search

crowdfightcovid19 / req-550-Syria

Repository to host the code of request 550 coming from the Pax Syriana Foundation
Other
3 stars 2 forks source link

Estimation of parameters #8

Closed apascualgarcia closed 4 years ago

apascualgarcia commented 4 years ago

Hi @Jennifer-Villers, @jordan-klein, @ecam85 and @Chamsy-Sarkis

Jennifer and Jordan, for launching massively the simulations I will need confidence intervals in which I will randomize each parameter within the range given (if possible indicate if it is e.g. 95% CI).

Jennifer, could you please how is the map between the parameters provided and the ICL paper? I guess you used data from Table 1 but I do not see a clear mapping. Ages for us are 0-13, 14-50, >50 right?

For everyone, I am not sure either how legitimate is to translate the hospitalization parameters to our model due to the following caveats, let me know what you think.

Jennifer-Villers commented 4 years ago

I have a question regarding the confidence intervals. Based on the available data, the duration of the latency period was calculated as the difference between the duration of the incubation period (5.2 with 95% CI 4.1-7) and the pre-symptomatic period (2.3 with 95% CI 0.8-3.0). How do I calculate the CI for the latency period?

Jennifer-Villers commented 4 years ago

Hi @Jennifer-Villers, @jordan-klein, @ecam85 and @Chamsy-Sarkis

Jennifer and Jordan, for launching massively the simulations I will need confidence intervals in which I will randomize each parameter within the range given (if possible indicate if it is e.g. 95% CI).

I have modified the file to add 95% CI or interquartile range (IQR) for age-independent parameters. For some parameters, I have different estimates based on different publications. Some parameters are calculated as the difference between two other parameters, each of them having its own 95% CI or IQR. I don't know how to calculate the CI or IQR of the calculated parameter and would appreciate some help. (https://github.com/crowdfightcovid19/req-550-Syria/blob/master/docs/Parameter%20estimates.md)

Jennifer, could you please how is the map between the parameters provided and the ICL paper? I guess you used data from Table 1 but I do not see a clear mapping. Ages for us are 0-13, 14-50, >50 right?

Please check with Jordan (@jordan-klein). I am taking care of the parameters that are age-independent only.

For everyone, I am not sure either how legitimate is to translate the hospitalization parameters to our model due to the following caveats, let me know what you think.

  • The criteria used in Western countries to say that a patient requires hospitalization will possibly not apply here (more capacity implies a more conservative criteria).
  • Even if the criteria are the same, the capacity of the hospitals will collapse earlier.
  • Therefore, I don't think we can use the same hospitalization rates. Let's see what are the numbers and criteria that the Health Directorate provides for hospitalization and quarantined, but if the fraction of people that can be hospitalized is small, considering this compartment will complicate things because it will require a time-dependent parameter.

I agree that hospitals getting overwhelmed is a problem, especially accounting for the fact that once the disease reaches the camp, it probably already has reached the cities and villages around where people will also need hospitalization. Even if we use a time-dependent parameter, it will be hard to estimate based on the modeling of one camp how overwhelmed the hospitals might be due to the propagation of the disease in the surrounding cities/villages.

Perhaps should we start by determining how much bigger is the number of people who would need hospitalization compared to the number of ward beds available?

  • Even if we consider it, as soon as the hospitalized collapse we must decide what will happen with all those cases that would go to the hospital but that must stay in the camp because there is no hospital space. We can say that the critical cases will die (actually we do not have a critical compartment) but we cannot say that all the cases that would have been hospitalized following Western rates will die (a fraction may neither become critical).

Regarding the course of the disease following hospitalization, even people who don't need critical care or a ventilator may die. The number one cause of death for younger people (< 50 yo) with covid is thrombosis or pulmonary embolism due to blood clots. You don't need ICU to treat that but you need blood thinners and sometimes surgery, both of which usually being administered at the hospital only.

  • So the alternative I see is either to create a new compartment between symptomatic and dead (different than hospitalized that will be a neglectable fraction), but we do not have parameters for those, or to increase the recovery rates for symptomatic (and also the question is how much).
  • We should also take into account that each new compartment/parameter will add nonlinearities and noise to the model so will reduce the confidence intervals of the prediction. So my position is to avoid adding variables unless we have good estimates. For instance, having a single recovery rate and death rate from the symptomatic compartment with a larger confidence interval for the parameters may bring more robust estimates than having more compartments behind the symptomatic one.
  • Another thing is if a given "age" in a Western country can be mapped to a Syrian one. This will possibly be circumvented if we gather comorbidity data.

Thank you for starting this discussion. I generally agree with your points and I don't have a good answer, except perhaps starting by estimating what the difference is between the number of hospital beds available and the number of beds we expect to be needed at the level of the entire population of NW Syria.

apascualgarcia commented 4 years ago

Hi @Jennifer-Villers, thanks for the answers!

Regarding the CI of the difference. If X and Y are two normal randomly distributed variables the variance of the new stochastic variable Z=X-Y is equal to the sum of the variances of X and Y. So the procedure I guess would be: from the CI of X and Y you estimate the standard deviation of each of them, then compute the variance of Z, and from that its CI. BUT, this works if the distributions are normal, however it doesn't seem to be the case because the intervals seem to be left-tailed. So let's leave this question in standby for the moment because I should also think how to generate random numbers from this uneven CI, and perhaps the easiest way to address the CI of Z is numerically.

Regarding the Hospital and Isolation Center compartment(s), the situation is the following. We have this model:

Model

Here a fraction q of the population is pulled from the camp at a rate \eta (we are agnostic on where this people will be evacuated, it might be to a hospital or isolation center) and it is assumed that the more infections the more people will be removed. Considerations:

So what I would suggest is to directly try to estimate a larger \gamma{I} and \alpha{I} (perhaps like you did with P, aggregating the posterior stochastic variables, which would mean adding the means and the variances), then assume that a number of people will leave the camp per day (whatever the evacuation capacity is) and either consider only one variable for them or not model them at all.

I see the point of modelling all the people leaving the camp and their return for questions like herd immunity. But I think that this estimation, although important, might not be the urgent estimation at this point (herd immunity may take several months and the decision of whether a rearrangement worths it will be made on the basis of what happens in <2 months) so we could just focus on the fact that this people is no longer in the camp and hence reduce infectivity.

Other thoughts? I call also @JudithBouman2412 that is not in the thread.

Jennifer-Villers commented 4 years ago

The proportion that would go to isolation centers for mild cases is even trickier to calculate because, as Chamsy explained, only people whose test came back positive would be sent to isolation centers. Until their tests come back positive, they will have plenty of time to infect other people in the camp and they may not even get tested at all if testing capacity is insufficient. In an attempt to prevent that, Chamsy has suggested to make isolation centers inside of the camp based on symptoms such as fever and cough but that could come with a large number of false positives as fever is a frequent symptom for other diseases that are prevalent in camps.

Overall, perhaps I would leave external isolation centers for mild cases out of the equation (for the reasons described above). We may instead include one intra-camp isolation center based on symptoms that would help decrease the contact rates between symptomatic people and the rest of the camp (but contact would still not be zero).

Regarding hospitals, I don't know yet how to deal with that. I feel like people who need to go to the hospital are likely to die if they don't get access to it, so they would either be taken away to the hospital or not survive. In either case, once the disease is serious enough to require hospitalization (by occidental standards), they should be removed from our total camp population. And I agree that once they're out, their recovery time may not be that important in the timescale we're interested in.

One option could be to assess the efficiency of the shielding strategy based on the number of hospitalizations (instead of the number of deaths). Since those numbers are correlated and since our strategy won't have any impact on the survival chances of those who need hospitalization, it might be enough to determine whether it makes sense to shield vulnerable people or not.

***ADDED from slack: About camp populations being more vulnerable, one option that they used in the African modeling paper is the following: "To account for this, we shifted age-specific severity risks (probability of becoming a severe case and critical case) towards younger age by 10 years." I don't know if that would capture the differences between western countries and Syrian camps or if we would be better off with real comorbidity estimates, but it is an option we can consider in the absence of good comorbidity data

jordan-klein commented 4 years ago

I've provided updates for my parameter estimates with confidence intervals: here

For the age specific hospitalization rates, I state my assumptions in my code: -Assume average in age group 0-12 approx % aged 0-9 requiring hospitalization in paper -Assume average in age group 13-50 approx % aged 20-29 requiring hospitalization in paper -Assume average in age group over 50 approx % aged 60-69 requiring hospitalization in paper

apascualgarcia commented 4 years ago

Chamsy notes:

Normal process if suspected case in a camp: 1- What is the process for suspected cases? First, the patient goes by himself to a medical camp outside of the camp clinic/hospital/covid center. Each medical camp has been organize with 3 areas: first a triage tent (temperature check). If positive at triage center, the patient goes to an observation tent, where he meets a doctor who does a brief clinical diagnosis. If thought positive, the doctor calls a “Covid-specific ambulance” to transfer the patient to: • Hospital for severe cases • Community Based Isolation center (CBI) for mild cases

What is the process at hospitals, and what are the figures about hospitals capacity? At hospital, there are PCR tests that are performed (currently one PCR machine in the NW, soon will be 3 machines, for a total capacity of 300-500 tests per day at max). If positive, patient stays at hospital during all treatment. I did not ask about when the patient should return to the camp, and whether the patient is quarantine after treatment in CBI. Currently, the capacity to receive Covid-19 patients in ICU beds with ventilators is 32 in total. The plan A is to set 3 Covid-19 care centers with 30 ICU beds with ventilators and 40 ICU observation beds, without ventilators. So in total, there should be 90 ICU beds with ventilators and 120 ICU beds without ventilators. Timing: plan A should be set around July (optimistic, personal thought).

Details about Community-based-isolation (CBI) centers 1- Hosting capacity of CBI centers A CBI capacity is around 50 beds (just beds, not ICU beds). The current capacity is 70 beds only, should be soon extended to 140 beds. The objective is to have 1500 beds distributed in 30 CBI centers. CBI centers will be placed in the areas with the highest concentration of IDPs (i.e. in areas with a lot of camps). According to health directorate, it is expected that in June, 50% of the CBI capacity is implemented, and fully implemented in July. Again, I think this is very optimistic. CBI capacity 70 beds now (should be extended to 140) for observation => in total 30 centers 1500 beds (in future, 50% in june 100% in July if according to expected). 15-20 days in quarantine. Phase II, when outbreak starts = doubling capacity. When the outbreak starts, there is a phase II for Plan A, with the objective of doubling the capacity (timing is not known, as Plan A was not set up in due time…). There is also a phase III, for which I could not get figures.

2- Measures in CBI CBI have three compartments: observation, treatment and recovery. Persons hosted in CBI should stay for 15-20 days if positive. When released, they return to the camp. They may also be transferred to hospital if case becomes severe and requires ICU (with or without ventilator). People who want to leave the CBI, should not be allowed to do it, but there will be no enforcement for that. Concretely, people can go out if they wish.

Data about comorbidities Before the war, cardiovascular diseases and diabetes was present in 11% of the population (I guess that he meant adult population). Exact numbers in camps are unknown, but they are much higher, due to general deterioration of health care and lack of patient follow-up. There will be specific measures for people affected by kidney diseases, tuberculosis, asthma and respiratory diseases (but not diabetes, cardiovascular and hypertension, because there are too many): when these people go to a medical center, they will be tested in routine by PCR (an average of once a month).

apascualgarcia commented 4 years ago

I am going to try to summarize the agreements on the variables and parameters and the questions still open.

  1. The model is age-structured with three age layers: 0-13, 14-50 and >50. Given the living conditions we will take parameters from the literature that consider populations 10 years older for each interval.. @jordan-klein please confirm that the parameters provided follow this criteria.

  2. We will consider one additional class for age 2 for comorbidities because they seem to be high. According to Chamsy comorbidities in the country are around 13% but in the camps "much higher". In this issue we should discuss the percentage and parameters for this class and other possibilities.

  3. The model has a pathway S --> E --> P --> I. The parameters S-->E are known and E-->P can be inferred from the incubation period, as discussed in this issue.

  4. An optimistic scenario rises the number of beds in Isolation Centers + Hospitals to 1500 beds for 4M inhabitants, which is ~4 beds for 10K. Since the informal camps are 2K, in a global outbreak the camp will have access to <1 bed, so we can assume that the existence of these facilities can be neglected.

  5. As a consequence, all cases should be handled within the camp. This opens a challenge to adapt the parameters found in the literature to this new scenario, e.g. we cannot use the fatality rates of hospitalized cases for all the infectious cases. I will propose in this issue how to circumvent this problem.

  6. The contacts matrix parameterization and the analytic estimation of R0 are the elephant in the room. We proposed to use R0 values from the literature and solve for beta. However, I think it would be better to estimate beta from the literature and solve for the average rage of contacts. I discuss this in this issue.

If nobody raises objections I will close this issue so that we can move the open discussions to the more specific issues opened in the links.

Jennifer-Villers commented 4 years ago

@apascualgarcia Hi Alberto, thank you very much for the summary!

I just wanted to clarify your point #3:

jordan-klein commented 4 years ago

@apascualgarcia

  1. Confirming that the parameters follow this criteria (for proportion requiring hospitalization, the only age-dependent parameter in the model)

  2. I analyzed the Syrian & Jordanian data & estimate about 17% of the adult population (over 12) has a comorbidity. The population structure for the model (3 age groups, comorbid yes/no) is here