from that, pandas and scikit-learn are utilized for data transformation and wrangling. All four required packages can be installed by executing pip install -r requirements.txt from the main directory via the command line. Note that our code is only compatible with PyTorch with CPU. Additionally, we make use of the external module mimic3benchmark, which is separately provided by [1].

Package	Description
PyTorch	For implementing the VRNNs.
SciPy	For sampling from probability distributions.
pandas	For data manipulation and transformation.
scikit-learn	Also, for data manipulation and transformation.

In the following, we describe how to reproduce the experimental setup and results mentioned in our paper.

1. Initial Pre-processing of MIMIC III

The initial pre-processing of MIMIC III data set is based on the code provided by [1]. For the details on the way their pre-processing works, please refer to their work and our paper. After the initial pre-processing (based on the methodology of [1] is completed, we have to load the pre-processed data set into memory. For this purpose, please use the jupyter notebook read_data_ihm.ipynb inside the directory Preprocess 2.0. Next, re-sample the time series based on the notebook resample_ts.ipynb. Re-sampling is a crucial step since the original time series is ashychrnous, e.g., not all the temporal features are observed at the same time. In our study, we re-sample the time series to ensure that we have exactly one entry of all the temporal variables in one hour. After this, we have to prepare the data set for learning. To this end, we utilize the notebook prepare_data.ipynb in the same directory. At the end of this, the quality of the data is enough such that it can be utilized in learning.

2. Constructing the VRNNs, and Forecasting the Time Series

We can construct the VRNNs (extended and vanila) based on their goal, i.e., how many steps ahead must we forecast? In general, we have 10 different learning tasks, i.e., one-ten steps ahead forecasting. For the first task, we have the code inside the directory 1 Step Ahead Forecasting. This directory contains four different jupyter notebooks, all of which implement the four models described earlier. For forecasting 2-5 steps ahead, please refer to the code inside 2-5 Step Ahead directory, which has the same file structure as the 1 Step Ahead Forecasting. Finally, to perform 6-10 steps ahead forecasting, the code inside 6-10 Step Ahead can be utilized.

Citation

Paper Reference

S. Ullah, Z. Xu, H. Wang, S. Menzel, B. Sendhoff and T. Bäck, "Exploring Clinical Time Series Forecasting with Meta-Features in Variational Recurrent Models," 2020 International Joint Conference on Neural Networks (IJCNN), 2020, pp. 1-9.

BibTex Reference

@inproceedings{ullah2020exploring,
title={Exploring clinical time series forecasting with meta-features in variational recurrent models},
author={Ullah, Sibghat and Xu, Zhao and Wang, Hao and Menzel, Stefan and Sendhoff, Bernhard and B{\"a}ck, Thomas},
booktitle={2020 International Joint Conference on Neural Networks (IJCNN)},
pages={1--9},
year={2020},
organization={IEEE}
}

Acknowledgements

This research has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement number 766186 (ECOLE).

References:

[1] Chung, Junyoung, et al. "A recurrent latent variable model for sequential data." Advances in neural information processing systems. 2015.

[2] Harutyunyan, Hrayr, et al. "Multitask learning and benchmarking with clinical time series data." arXiv preprint arXiv:1703.07771 (2017).