search

thibautjombart / covid19_bed_occupancy

Shiny app providing estimates of future bed occupancy given recent admissions
https://cmmid-lshtm.shinyapps.io/hospital_bed_occupancy_projections/
Other
12 stars 5 forks source link

Look for alternative distributions for LoS #49

Open thibautjombart opened 4 years ago

thibautjombart commented 4 years ago

Is mentioned elsewhere (https://github.com/thibautjombart/covid19_bed_occupancy/issues/47) we should not rely on a single set of LoS distributions. It would be good to scan the literature for new distributions, select relevant ones, and add them to the app.

thibautjombart commented 4 years ago

@mert0248 not sure if you are happy looking into this, feel free to un-tag yourself if not

erees commented 4 years ago

Did a literature search and found these different estimates of LOS which might be useful:

https://www.medrxiv.org/content/10.1101/2020.03.21.20038778v1 Tian et al. The average length of stay from onset to hospitalization was 4.1±3.7 days, and hospitalization duration average 16.1± 6.2 days. Not completely clear where patients came from, but outside Wuhan

https://jamanetwork.com/journals/jama/article-abstract/2761044 Wang et al Retrospective, single-center case series of the 138 consecutive hospitalized patients with confirmed NCIP at Zhongnan Hospital of Wuhan University in Wuhan, China, from January 1 to January 28, 2020 Among those discharged alive (n = 47), the median hospital stay was 10 days (IQR, 7.0-14.0).

https://www.imperial.ac.uk/media/imperial-college/medicine/sph/ide/gida-fellowships/Imperial-College-COVID19-symptom-progression-11-03-2020.pdf HK, Japan, Singapore, S. Korea Gaythorpe
Table 5

Zhang et al 221 patients who were confirmed diagnosed as COVID-19 according to WHO interim guidance, from January 2 to February 10, 2020 at Zhongnan Hospital of Wuhan University,Wuhan, China https://www.medrxiv.org/content/10.1101/2020.03.02.20030452v1.full.pdf Table 5 Duration of ICU stay (days) ICU to ward tranfers: 8.0(5.0-13.0); Death in ICU: 11.0(4.5-14.5)

Yang et al https://www.thelancet.com/journals/lanres/article/PIIS2213-2600(20)30079-5/fulltext Other durations, but not length of stay explicitly - but might be possible to work out from other numbers reported.

erees commented 4 years ago

Also possible to caculate LOS from deaths (uk data) but will be biased to more severe cases?

mert0248 commented 4 years ago

@erees thanks I've been trying to collate results into a word doc an will share as a googledoc on slack.

mert0248 commented 4 years ago

Also possible to caculate LOS from deaths (uk data) but will be biased to more severe cases?

Which UK dataset? But yes will be biased if only deaths.

erees commented 4 years ago

Also possible to caculate LOS from deaths (uk data) but will be biased to more severe cases?

Which UK dataset? But yes will be biased if only deaths.

The dataset available in covid19 automation. I think it is all reported uk deaths but has admission dates and death date.

johnurbanik commented 4 years ago

Pulling it back in here, but the wuhan empirical dataset suggests mean of 15-17 days:

https://github.com/understand-covid/epi-modeling-toolkit/blob/master/parameter%20estimation/hospital_stay_analysis.ipynb

I'm inclined to believe that the studies coming from Wuhan are likely to be off, especially given the information intelligence report about data manipulation in China Bloomberg.

The study for outside of Hubei seems (slightly) more promising, and it's possible that we could back out an empirical mean from the same dataset as I used above (which has Guangzhou cases: https://github.com/c2-d2/COVID-19-wuhan-guangzhou-data).

If anyone aware of more datasets that include time series of hospitalizations, it's possible that one could fuse those datasets with the JHU CSSEGISandData data for deaths and recoveries (where countries have been reporting it).

thibautjombart commented 4 years ago

As of 8bae660e39762336871793c1202d47c63987f322 the app defaults to a custom distribution, with the 2 Zhou et al as other options. Further additions will be merely incremental, provided the distribution is already defined.

Also see this issue, which will be relevant for creating new distributions from published info.

samclifford commented 4 years ago

Is there a reason that the coefficient of variation maxes out at 1? Having CV > 1 gives you distributions which have a mode at the far left extreme. Also how well do small CVs play with the mean-1 trick, particularly at small mean lengths of stay?

samclifford commented 4 years ago

With the spin-off of the length of stay review into a new paper (https://www.medrxiv.org/content/10.1101/2020.04.30.20084780v1) can we close this