A Machine Learning Model for Predicting Deterioration of COVID-19 Inpatients

Public repository containing research code for the COVID-19 prediction model, described in the manuscript "A Machine Learning Model for Predicting Deterioration of COVID-19 Inpatients"

Authors:

Omer Noy*, Dan Coster*, Maya Metzger, Itai Attar, Shani Shenhar-Tsarfaty, Shlomo Berliner, Galia Rahav, Ori Rogowski, Ron Shamir

Code:

Install requirements

python -m pip install -r requirements.txt

This code was tested with python 3.7.

Modules

The repository is organized as follows:

• anomaly_scores Anomaly detection approaches, used as anomaly features
• data_preprocessing Code for preprocessing the parsed data, including time-series formatting, imputation, etc.
• feature_generation Code for features engineering, including historical summary statistics and trend features.
• feature_selection Code for feature selection strategies.
• ml_models ML model classes, including pre-training processing, fit and evaluation methods.
• outlier_removal Values Removal according to predefined clinical ranges.

Data:

The data used in our study cannot be shared. This section describes the data format used for the code.

Our data-specific parser generates 3 main pandas dataframes:

• Baseline_df Contains demographics and background disease (static features).
• vital_df Vital signs (longitudinal features).
• labs_df Lab tests (longitudinal features).

Baseline dataframe format (Baseline_df):

Columns	Data type
Patient ID	object
Admission Date	datetime64[ns]
Gender	bool
Age	float64
...	String
Background diseases	bool

Longitudinal dataframes format (vital_df, labs_df):

Columns	Data type
Patient ID	object
Date Time	datetime64[ns]
Feature Name	String
Value	float64

Using data_preprocessing/create_time_series_data, the dataframes can be merged and pivoted into a time-series format, with columns representing features and rows representing the longitudinal patients' observations.