Eclipse Community Forums
Forum Search:

Search      Help    Register    Login    Home
Home » Eclipse Projects » Spatiotemporal Epidemiological Modeler (STEM) » Handling multiple diseases in STEM
Handling multiple diseases in STEM [message #3224] Fri, 22 May 2009 17:25 Go to next message
Stefan Edlund is currently offline Stefan EdlundFriend
Messages: 127
Registered: July 2009
Location: IBM
Senior Member
As a first step towards handling a true multi-serotype scenario in STEM,
we need to correctly handle two or more isolated diseases circulating
among the population in the same scenario. While it is possible to add
more than one disease in a scenario today, there's some problems:

1. Since birth/death rates are defined inside each disease, they can be
different so the models would disagree on how many people are around at a
given time.
2. The disease death rate for one of the diseases would not be taken into
account in the other disease, so it would tend to overestimate the number
of people availabe.

Solving this is not straightforward at all, here's some thoughts.

Tackling the first problem, the diseases are not truly isolated since they
should work off the same assumptions about birth and death rate, and one
disease needs to know how many people the other disease killed off. So one
idea would be to pull out the birth/death rate into a separate decorator
that models the demographic changes for a scenario. This
"DemographicModel" (I'm open to better name suggestions, it's just a
decorator) simply determine how many individuals are born and die at a
given time (for any species, humans, mosquitos etc.). This is not
necessarily just a fixed birth/death rate. For instance, I can envision
writing a special DemographicModel for mosquitos increasing birth rate
during rainy conditions etc. The DemographicModel will also consult all
disease models in the scenario at each step to see how many people were
killed by diseases to make sure population count is correct.

In this sense, I think a DemographicModel is similar to an infector (or
innoculator). We would be able to create one using the wizard, specify
parameters (birth/death rates) and a region. A top level region would
propagate down to lower level regions just like for innoculators today. If
we can get global demographic data for various regions we can even add a
DemographicModel to the STEM library for a bunch of regions.

As for the second problem, I propose we implement (change) the API's:

A DiseaseModelLabelValue would only have these variables:

1. S, I, R etc. depending on the type of the disease
2. DiseaseDeaths. Each disease model is responsble for determining how
many people die from the disease
3. Incidence How many people become infected in this time period. We still
need to keep this around since it's an important epi parameter

A DemographicLabelValue has these variables:

1. Population. Current total size of the population
2, OriginalPopulation. Needed so we can reset population back to the
original value when a simulation is restarted
3. Births. How many people are born this time period
4. Deaths. How many people die (for any reason) this time period

A DiseaseModel has only two methods it needs to implement:

1. calculateDelta(STEMTime time, long timeDelta,
EList<DiseaseModelLabel>list). When this method returns, the delta value
of each of the passed in labels must have been set so that it can be
applied and is ready to advance the state of the disease from the current
state to the next. At this stage, this method includes lots of things like
it does today like transportation/mixing etc. It does NOT handle births or
deaths (from other than deaths by the disese).
2. applyBirthDeaths(EList<DemographicLabelValue>birthDeaths,
EList<DiseaseModelLabelLabel>list). This is the method that will add the
births and deaths from the passed in list in the first argument to the
delta value for each label in the second argument. Only labels with
overlapping nodes are changed. Important With the algorithm below the
deaths will include disease deaths from this disease model so it will need
to be substracted before applying.

A DemographicModel has two methods to implement:

1. calculateDelta(STEMTime, long timeDelta, EList<DemographicLabel>list).
Similar to the disease model, but the delta value calculated is the birth
and death given the rates associated with the demographic model.
2. applyDeaths(EList<DiseaseModelLabel>list,
EList<DemographicLabelValue>birthDeaths) Additional deaths are applied to
the delta value here from the passed in list of disease model labels. Only
labels with overlapping nodes are changed. This is to adjust for deaths
caused by diseases.

Assuming we have two disease models DM1 and DM2, and one demographic
decorator D, the process of figuring out the change (delta) for DM1, DM2
and D for one time step would be something like:

DM1.calculateDelta(time, timeDelta, DM1List); // DM1List is every label
updated by DM1
DM2.calculateDelta(time, timeDelta, DM2List);
// Now the deltas has been set for every label for both decorators, but
the birth/deaths haven't been applied yet
D.calculateDelta(time, timeDelta, DList); // DList is every label updated
by D
// Now we need to correct for for the disease deaths for each decorator
D.applyDeaths(DM1List, DList);
D.applyDeaths(DM2List, DList);
// And finally adjust the birth/deaths for each disease model
DM1.applyBirthDeaths(DList, DM1List);
DM2.applyBirthDeaths(DList, DM2List);

The error in the time step (for the integration solution) would now be
considered for the delta calculated for each label by DM1, DM2 and D, and
if any exceed the tolerance we'd reduce the step size etc.

Still trying to think if there's a more elegant solution to this, but I
think the method above would work...
Re: Handling multiple diseases in STEM [message #3260 is a reply to message #3224] Tue, 26 May 2009 12:37 Go to previous messageGo to next message
Eclipse UserFriend
Originally posted by: daford.almaden.ibm.com.invalid

Stefan,
I wonder if we couldn't make a kind of "wrapper" disease model that would
consolidate an arbitrary number of disease models. The intent would be to
completely charaterize the entire disease state of a population with respect
to all of the diseases and the history of infection->recovery.

The wrapper would have to "query" the disease models it tracks for their
disease states to set up a (combinatorial) state space. Then each disease
model would need to report the numbers of population members that transition
from one state to the next. The wrapper would then aggregate those
transitions and update the "gloabl" disease state space for the population
with repsect to all of the diseases.

It might require a change in how a disease model gets access to the disease
state values it needs for its computations. At the moment each disease model
defines/uses a single vector and is written assuming a particular defintion
for that vector. If we were to change to have a single global state space
(for each node) we could have the disease models simply provide a token for
the state value they want (eg. "S") and the combined state would return the
correct value with respect to that disease. For instance, if we have two
diseases, population members infected with the first disease would still be
returned as susceptible for the second. This approach would also handle the
birth/death issue properly as well.

Dan



Stefan Edlund wrote:
>
> As a first step towards handling a true multi-serotype scenario in STEM,
> we need to correctly handle two or more isolated diseases circulating
> among the population in the same scenario. While it is possible to add
> more than one disease in a scenario today, there's some problems:
>
> 1. Since birth/death rates are defined inside each disease, they can be
> different so the models would disagree on how many people are around at a
> given time.
> 2. The disease death rate for one of the diseases would not be taken into
> account in the other disease, so it would tend to overestimate the number
> of people availabe.
>
> Solving this is not straightforward at all, here's some thoughts.
>
> Tackling the first problem, the diseases are not truly isolated since they
> should work off the same assumptions about birth and death rate, and one
> disease needs to know how many people the other disease killed off. So one
> idea would be to pull out the birth/death rate into a separate decorator
> that models the demographic changes for a scenario. This
> "DemographicModel" (I'm open to better name suggestions, it's just a
> decorator) simply determine how many individuals are born and die at a
> given time (for any species, humans, mosquitos etc.). This is not
> necessarily just a fixed birth/death rate. For instance, I can envision
> writing a special DemographicModel for mosquitos increasing birth rate
> during rainy conditions etc. The DemographicModel will also consult all
> disease models in the scenario at each step to see how many people were
> killed by diseases to make sure population count is correct.
>
> In this sense, I think a DemographicModel is similar to an infector (or
> innoculator). We would be able to create one using the wizard, specify
> parameters (birth/death rates) and a region. A top level region would
> propagate down to lower level regions just like for innoculators today. If
> we can get global demographic data for various regions we can even add a
> DemographicModel to the STEM library for a bunch of regions.
>
> As for the second problem, I propose we implement (change) the API's:
>
> A DiseaseModelLabelValue would only have these variables:
>
> 1. S, I, R etc. depending on the type of the disease
> 2. DiseaseDeaths. Each disease model is responsble for determining how
> many people die from the disease
> 3. Incidence How many people become infected in this time period. We still
> need to keep this around since it's an important epi parameter
>
> A DemographicLabelValue has these variables:
>
> 1. Population. Current total size of the population
> 2, OriginalPopulation. Needed so we can reset population back to the
> original value when a simulation is restarted
> 3. Births. How many people are born this time period
> 4. Deaths. How many people die (for any reason) this time period
>
> A DiseaseModel has only two methods it needs to implement:
>
> 1. calculateDelta(STEMTime time, long timeDelta,
> EList<DiseaseModelLabel>list). When this method returns, the delta value
> of each of the passed in labels must have been set so that it can be
> applied and is ready to advance the state of the disease from the current
> state to the next. At this stage, this method includes lots of things like
> it does today like transportation/mixing etc. It does NOT handle births or
> deaths (from other than deaths by the disese).
> 2. applyBirthDeaths(EList<DemographicLabelValue>birthDeaths,
> EList<DiseaseModelLabelLabel>list). This is the method that will add the
> births and deaths from the passed in list in the first argument to the
> delta value for each label in the second argument. Only labels with
> overlapping nodes are changed. Important With the algorithm below the
> deaths will include disease deaths from this disease model so it will need
> to be substracted before applying.
>
> A DemographicModel has two methods to implement:
>
> 1. calculateDelta(STEMTime, long timeDelta, EList<DemographicLabel>list).
> Similar to the disease model, but the delta value calculated is the birth
> and death given the rates associated with the demographic model.
> 2. applyDeaths(EList<DiseaseModelLabel>list,
> EList<DemographicLabelValue>birthDeaths) Additional deaths are applied to
> the delta value here from the passed in list of disease model labels. Only
> labels with overlapping nodes are changed. This is to adjust for deaths
> caused by diseases.
>
> Assuming we have two disease models DM1 and DM2, and one demographic
> decorator D, the process of figuring out the change (delta) for DM1, DM2
> and D for one time step would be something like:
>
> DM1.calculateDelta(time, timeDelta, DM1List); // DM1List is every label
> updated by DM1
> DM2.calculateDelta(time, timeDelta, DM2List);
> // Now the deltas has been set for every label for both decorators, but
> the birth/deaths haven't been applied yet
> D.calculateDelta(time, timeDelta, DList); // DList is every label updated
> by D
> // Now we need to correct for for the disease deaths for each decorator
> D.applyDeaths(DM1List, DList);
> D.applyDeaths(DM2List, DList);
> // And finally adjust the birth/deaths for each disease model
> DM1.applyBirthDeaths(DList, DM1List);
> DM2.applyBirthDeaths(DList, DM2List);
>
> The error in the time step (for the integration solution) would now be
> considered for the delta calculated for each label by DM1, DM2 and D, and
> if any exceed the tolerance we'd reduce the step size etc.
>
> Still trying to think if there's a more elegant solution to this, but I
> think the method above would work...
>
Re: Handling multiple diseases in STEM [message #3294 is a reply to message #3260] Tue, 26 May 2009 18:11 Go to previous messageGo to next message
Stefan Edlund is currently offline Stefan EdlundFriend
Messages: 23
Registered: July 2009
Junior Member
Right, to handle a true multi-serotype model like dengue fever where the
disease death rate varies depending on your infection history we would
need to do something like that. Abstracting the compartment states would
require a huge overhaul of STEM though, which is something I wasn't really
ready to do at this stage.

I like the wrapper model idea though, I think the demographic model could
be turned into a type of wrapper that contains everything affecting a
population like births, deaths, diseases.

What I'm still trying to wrap my head around is exactly what is it the
differential equation solver is "solving". The pseudo code below was an
example of what needs to be done to advance the solution one time step,
where the time step varies depending on how much activity is going on. So
just as an example, if a rain storm causes mosquito population to explode,
the solver will step very slowly (perhaps calculating a solution every
simulated minute or less!) through that part of the simulation. We need an
abstraction of all dynamic activity that pretty much affects anything in
STEM.
Re: Handling multiple diseases in STEM [message #3326 is a reply to message #3294] Thu, 28 May 2009 20:59 Go to previous message
James Kaufman is currently offline James KaufmanFriend
Messages: 167
Registered: July 2009
Senior Member
Stefan,
I agree completely with your analysis. Instead of "Demographic Model" we
might call it a "Population Model" which has, as some of it's properties,
"Demographic Properties". I agree with your list of properties but this
list could be longer.
For example, for humans we might want
1. Population. Current total size of the population
2A, OriginalPopulation. Needed so we can reset population back to the
original value when a simulation is restarted.
2B OriginalPopulation data Valid
3. Births. How many people are born this time period
4. Deaths. How many people die (for any reason) this time period

For humans, we might also want as part of a larger population model object
5. road transportation rate
6. air transportation
7. common border mixing
8. etc.

For Mosquitoes we might wan
5. some rainfall function by date
6. etc.

We might want to add properties files with Birth rates/Death rates for
Humans. I think the idea of infector/innoculator wizard is a good one too.
In fact maybe we can generalize this so that, in principle, any properties
that derive from our property files could be modified with such a wizard
so users can override the default data (for example, in event of a war a
border might be closed, in the event of an economic boom - air travel
might increase). This could be a very useful tool and would let us modify
the decorators in experiments as well.

Regarding a wrapper, we would need it for dengue fever but it might add
unnecessary complexity for simpler diseases that we want to keep ez to
use. I agree it's best to take this on incrementally and start by creating
this more general and powerful Population or Demographic Model object.
Re: Handling multiple diseases in STEM [message #561056 is a reply to message #3224] Tue, 26 May 2009 12:37 Go to previous message
Daniel Ford is currently offline Daniel FordFriend
Messages: 138
Registered: July 2009
Location: New York
Senior Member
Stefan,
I wonder if we couldn't make a kind of "wrapper" disease model that would
consolidate an arbitrary number of disease models. The intent would be to
completely charaterize the entire disease state of a population with respect
to all of the diseases and the history of infection->recovery.

The wrapper would have to "query" the disease models it tracks for their
disease states to set up a (combinatorial) state space. Then each disease
model would need to report the numbers of population members that transition
from one state to the next. The wrapper would then aggregate those
transitions and update the "gloabl" disease state space for the population
with repsect to all of the diseases.

It might require a change in how a disease model gets access to the disease
state values it needs for its computations. At the moment each disease model
defines/uses a single vector and is written assuming a particular defintion
for that vector. If we were to change to have a single global state space
(for each node) we could have the disease models simply provide a token for
the state value they want (eg. "S") and the combined state would return the
correct value with respect to that disease. For instance, if we have two
diseases, population members infected with the first disease would still be
returned as susceptible for the second. This approach would also handle the
birth/death issue properly as well.

Dan



Stefan Edlund wrote:
>
> As a first step towards handling a true multi-serotype scenario in STEM,
> we need to correctly handle two or more isolated diseases circulating
> among the population in the same scenario. While it is possible to add
> more than one disease in a scenario today, there's some problems:
>
> 1. Since birth/death rates are defined inside each disease, they can be
> different so the models would disagree on how many people are around at a
> given time.
> 2. The disease death rate for one of the diseases would not be taken into
> account in the other disease, so it would tend to overestimate the number
> of people availabe.
>
> Solving this is not straightforward at all, here's some thoughts.
>
> Tackling the first problem, the diseases are not truly isolated since they
> should work off the same assumptions about birth and death rate, and one
> disease needs to know how many people the other disease killed off. So one
> idea would be to pull out the birth/death rate into a separate decorator
> that models the demographic changes for a scenario. This
> "DemographicModel" (I'm open to better name suggestions, it's just a
> decorator) simply determine how many individuals are born and die at a
> given time (for any species, humans, mosquitos etc.). This is not
> necessarily just a fixed birth/death rate. For instance, I can envision
> writing a special DemographicModel for mosquitos increasing birth rate
> during rainy conditions etc. The DemographicModel will also consult all
> disease models in the scenario at each step to see how many people were
> killed by diseases to make sure population count is correct.
>
> In this sense, I think a DemographicModel is similar to an infector (or
> innoculator). We would be able to create one using the wizard, specify
> parameters (birth/death rates) and a region. A top level region would
> propagate down to lower level regions just like for innoculators today. If
> we can get global demographic data for various regions we can even add a
> DemographicModel to the STEM library for a bunch of regions.
>
> As for the second problem, I propose we implement (change) the API's:
>
> A DiseaseModelLabelValue would only have these variables:
>
> 1. S, I, R etc. depending on the type of the disease
> 2. DiseaseDeaths. Each disease model is responsble for determining how
> many people die from the disease
> 3. Incidence How many people become infected in this time period. We still
> need to keep this around since it's an important epi parameter
>
> A DemographicLabelValue has these variables:
>
> 1. Population. Current total size of the population
> 2, OriginalPopulation. Needed so we can reset population back to the
> original value when a simulation is restarted
> 3. Births. How many people are born this time period
> 4. Deaths. How many people die (for any reason) this time period
>
> A DiseaseModel has only two methods it needs to implement:
>
> 1. calculateDelta(STEMTime time, long timeDelta,
> EList<DiseaseModelLabel>list). When this method returns, the delta value
> of each of the passed in labels must have been set so that it can be
> applied and is ready to advance the state of the disease from the current
> state to the next. At this stage, this method includes lots of things like
> it does today like transportation/mixing etc. It does NOT handle births or
> deaths (from other than deaths by the disese).
> 2. applyBirthDeaths(EList<DemographicLabelValue>birthDeaths,
> EList<DiseaseModelLabelLabel>list). This is the method that will add the
> births and deaths from the passed in list in the first argument to the
> delta value for each label in the second argument. Only labels with
> overlapping nodes are changed. Important With the algorithm below the
> deaths will include disease deaths from this disease model so it will need
> to be substracted before applying.
>
> A DemographicModel has two methods to implement:
>
> 1. calculateDelta(STEMTime, long timeDelta, EList<DemographicLabel>list).
> Similar to the disease model, but the delta value calculated is the birth
> and death given the rates associated with the demographic model.
> 2. applyDeaths(EList<DiseaseModelLabel>list,
> EList<DemographicLabelValue>birthDeaths) Additional deaths are applied to
> the delta value here from the passed in list of disease model labels. Only
> labels with overlapping nodes are changed. This is to adjust for deaths
> caused by diseases.
>
> Assuming we have two disease models DM1 and DM2, and one demographic
> decorator D, the process of figuring out the change (delta) for DM1, DM2
> and D for one time step would be something like:
>
> DM1.calculateDelta(time, timeDelta, DM1List); // DM1List is every label
> updated by DM1
> DM2.calculateDelta(time, timeDelta, DM2List);
> // Now the deltas has been set for every label for both decorators, but
> the birth/deaths haven't been applied yet
> D.calculateDelta(time, timeDelta, DList); // DList is every label updated
> by D
> // Now we need to correct for for the disease deaths for each decorator
> D.applyDeaths(DM1List, DList);
> D.applyDeaths(DM2List, DList);
> // And finally adjust the birth/deaths for each disease model
> DM1.applyBirthDeaths(DList, DM1List);
> DM2.applyBirthDeaths(DList, DM2List);
>
> The error in the time step (for the integration solution) would now be
> considered for the delta calculated for each label by DM1, DM2 and D, and
> if any exceed the tolerance we'd reduce the step size etc.
>
> Still trying to think if there's a more elegant solution to this, but I
> think the method above would work...
>
Re: Handling multiple diseases in STEM [message #561071 is a reply to message #3260] Tue, 26 May 2009 18:11 Go to previous message
Stefan Edlund is currently offline Stefan EdlundFriend
Messages: 23
Registered: July 2009
Junior Member
Right, to handle a true multi-serotype model like dengue fever where the
disease death rate varies depending on your infection history we would
need to do something like that. Abstracting the compartment states would
require a huge overhaul of STEM though, which is something I wasn't really
ready to do at this stage.

I like the wrapper model idea though, I think the demographic model could
be turned into a type of wrapper that contains everything affecting a
population like births, deaths, diseases.

What I'm still trying to wrap my head around is exactly what is it the
differential equation solver is "solving". The pseudo code below was an
example of what needs to be done to advance the solution one time step,
where the time step varies depending on how much activity is going on. So
just as an example, if a rain storm causes mosquito population to explode,
the solver will step very slowly (perhaps calculating a solution every
simulated minute or less!) through that part of the simulation. We need an
abstraction of all dynamic activity that pretty much affects anything in
STEM.
Re: Handling multiple diseases in STEM [message #561088 is a reply to message #3294] Thu, 28 May 2009 20:59 Go to previous message
James Kaufman is currently offline James KaufmanFriend
Messages: 167
Registered: July 2009
Senior Member
Stefan,
I agree completely with your analysis. Instead of "Demographic Model" we
might call it a "Population Model" which has, as some of it's properties,
"Demographic Properties". I agree with your list of properties but this
list could be longer.
For example, for humans we might want
1. Population. Current total size of the population
2A, OriginalPopulation. Needed so we can reset population back to the
original value when a simulation is restarted.
2B OriginalPopulation data Valid
3. Births. How many people are born this time period
4. Deaths. How many people die (for any reason) this time period

For humans, we might also want as part of a larger population model object
5. road transportation rate
6. air transportation
7. common border mixing
8. etc.

For Mosquitoes we might wan
5. some rainfall function by date
6. etc.

We might want to add properties files with Birth rates/Death rates for
Humans. I think the idea of infector/innoculator wizard is a good one too.
In fact maybe we can generalize this so that, in principle, any properties
that derive from our property files could be modified with such a wizard
so users can override the default data (for example, in event of a war a
border might be closed, in the event of an economic boom - air travel
might increase). This could be a very useful tool and would let us modify
the decorators in experiments as well.

Regarding a wrapper, we would need it for dengue fever but it might add
unnecessary complexity for simpler diseases that we want to keep ez to
use. I agree it's best to take this on incrementally and start by creating
this more general and powerful Population or Demographic Model object.
Previous Topic:Handling multiple diseases in STEM
Next Topic:STEM Migration Plans and Release Schedule
Goto Forum:
  


Current Time: Thu Dec 18 09:49:08 GMT 2014

Powered by FUDForum. Page generated in 0.02250 seconds
.:: Contact :: Home ::.

Powered by: FUDforum 3.0.2.
Copyright ©2001-2010 FUDforum Bulletin Board Software