Stream: Covid-19 Response
Topic: Modeling risk
Josh Mandel (Mar 13 2020 at 15:55):
I wish there was case-level data. I've been noodling on the "right" bayesian treatment of the information we do get from the Wuhan data set, in terms of inferring a logistic regression based on some explicit priors + observed marginal data (e.g., summed over all HTN cases or summed over all DM cases). It's not clear to me how to do this though (I have only a tiny bit of experience in this space).
Arien Malec (Mar 13 2020 at 16:55):
Here are some observations:
1) The median age in China is 39, but the highest proportion of cases is in the 50-59 segment, which either indicates reduced transmissibility/infection the younger you are, or (more likely) the bias towards discovery of more severe cases
2) One way to attack this problem is to assume that co-morbidity is the main driver of risk, and use the prevalence figures to estimate CFR in the >60 cohort, and see where the residual risk is, and do the same assuming age is the main risk driver.
I assume this is like epidemiology 101 and we are just too dumb to apply what would be perfectly obvious to anyone else. Time for journal search.
Josh Mandel (Mar 13 2020 at 16:56):
Yes -- I have to do some reading from the 1970s -80s to come up to speed
Arien Malec (Mar 13 2020 at 16:57):
https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(20)30566-3/fulltext#seccestitle150
Arien Malec (Mar 13 2020 at 16:58):
https://jech.bmj.com/content/71/2/154
(maybe explains the male/female split)
Arien Malec (Mar 13 2020 at 16:58):
https://link.springer.com/article/10.1007%2Fs00134-020-05991-x
Carl Anderson (Mar 13 2020 at 17:11):
Sorry if I'm late to the party, but I'm wondering if we've already discussed including tobacco use in the risk model?
Josh Mandel (Mar 13 2020 at 17:18):
Smoking status is captured in the available data sets.
Josh Mandel (Mar 13 2020 at 17:19):
not captured.
Carl Anderson (Mar 13 2020 at 17:26):
Ah, so (speculating) if there is a gender difference in smoking prevalence - that might explain there being some signal in gender as a risk factor.
Arien Malec (Mar 13 2020 at 19:10):
Right, the BMJ article above documents large gender differences in smoking status.
The Lancet article documents a weak smoking status association, but the data are super suspect. In a population with a 64% rate of male prevalence, and ~5% female prevalence, and a population that's 62% male , we'd expect 41 smokers, but we only see 11 reported.
Arien Malec (Mar 13 2020 at 19:11):
Sorry, women are 3.4% ever smoked. But we are still missing a bunch of smokers in the Lancet data.
Arien Malec (Mar 13 2020 at 19:12):
But yes, discrepancy in smoking status is a more plausible driver of COVID-19 CFR than estrogen herp derp.
Arien Malec (Mar 13 2020 at 19:22):
It's possible that the simplest way to address the risk modeling question is to acknowledge there isn't enough data to model risk quantitatively, and simply model # of risk factors.
70+, smoking status, # of documented co-morbidities (CVD, diabetes, HTN, cancer, COPD)
Arien Malec (Mar 13 2020 at 20:06):
Slightly more nuanced but still qualitative:
70+/CVD are 1.5 risk points, smoking & other comorbidities are 1 risk point.
0 risk points = low risk, (0,1] risk point = moderate risk, (0,2] = high risk, (2,∞) = very high
Josh Mandel (Mar 13 2020 at 21:30):
Yeah, I've been thinking about that, and I bet it's the practical/good-enough thing to do. I'm just intellectually curious about the modeling side.
Arien Malec (Mar 13 2020 at 23:30):
KFF did a binary "is at risk" assessment, based on 60+ and presence of a comorbidity
Arien Malec (Mar 13 2020 at 23:30):
Josh Mandel (Mar 13 2020 at 23:31):
Yeah that's pretty much the spirit
Arien Malec (Mar 14 2020 at 00:06):
OK, I fixed a bunch of VLOOKUP range errors (would appreciate more eyes on this), and all the risk scores seem to line up, with some variation.
I didn't put in smoking b/c the p value was under the cutoff, but I'm pretty sure it's a factor.
Cases:
Healthy child: 9yo, no comorbidities: all low
Child with diabetes, 7yo with Diabetes: KFF and points say high risk, CFR says Medium
Child with cancer, 8yo, Cancer: same
Healthy GenZ/Millennial: 23, no dx: all low
Young Millennial with Diabetes and Cancer: CFR says High, points says Very High
Gen Y, Healthy: 49yo, no dx: All low
Gen Y, Cancer (me): 49yo, Cancer: All medium
Gen Y, T2DM + early CVD: All Very High
60, healthy: Medium (KFF TRUE)
60, diabetes: points, High, CFR Very High (could tweak cutoff here)
89, healthy: high
89, any dx: Very High
Arien Malec (Mar 14 2020 at 00:07):
This has the CFR, points, and KFF risk scoring approaches.
I'm done for now until @Farzad Mostashari gives us feedback.
Arien Malec (Mar 15 2020 at 16:50):
Lancet paper on risk factors: https://t.co/VZUD9ydzmH?amp=1
image.png
Arien Malec (Mar 15 2020 at 16:56):
Biased by being a cohort of people hospitalized.
Tiers of risk:
1) age == CVD
2) HTN
3) diabetes
4) Kidney disease, COPD
Abbie Watson (Mar 20 2020 at 00:01):
Some preliminary reporting coming in that Type A blood group may be a risk factor.
Nancy Lush (Mar 22 2020 at 03:04):
FYI, we implemented a risk assessment for stroke, based on multi-chronic issues, that also doubled as an educational piece. This was developed for a Neurologist out of CT. If it would be helpful to use this as a base and modify for this purpose, that might be an option. You can find it here: https://stroke.sunrockhealthsolutions.com/ - it is free to use. There is a similar assessment for Dementia. Reach out if that would be helpful.
Farzad Mostashari (Apr 01 2020 at 12:02):
Arien Malec said:
Biased by being a cohort of people hospitalized.
Tiers of risk:
1) age == CVD
2) HTN
3) diabetes
4) Kidney disease, COPD
Farzad Mostashari (Apr 01 2020 at 12:02):
looks like underlying health factors are indeed likely to be risks for severe COVID19 disease independent of age.
https://www.cdc.gov/mmwr/volumes/69/wr/pdfs/mm6913e2-H.pdf
we should incorporate these as rough RRs in our risk score @Josh Mandel @Arien Malec
Arien Malec (Apr 01 2020 at 19:33):
How plausible is it that only 2.6% of patients are current and former smokers? Either smoking is highly protective, or that data might be wrong.
Jason Walonoski (Apr 02 2020 at 00:17):
Or no one smokes any more, or many smokers have already died from cancer, or... Not enough good data to draw a lot of conclusions.
Last updated: Apr 12 2022 at 19:14 UTC
