Early Triage of Critically-ill COVID-19 Patients using Deep Learning

This package provides an implementation of the prediction and analysis of early triage of critically-ill COVID-19 patients using deep learning.

More details can be found in paper: https://www.nature.com/articles/s41467-020-17280-8

Setup

Dependencies

Processing pipelines are implemented in python. If there is missing value, data imputation requires R.

Python 3.6+

• pandas 0.23.0
• tensorflow 1.14
• sklearn 0.21.3
• sksurv 0.11
• scipy 1.4.1

R 3.6.2+

• mice_3.8.0
• dplyr_0.8.5

Data

Data are all in csv format. Results will be saved in csv format as well.

Input file format

A csv file in the following formate is needed for prediction:

patient_ID	X.ray.abnormality	Dyspnea	Number.of.comorbidities	Malignancy	COPD	Age	NLR	Lactate.dehydrogenase	Direct.bilirubin	Creatine.kinase
1	1	0	0	0	0	59	3.428571429	201	1.7	26
6		0	0	0	0	65	6.545454545	255	3.3
9		0	4	0	0	61	3.888888889	211	3.2	113

If value is unknow, you can leave the cell empty. A imputation method will be used to fill the value. But please note that complete values will result in better prediction performance.

Explaination of each value is as following:

column	data_type	unit	explain
patient_ID	any		ID of patients
X.ray.abnormality	binary	0/1	1 indicate x ray abornmality
Dyspnea	binary	0/1	1 indicate dyspnea
Number.of.comorbidities	intiger	0-9	count of comorbidities
Malignancy	binary	0/1	1 indicate cancer (maligant)
COPD	binary	0/1	1 indicate COPD
Age	intiger	0-150	age of patient
NLR	float	ratio	Neutrophil-Lymphocyte Ratio (NLR)
Lactate.dehydrogenase	float	U/I	Lactate dehydrogenase test result
Direct.bilirubin	float	μmol/l	Direct bilirubin test result
Creatine.kinase	float	U/l	Creatine kinase test result

For more details, please refer to paper.

Truth file format

To evaluate the result, a right censored file containing the truth of each observation is required. A csv file in the following format is needed:

patient_ID	critically_ill	critically_ill_time
1	0	5.0
2	0	10.0
3	0	6.0

All values are needed and the samples need to be in the same order as input csv file (patient_ID will NOT be checked in analysis). Explaination of each value is as following:

column	data_type	unit	explain
patient_ID	any		ID of patients
critically_ill	binary	0/1	1 indicate critically-ill patient, right censored
critically_ill_time	float	days	number of days from data collection to the event

Truth file is not needed if you only want to predict the survival ratio.

Model checkpoints

All models should be stored in models fold. Please unzip the models.zip file and make sure the files follows rule:

• code_root
  • models
    • fold_0
      • model.meta
    • fold_1
      • model.meta
    • fold_2
      • model.meta
    • fold_3
      • model.meta
    • fold_4
      • model.meta
    • imputation.csv

Preprocess

Preprocessing code will impute missing value and normalize results. To run the code:

python preprocess.py input.csv

"input.csv" can be replaced with any csv file that meets input file format requirement. A file with name [INPUT]_processed.csv will be saved containing the processed file. Use this file as the input of next step.

Prediction

Prediction code will predict risk scores and survival probability for each sample.

python prediction.py input.csv_processed.csv

"input.csv" can be replaced with the name of your input file. A file with name [INPUT]_processed.csv_prediction.csv will be saved containing the prediction results. Use this file as the input of next step.

5 values are generated:

Value	Explaination
score_dl	risk score of deep learning model
score_final	final risk score of nomogram
survival_5days	5 days critically ill probability
survival_10days	10 days critically ill probability
survival_30days	30 days critically ill probability

Evaluation

Evaluation code will analysis predictions. To run:

python evaluation.py input.csv_processed.csv_prediction.csv truth.csv

"input.csv" can be replaced with the name of your input file. "truth.csv" can be replaced with the truth file that meet the format requirement. A file with name [INPUT]_processed.csv_prediction.csv_evaluation.csv will be saved containing the evaluation results. Use this file as the input of next step.

Following analysis will be conducted on "survival_30days" with 95% confidence inter computed as well:

Value	Explaination
c-index	C-index of survival analysis
auc	ROAUC value assuming all right-censored samples are negatives
recall	recall value with different cut-offs assuming all right-censored samples are negatives

Disclaimer

This tool is for research purpose and not approved for clinical use.

This is not an official Tencent product.