For a set of Covid+ patients (PCR positive), images were extracted from the Radiology PACS at Stony Brook Medicine and de-identified using POSDA. Images were matched with clinical data from the local Covid Data Commons. The Covid Data Commons is based on data captured from the electronic health records (EHR) at Stony Brook Medicine and manual review of clinical charts.

The main data file is named ‘deidentified_overlap_tcia.csv.cleaned.csv’. The file contains one row per patient whose images have been extracted. For each patient one encounter is selected using an algorithm (see “Encounter/visit selection steps” below for more detail). The algorithm is designed to select the Covid+ encounter where the patient had their most severe encounter. Images should be interpreted and aligned with the date-shifted field visit_start_datetime to correlate severity with the imaging data.

Clinical Data key

A description of fields in the de-identified files are provided in the file named ‘deidentified_overlap_tcia.csv.cleaned.csv.template.csv’. The column in the description file is_chart_abstracted indicates whether the column is derived from the manual chart review. Some field names are descriptive and so no additional information is provided. For laboratory and vital measurements the first value for the patient is selected.

Values of NA indicate that the value is missing, TRUE is a boolean True, FALSE is a boolean False. Original encoding from the source data of {Yes, No} are preserved in the final file. Some numeric measurement fields are constructed as: 2075-0_Chloride [Moles/volume] in Serum or Plasma where 2075-0 is the LOINC code and Chloride [Moles/volume] is the description associated with the LOINC code. LOINC codes and descriptions can be found on the LOINC website, for example, 2075-0.

Encounter/visit selection steps

The first steps of the algorithm is to find Covid+ patients and their potential encounters associated with infection:

1. Apply date cut-off of February 1, 2020 for either the start or the end of an encounter.
2. Remove future visits and remove any non-discharged (active) encounters.
3. Identify the patient encounters where there are Covid+ PCR tests.
4. Select visits which occur up to 7 days after the Covid+ PCR test.
5. Identify Covid+ patients with encounters who have the ICD-10 code (U07.1) for Covid-19 virus identified.

In the second part of the algorithm we filter the encounters down to a single encounter –  the most severe encounter:

1. If a patient has only one encounter select this encounter.
2. If a patient has multiple encounters, first select the inpatient encounters.
3. If the patient has remaining encounters, select the hospital observation encounters.
4. If the patient has remaining encounters, select the emergency department encounters.
5. If the discharge disposition is death or hospice for an encounter, select that encounter and drop the others for that patient.
6. If there is an encounter where the patient required invasive ventilation or ECMO, select that encounter.
7. Pick the encounter with the longest length of stay.
8. If there are still multiple encounters remaining for a patient, select the most recent one.

Data collection was enabled by the Renaissance School of Medicine at Stony Brook University’s “COVID-19 Data Commons and Analytic Environment”, a data quality initiative instituted by the Office of the Dean, and supported by the Department of Biomedical Informatics.