A machine-learning framework for predicting outcomes from time-series history.

pip install timesias==0.0.4



Forcast outcomes from time-series history. This is the top-performing algorithm for DII National Data Science Challenge.


Install this package via pip:

pip install timesias

or clone this program to your local directory:

git clone https://github.com/GuanLab/timesias.git


For visualization:

Input data format

The example data in the data/ are randomly generated data for the demonstration of the algorithm.

Two types of data is requied for model training and prediction:

  • gs.file: gold standard file with two columns. The first column is paths for time-series records. The second column is the gold standard (0/1), representing the final outbreak of sepsis
  • *.psv: time series record files. .psv table files separated by |, which are the time-series records. The header of psv file are the feature names. To note, the first column is the time index.

Model training and cross validation

timesias -g [GS_FILE_PATH] -t [LAST_N_RECORDS] -f [EXTRA_FEATURES] -e [EVA_METRICS] --shap
  • GS_FILE_PATH: the path to the gold-standard file; for example, /data/gs.file;
  • LAST_N_RECORDS: last n records to use for prediction. default: 16;
  • EXTRA_FEATURES: addtional features used for prediction. default: ['norm', 'std', 'missing_portion', 'baseline'], which are all features we used in DII Data challenge.
  • EVA_METRICS: evaluation metrics to use. Available choices: auroc auprc cindex pearsonr spearmanr. For binary classification, AUROC and AUPRC are recommended; for regression, we recommend: C-index, Pearsonr and Spearmanr. default: AUROC AUPRC

also use:

 timesias --help

to get instructions on the usage of our program.

The above one-line command will yield the following results automatically:

  1. ./models.: where all hyperparameters of trained models will be saved.

  2. ./results: where all results mentioned below will be stored:

    1. eva.tsv: Evaluation results during five-fold cross validation.
    2. all results from top feature evaluations if --shap is used. the details will be mentioned in the next section.

Top feature evaluation

if --shap is indicated, SHAP analysis will be carried out to show top contributing measurements and last nth time points. This will generate an html report (./results/top_feature_report.html) like the following:

The corresponding shap values will be stored in ./results/shap_group_by_measurment.csv and ./results/shap_group_by_timeslot.csv.

Other applications of this method

This method can be generalized to be used on other hospitalization data. One application of this method is the COVID-19 DREAM Challenge, where this method also achieves top performance.