Model Cards - AIforGood Simulator

1. Agent-based modelling (ABM)
Model Description: Overview, Input, Output, Architecture

Model Overview
Agent-based modeling for the outbreak of cholera has been done before in refugee camps and this is a simplified version of the agent-based modeling (constrained by the time of the hackathon) with 500 people on a realistic map of the Moria Camp in Greece. The data about the population breakdown and their daily activities are obtained through UNHCR and NGO workers who work on the ground in the camp. The model is implemented in Netlogo and hosted on the website
ABM: Input parameters used for all simulations
Demographic inputs:
population_size 500 people (constrained by the computation power) percent_under18 62%
percent_18-49 35%
percent_50+ 3%
Health inputs:
transmission_probability 15% (assumed)
incubation_time 5 days
infection_prob_under18 25% (assumed)
infection_prob_18-49 45% (assumed)
infection_prob_50+ 60% (assumed)
infection_time 5 days
infection_probability_recovery 50% (assumed)
ABM: Model logic and assumptions
Basic daily routine of refugees:
wake up between 7am and 10am
go to nearest wash facility (W in blue) to shower, go to the toilet, etc.
return home
go to food distribution center (F in orange) to get food between 12-2pm
return home
go to nearest office (O in grey) to get support, consultation between 3-6pm
return home
Exposure to COVID-19:
one refugee at random is exposed to the virus
virus spreads only when people are in the same building with a certain transmission probability set by the slider - they then get exposed and turn pink
after the incubation time - set by the slider - they either recover or they get infected and show serious symptions - infection rate is based on their age group and the sliders in the model
if refugees are infected with the virus and experience serious symptoms, they turn red and are sent to one of the medical service areas ("+" in red)
refugees infected with the virus recover after the infection period and with a probability of recovery - both set by the sliders
if refugees recover, they turn green and resume their daily routine
ABM Intervention assumptions
PPE:
everyone gets PPE
PPE reduces transmission by percentage set in the slider in the model
Lockdown:
refugees allowed to wake up and use wash facilities
between 12pm and 2pm only one refugee is allowed to go get food for the housing unit for 50% of the housing units
between 3pm and 6pm only one refugee is allowed to get food for the housing unit for the remaining 50% of the housing units
offices are closed and no consultation hours are allowed
Testing of Infected:
refugees go about their normal daily routine
during the consultation time between 3pm and 6pm a certain percentage of refugees - set by the slider - go to the medical center instead to get tested
if tested positive for COVID-19 (if they are exposed) then they get sent to an isolation area set up outside the camp
refugees who are not tested still go about their daily routine
refugees who are infected get sent to the hospital regardless of if they are isolated or not
refugees who were isolated and recover come back to the camp to resume their normal activities
Isolation of Vulnerable:
refugees go about their normal daily routine
during the consultation time between 3pm and 6pm a certain percentage of refugees - set by the slider - go to the medical center instead to get checked
if they are assumed to be vulnerable to the infection then they are quarantined
refugees who are not tested still go about their daily routine
refugees who are infected get sent to the hospital regardless of if they are isolated or not
refugees who were isolated and recover come back to the camp to resume their normal activities
ABM: Model performance
(pending some plots and insights)
ABM: Pros and cons
Pros:
Individuals within the population can be represented. More flexibility in representing individual heterogeneity
More visual representation on how the virus could spread given geographical territory.
Mimic the real-life interactions within the camp by assigning different roles to agents according to their routine as activities within camps usually follow a timeline
Cons:
Challenging to simulate real-life contact. The activities followed by the agents are oversimplified therefore might deviate from how people interact in real-life.
Heavy computation by the model as the goal of modelling 20,000 agents cannot be done on the local laptop as the speed becomes extremely slow.
Cascading of probabilities induced by the assumptions might skew the effectiveness of the intervention mechanisms.
2. Compartment Modelling
Model Description: Overview, Input, Output, Architecture

Model Overview
Treating a population as a whole enables the full expressivity of the SEIR model where all parameters within the equation can be captured. We repurpose this model in python’s streamlite module to build our own mobile dashboard. In particular, we fix the R0, Tinc, Tinf to be the same as the parameters estimated from the Princess Diamond cruise ship (although the population density in Moria 200,000/km2 is still much higher than that of Princess Diamond cruise ship 24,400/km2). The parameters for the Princess Diamond cruise ship is taken from this study.
Compartment Model: Description
A simple SEIR (Susceptible -> Exposed ->Infected - > Removed) model that is used in research to simulate epidemics. In its core, it uses four differential equations to follow the change at each stage of the disease’s progression.
Compartment Model: Input
Population size
R0: Reproduction number
R1: Reproduction number after an intervention
T_inf: Infectious period
T_inc: Incubation period
Compartment Model: Output
A dataset of 5 variables:
S: number of Susceptible population at each day of the simulation
E: number of people exposed to the virus
I: number of people that have been infected
R: number of people that have been removed(immune or isolation)
D: number of people that died from the virus
Compartment Model: Logic and assumptions
a homogeneous population with uniform mixing (thus a constant R-value)
Compartment Model: Intervention assumptions
The model assumes the camp has three range of intervention parameters at disposal (can be a combination of mitigation strategies. The effectiveness range from very effective 94% to fairly effective 80%. We found that any mitigation strategy that reduce R0 by less than 80% (R0>2.6) to be ineffective as the exponential nature of the growth would just mean the most of the population will be infected with a short time delay compared to the base model.
Compartment Model: Model performance
With an R0 so high as 14 any measure taken needs to be drastic.