Ensemble Learning for COVID-19 Risk Prediction

• implemented 7 prognosis risk prediction models for COVID-19. Detailed info in this paper: DOI:10.1093/jamia/ocaa295
• introduced a competence quantification framework for assessing the competence/confidence of a model in predicting a given data entry (i.e. a digital representation of a covid patient)
• ensembled 7 prediction models for prediction using fusion strategies based on their competences
• evaluated single models and the ensembled mode on two large COVID-19 cohorts from Wuhan, China (N=2,384) and King's College Hospital (N=1,475)

usage

Prerequisite

install python libraries using requirements.txt

pip install -r requirements.txt

Setup and Run

1. edit the experiment setting file: ./test/test_config.json, changing the following parameters accordingly.
   • data_file - the csv file which is the patient level data for prediction. check full list of variables below, if numeric columns are missing, they will be automatically imputed using model related distributions.
   • (optional) sep - the separator, changing it to comma if your file is comma separated.
   • (optional) mapping - a dictionary to map column names
   • comorbidity_cols - the list of column names (binary valued) denoting comorbidities. NB: if you don't have morbidity_Hypertension (a mandatory variable), please add an entry to binary_columns_to_impute. This will populate all zero column for binary_columns_to_impute.
2. run the models

   python test_util.py

3. check the result files, which are to be saved to ./test folder.

   # performance tables
   death_result.tsv
   poor_prognosis_result.tsv
   
   # figures
   auc_fig_death.png
   auc_fig_poor_prognosis.png
   calibration_fig_death.png
   calibration_fig_poor_prognosis.png

variables

*demographics*
age
Male

*underline conditions*
morbidity_Hypertension

*bloods*
Lactate dehydrogenase
Albumin
C-reactive protein
Serum sodium
Serum blood urea nitrogen
Red cell distribution width
Lymphocyte count
Neutrophil count
Direct bilirubin

*vitals*
Oxygen saturation

*outcome (binary, 1 means event happened)*
death
poor_prognosis - defined as either death or ICU admission

generate cohort descriptive table

1. change the display configuration file ./conf/ensemble_vis_configs.json to specify where your data file is by changing raw_data value there.
2. run the following script

   python fig_tab_generator.py

3. result file will be at test/desc_tab1.tsv.

contact

email: [email protected]