Datathon: Predictors of severe COVID-19 outcomes

1. Characterize the BCM experience with COVID, including hospitalization rate by comorbidity, and ICU utilization.

2. Train a multivariable predictive model for severity of COVID (ICU admission, incidence of ARDS criteria, length of stay, and in-hospital mortality) as a function of known and novel factors such as comorbidity.

3. Study the covariation in severe outcomes with treatment modalities to evaluate population-level changes.

The file you probably want

bslmc_v6.R (click on it up above).

Routputs_v6.txt for results (although you should treat them as proof of concept only and not believe them. Not even hypothesis-generating.)

Also I feel the outputs/makefile file is useful and nifty.

Analytic data set sketch

• one row per ED encounter
  • patient ID of course
  • ER visit diagnoses
  • admission diagnoses, if admitted
  • covid test date(s) & results
  • ER visit index date
• outcomes as follows:
  • admitted yes/no
  • ER directly to ICU yes/no
  • length of stay, if admitted (continuous)
  • pao2:fio2 ratio (future summary measure, of oxygenation)
  • mortality yes/no (and date)
  • intubated yes/no (and date)
  • maybe future ICU admit & date if I can manage it
• predictors as follows
  • demographics
    • age
    • sex
    • race
    • ethnicity
    • ZIP
  • comorbidities (two columns for each: N prior visits or rate, and prob list yes/no)
    • diabetes
    • copd
    • asthma
    • hypertension
    • coronary disease
    • cancer
  • number of prior hospital admissions (or rate)
  • number of prior ER visits (or rate)
  • vitals (summary meas if needed)
    • temp
    • pulse
    • respirations
    • BP
    • SpO2
    • height
    • weight
  • labs (summary measure of labs just before/on index date)
    • wbc
    • hgb
    • plt
    • sodium
    • K
    • bicarb
    • creatinine
    • d-dimer
    • CRP
    • LDH
    • BUN
    • HDL
    • direct bilirubin
    • RDW
    • albumin
    • neutrophils
    • lymphocytes
    • ALT
    • P:F ratio can be predictor for more "hard" downstream events

Data pull spec

Include: anyone with positive covid test. BSLMC and BCM outpatient. If more than 10,000 patients, OK to randomly sample. (Seems to be 1100 to 1800). 700 ish at office visit in person.

Tables (inpatient/BSLMC):

• PAT_ID

• ENC_DX including distant past

• PROBLEM including distant past

• ORDER_RESULTS only need recent dates like 2020

• HSP, including all types of visits (ER, inpatient, etc.). OK to limit to 2020 only, but not necessary.

• possibly flowsheet: height weight systolic diastolic pulse respirations spo2. limit to intubated or other o2 params, and spo2 po2.

• discharge disposition (looking for discharge disposition)

Tables (outpatient):

• TBD

Where to look in this code

High value areas would be bslmc_v4_DataSets_pipe.R and bslmc_v6.R for inpatient data and analysis_outpat.R for outpatient. Pay the most attention to variables & values mentioned within select() and filter() statements to get a sense of how the data is structured.

Requirements for current repo

• R, Rscript
• R packages: dplyr, ggplot2, tidyr, lubridate, here, earth, ROCR
• make and usual UNIX-like toolchain (mv, rm, cp)
• pandoc (only for documentation)

Datathon "alpha phase" use case examples

ID	Hard?	Waiting on:	Description
1	Easy	Done	Test volume, positive tests, by date. Foundational.
2	Easy	Done	Count tests, positive tests, by comorbidity (see below).
3	Easy	Andy	Retest volume, likelihood of positivity. By clinic.
4o	Int.	Done	Pulse oximetry (SpO2) by positive/negative.
4i	Int.	Done/Rory	" " "
5	Int.	Andy	Basic labs (see below) by positive/negative.
6	Adva.	Rory	I: People who "touch" the chart (PPE estimate).
7	Adva.	Andy	I: Rate of testing late in an admission, rate of positive.
8	Adva.	Rory	I: Basic descriptives (LOS, floor/hosp census by date).
9	Adva.	Andy	O: Rt. of admission/ER (manual rev.?); risk factors.
10	Adva.	Andy	Anything to do with mortality.

Capture what Baylor Medicine clinic it was sent from. Capture what lab (vendor) it went to (LabCorp, or whoever), called "test perfomed by." Most common "lab facility" are CPL and LabCorp. Slicer has the confirmed and the suspected registries for COVID.

Define "comorbidity"

1. The "big four": Asthma, COPD, DM, HTN.¹

   a. Define first using just the problem list. This is "middle school level."

   b. Then define using encounters, "high school:" anyone with 2 or more encounters, from the beginning of time, is defined has having the diabetes (copd, htn, etc.) phenotype. Also note that we want to know the value of the counts: really the distribution of the count over patients. (100 patients have 2 COPD encounters, 10 have 3 encounters, and 1 patient had 10 encounters, since beginning of the database, which means 5 years.) Final note: for big four, this depends only on the encounters, not on problem list. (A woman with 3 COPD office visits in past year should be considered as having COPD according to this rule, even if she does not have it in the problem list.)

   c. Then meds, "college level." Requires increasingly more work.

   d. Then more fancy, such as moving beyond rules based phenotype definitions. Probably don't do this for purposes of this stage in the datathon.

2. Also want a broad view of all "medical history" items in problem list (not just big four). This is shown on one of Gloria's SlicerDicer slides as a bar graph.

Epic has notion of encounter for office, procedure, and rx. Gloria knows someone for data quality of tobacco use.

Define "labs"

Labs are called "procedure." Labs are of definite interest. Big ones (values are interesting): cmp, cbc, flu, bmp, crp, d-dimer, ferritin. Within those lab panels, the values I care most about are: sodium, creatinine, and white blood cell count, detailed white cell differential (neutrophils, lymphocytes, monocytes, eosinophils, basophils).

In broader "non big-one" sense, I do want to know whether they have the lab ordered (e.g. A1c, lipid panel: I don't care so much about the exact value, but I want to know whether or not the patient had the lab ordered/done). This is shown on one of Gloria's SlicerDicer slides.

Footnotes

1. Epic Health Research Network, https://ehrn.org/ ↩