Improving Medical Predictions by Irregular Multimodal Electronic Health Records Modeling

This repository contains the PyTorch implementation for the paper Improving Medical Predictions by Irregular Multimodal Electronic Health Records Modeling. This work has been accepted at the International Conference on Machine Learning, 2023.

Set up environment

Environment

Run the following commands to create a conda environment:

conda create -n MulEHR python=3.8
source activate MulEHR
pip install -r requirements.txt

Data

We uilize open-source EHR MIMIC-III to conduct experiment. This dataset is a restricted-access resource. To access the files, you must be a credentialed user and sign the data use agreement (DUA) for the project. Because of the DUA, we cannot provide the data directly.

After download MIMIC-III, we obtain time series and clinical notes following:

1. Generate time serise data following MIMIC-III Benchmarks. Note, there are five tasks in the MIMIC-III Benchmarks: in-hospital-mortality, decompensation, length-of-stay, phenotyping and multitask. We conduct experiments on in-hospital-mortality and phenotyping, which are more important based on clinicans' suggestion. Effetiveness of our model on other tasks are leveage as furture works.
2. Extract clinical note data following ClinicalNotesICU
3. Process regular time series, irregular time series and clinical notes. For example, for 48 IHM task, runs

python preprocess.py

with defualt setting.

Build your own data

To build your own task, you need a dataset of a list of instances. For each instance, the following information is required:

• name: ID number.
• irg_ts: Irregular time series matrix, which is a d_m x l_ts np.array. Here, d_m is the number of features, and l_ts is the total number of distinct time points.
• irg_ts_mask: Irregular time series mask matrix, the same shape as irg_ts. When there is an existing value in the corresponding position of irg_ts, the mask value is 1; otherwise, it is 0.
• reg_ts: Imputed regular time series matrix, which is a (d_m x 2) x l_reg np.array. Here, (d_m x 2) is the number of features with corresponding masks, and l_reg is the length of the predefined time interval.
• ts_tt: A list of time points of irregular time series, whose length is l_ts.
• text_data: A list of clinical notes, whose length is l_txt, the number of clinical notes a patient has.
• text_time_to_end: A list of time points of irregular clinical notes, whose length is l_txt.
• label: The predicted output.

Experiment

Run following bash file to conduct experiment:

sh run.sh

Citation

If you found this repository useful, please consider cite our paper:

@misc{zhang2022improving,
      title={Improving Medical Predictions by Irregular Multimodal Electronic Health Records Modeling}, 
      author={Xinlu Zhang and Shiyang Li and Zhiyu Chen and Xifeng Yan and Linda Petzold},
      year={2022},
      eprint={2210.12156},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Acknowledgements

We would like to acknowledge the following open-source projects that were used in our work:

• MIMIC-III Benchmarks for generating the time series data from the MIMIC-III dataset.

• ClinicalNotesICU for extracting the clinical notes data from the MIMIC-III dataset.

The use of these open-source projects has been instrumental in our research and we are grateful for the contributions made by their authors.