Introduction

This repository contains reproducible code for our paper, Identifying who has long COVID in the USA: a machine learning approach using N3C data, which uses data from the National COVID Cohort Collaborative’s (N3C) EHR repository to identify potential long-COVID patients. If you use this code in your work, please cite:

Pfaff ER, Girvin AT, Bennett TD, et al. Identifying who has long COVID in the USA: a machine learning approach using N3C data. Lancet Digital Health. 4(7),E532-E541. doi:10.1016/S2589-7500(22)00048-6

Note that the code on the main branch of this repository has been updated since the publication of the paper, and will continue to be updated. For a snapshot of the code as it stood when the paper was published, please use the "published_paper_code" branch.

Purpose of this code

This code is designed to identify possible long COVID patients using electronic health record data as input. As of 7/11/2022, our feature table engineering code and our pretrained model are available in this repository. The model and its intent are described in detail in the paper linked above.

Prerequisites

In order to run this code, you will need:

• EHR data in the OMOP data model
• At least some COVID positive patients in your data, indicated through positive PCR or antigen tests (LOINC-coded) or U07.1 diagnosis codes.
• The ability to run Python against your OMOP data model

The SQL code in this repository is written in the Spark SQL dialect. If you have a different RDBMS, most of the SQL will work but you will likely need to swap out a few functions here and there. The Python code in this repository is written in PySpark. As PySpark is not very common, we will provide a pandas translation soon (see Future version notes).

Running our code

This repository is intended to be run in a stepwise fashion, using the numbered folders. (I.e., first run all the scripts in 1_, then 2_, and so forth.) Each numbered folder has its own README inside with additional details.

Future version notes

In our next update, we will:

• release code to enable users to retrain the model on their own data, rather than using ours.
• provide a pandas translation of the PySpark model code for ease of use
• provide a list of Python packages/versions in the "Prerequsite" section above