More than 1 billion people live in informal settlements worldwide, where precarious living conditions pose unique challenges to managing a COVID-19 outbreak. Taking Northwest Syria as a case study, we simulated an outbreak in high-density informal Internally Displaced Persons (IDP) camps using a stochastic Susceptible-Exposed-Infectious-Recovered model. Expanding on previous studies, taking social conditions and population health/structure into account, we modelled several interventions feasible in these settings: moderate self-distancing, self-isolation of symptomatic cases and protection of the most vulnerable in ‘safety zones’. We considered complementary measures to these interventions that can be implemented autonomously by these communities, such as buffer zones, health checks and carers for isolated individuals, quantifying their impact on the micro-dynamics of disease transmission. All interventions significantly reduce outbreak probability and some of them reduce mortality when an outbreak does occur. Self-distancing reduces mortality by up to 35% if contacts are reduced by 50%. A reduction in mortality by up to 18% can be achieved by providing one self-isolation tent per eight people. Protecting the most vulnerable in a safety zone reduces the outbreak probability in the vulnerable population and has synergistic effects with the other interventions. Our model predicts that a combination of all simulated interventions may reduce mortality by more than 90% and delay an outbreak’s peak by almost 2 months. Our results highlight the potential for non-medical interventions to mitigate the effects of the pandemic. Similar measures may be applicable to controlling COVID-19 in other informal settlements, particularly IDP camps in conflict regions, around the world.

Since the onset of the COVID-19 pandemic, many studies have provided evidence for the effectiveness of strategies such as social distancing, testing, contact tracing, case isolation, use of personal protective equipment/face masks and improved hygiene to reduce the spread of the disease. These studies underlie the recommendations of WHO, but their implementation is contingent on local conditions and resources.

Mathematical modelling is the basis of many epidemiological studies and has helped inform policymakers considering COVID-19 responses around the world. Nevertheless, only a limited number of studies have applied these models to informal settlements.

We developed a mathematical model to study the dynamics of COVID-19 in Syrian internally displaced persons (IDP) camps, elaborating on previous efforts done in similar settings by explicitly parameterising the camps’ demographics, living conditions and microdynamics of interpersonal contacts in our modelling.

We designed interventions such as self-distancing, self-isolation and the creation of safety zones to protect the most vulnerable members of the population, through conversations with camp managers. We ensured that our proposed interventions would be feasible and have community buy-in.

Our results show how low-cost, feasible, community-led non-medical interventions can significantly mitigate the impact of COVID-19 in Northwest Syrian IDP camps.

Our model represents a step forward in the much-needed search for epidemiological models that are sufficiently flexible to consider specific social contexts. The model can also help inform similar interventions in refugee camps in conflict-torn regions, and potentially be adapted to other informal settlements and vulnerable communities around the world.

The spread of airborne infectious diseases with pandemic potential in regions immersed in protracted armed conflicts, with large displaced populations, is an important challenge.

This study focuses on the spread of COVID-19 in the Northwest region of Syria (NWS): a relatively small geographical area with 4.2 million people, of which 1.15 million (27.4%) are IDPs living in camps,

To investigate feasible COVID-19 prevention interventions in the camps, we considered a Susceptible-Exposed-Infectious-Recovered model similar to the one presented by Gatto

Building on the approach used to model the impact of these interventions in African cities, our model includes a parameterisation of the number of contacts each individual has per day.

We consider a model simulating a viral outbreak in a single camp over a 12-month period inspired by those proposed by Gatto

Diagram of the model. The model considers the following compartments: susceptible (S) exposed (E) infectious-presymptomatic (P) infectious asymptomatic (A) infectious symptomatic (I) infectious-requiring hospitalisation (H) recovered (R) and dead (D).

The susceptible population

Fixed parameters

Parameter | Description | Value | Distribution | Reference |

Incubation period (days) | 5.2 (95% CI 4.1 to 7.0) | Lognormal | ||

Presymptomatic infectious period (days) | 2.3 (95% CI 0.8 to 3.8) | Gaussian | ||

Latent period (days) | Derived | |||

Symptomatic infectious period (days) | 7 | --- | ||

Asymptomatic infectious period (days) | 7 | --- | ||

Time from symptom onset to requiring hospitalisation (days) | 7 (IQR: 4–8) | Gamma | ||

Time from symptoms onset to death (critical cases, days) | 10 (IQR: 6–12) | Gamma | ||

Time from requiring hospitalisation to recovery/death (days) | 10 (IQR: 7–14) | Gamma | ||

f | Probability an infectious individual is symptomatic | 0.84 (95% CI 0.8 to 0.88) | Gamma | |

σ | Indicator of whether hospitalised recover or die | --- | Assumed |

See

While we introduced the model as a classical system of ordinary differential equations (Equations 1–8) we considered a stochastic implementation,

The model splits the population into classes (indexed i) to account for heterogeneity with respect to clinical risk and behaviour. Working with population classes allows us to encode behavioural assumptions in the model and strike an appropriate balance between generality, computational tractability and the requisite specificity to realistically evaluate our proposed interventions.

Under some interventions, the demographic classes may be subdivided further into subclasses according to behaviour (‘behaviour classes’). Consequently, different interventions may require models with different numbers of classes. We refer to both demographic and behaviour classes generically as ‘classes’ (see section Interventions for the modelling of behaviour classes).

We parameterised the model with data from IDPs in NWS.

Demographic class-specific parameters

Parameter | Description | Demographic class | |||||

Age 1 (0–12) | Age 2 | Age 2 | Age 3 (>50) no comorbidities | Age 3 (>50) comorbidities | References | ||

Fraction in class | – | 0.407 | 0.471 | 0.0626 | 0.022 | 0.0373 | |

Mean contacts per day | 25 | 15 | 15 | 10 | 10 | From camp managers |

Estimated proportions of individuals in the population and mean number of contacts per individual per day for each demographic class. See

Although individuals in IDP camps share tents with other co-occupants, whom they may be more likely to infect than occupants of different tents, we ignore spatial structure in our model and assume a well-mixed population. This is justified because individuals from different tents share common spaces (e.g. latrines) and have frequent interactions with each other, especially among children. Consequently, our following derivation of the transmissivity parameter itself, τ, is not spatially explicit.

The rate at which susceptible individuals become exposed is

where

The probability of infection if there is a contact between a susceptible and an infected person is

Transmissibility parameters

Parameter | Description | Value | Distribution | Reference |

τ | Maximum transmissibility | 0.14 (95% CI 0.05 to 0.40) | Lognormal | Derived |

Presymptomatic transmissibility of individuals becoming symptomatic relative toτ | Reference stage (=1) | – | ||

Mean presymptomatic transmissibility relative toτ | 0.93 (95% CI 0.88 to 0.99) | Empirical | Derived | |

Asymptomatic transmissibility relative toτ | 0.14 (95% CI 0.05 to 0.40) | Lognormal | Derived | |

Clinical symptomatic transmissibility relative toτ | 0.24 (95% CI 0.11 to 0.60) | Lognormal | Derived | |

Hospitalised transmissibility relative toτ | 0.11 (95% CI 0.05 to 0.29) | Lognormal | Derived |

See

The τ parameter was estimated by randomly generating a value for the basic reproduction number,

In NWS, there are four active and two planed COVID-19 referral hospitals, with a current capacity of 66 ventilators, 74 ICU beds and 355 ward beds for 4.2 million people.

The fractions of symptomatic cases that are severe

Proportion of symptomatic cases that become severe and critical

Parameter | Description | Demographic class | ||||

Age 1 (0–12) | Age 2 | Age 2 (13–50) comorbidities | Age 3 | Age 3 (>50) comorbidities | ||

Fraction of symptomatic cases severe | 0.064 (0.064) | 0.067 (0.066) | 0.199 (0.191) | 0.183 (0.178) | 0.445 (0.406) | |

Fraction of symptomatic cases critical | 0.0065 (0.009) | 0.020 (0.028) | 0.094 (0.129) | 0.063 (0.088) | 0.222 (0.289) |

Since the rates at which these cases become severe, critical or recover are different, we introduced three parameters

Since the rates at which clinical symptomatic individuals (

The interventions we consider are modelled by modifying the rate at which individuals become exposed (the term

All interventions, self-distancing (see

Diagram of interventions.

Parameterisation of the interventions

Intervention | Details | |||||

Self-distancing | 1 | 1 | 1 | 1 | Constant | |

Self-isolation | 1 | 1 | 0.2 | |||

1 | 1 | 1 | ||||

0 | 0 | 0 | ||||

1 | 1 | 1 | ||||

Safety zone | {0,1} see (*) | {0,1} see (*) | 0.2 | |||

1 | 1 | 1 | ||||

Evacuation | 1 | 1 | 1 | 0 | 1 | — |

Values of

*The parameters are set to 0 if health checks to access to the buffer zone are implemented in the intervention and set to 1 otherwise.

The first non-medical intervention that we modelled is a reduction in the mean number of contacts per individual per day for the whole camp population (see

Self-isolation is a challenge in informal settlements, where households consist of a single (often small) space, water is collected at designated locations, sanitation facilities are communal and food supplies are scarce. We considered the possibility of those showing symptoms (i.e. in compartment I) self-isolating in individual tents in dedicated parts of the camps. We excluded individuals with severe symptoms (H compartment) from the intervention since they require additional care not compatible with self-isolation. Instead, we considered the possibility of evacuating these individuals from the camp as an additional intervention (see below). We simulated self-isolation with various numbers of isolation tents per camp, ranging from 10 to 2000 for a camp of 2000 people (see

In addition, interactions between carers and isolated individuals were restricted to buffer zones, which we envisioned as open spaces, with guidelines in place to limit occupancy to four individuals wearing masks with at least 2 m of distance between them, where we assume transmissivity is reduced by 80%

In this intervention, the camp is divided in two areas: a safety zone, in which more vulnerable people live (hereby referred to as a ‘green’ zone following previous studies),

Interactions between the two zones are limited to a buffer zone, reducing transmissivity (i.e.

The last intervention we simulated is the evacuation of severe cases (individuals in the hospitalisation compartment). Since they require more intensive care that cannot be delivered while adhering to the guidelines of a buffer zone, severe cases were assumed to be fully infectious and not able to self-isolate. Once severe cases are evacuated, their infectivity is reduced to zero (

The specific values of the parameters shown in

The next-generation matrix is also computed at each integration step from the parameters drawn and τ estimated. In the code provided, it is possible to fix the seed to exactly reproduce the results presented. Our simulations start with a completely susceptible population where one person in the younger adult population is exposed to the virus (who is also in the orange zone if the safety zone intervention is in place). We verified that a steady state was always reached before the end of each simulation. We did not consider migration, births, nor deaths due to other causes, since they are small enough in magnitude to not significantly impact the course of an outbreak, provided additional conflict does not erupt.

For each implementation of the interventions, we ran 2500 simulations and compared results between them. The main variables considered are the fraction of simulations in which at least one death is observed (a proxy for the probability of an outbreak), the fraction of the population that dies and the time until the symptomatic population peaks, as well as the infection fatality rate (IFR), and the fractions of the population that have recovered and remain susceptible at steady state. For consistency, we only considered simulations in which there was an outbreak when comparing the outcome of a variable between interventions. We used the Shapiro-Wilk test

In the absence of interventions, the mean IFR is

Our results show that self-distancing has a notable effect on reducing the probability of an outbreak which decreases roughly linearly as the number of contacts is reduced (see

Effect of interventions on outbreak probability, fatalities and time until symptomatic cases peak. (A) Self-distancing, probability of an outbreak. (B) Self-distancing, fraction of the population dying. (C) Self-distancing, time until peak symptomatic cases. (D) Self-isolation, probability of an outbreak. (E) Self-isolation, fraction of the population dying. (F) Self-isolation, time until peak symptomatic cases. (G) Safety zone, probability of an outbreak. (H) Safety zone, fraction of the population dying. (I) Safety zone, time until peak symptomatic cases. Triangles indicate the means and boxes IQRs. Note that in figures of the safety zone intervention (panels G–I), the mean of an outcome for the whole population is not the weighted mean of the exposed and safety zones, since outcomes are computed considering simulations in which at least one death was observed in the population class inhabiting the zone, that is, the number of simulations considered to compute each mean is different. In the safety zone figures (panels G–I) health checks are in place, in

With only 10 tents for a camp of 2000 people (i.e. 1 tent for every 200 people), self-isolation yields a marked decrease in the probability of observing an outbreak (~26%) (see

We observe no significant effects when severe cases requiring hospitalisation are evacuated (see

In this section, we consider the scenario in which all older adults, younger adults with comorbidities and their family members up to 20% of the camp population live in the green zone, unless otherwise specified. Creating a green zone improves the effect of the previous interventions overall, but with sometimes opposite outcomes in the exposed and protected populations. For example, the probability of an outbreak decreases for the protected population, by around 11%, if only two contacts are allowed per week in the buffer zone (see

Considering different scenarios for allocating people to the green zone, the lowest probability of an outbreak is achieved when only older adults or at most older adults and younger adults with comorbidities move there, with probabilities below 0.4 and 0.65, respectively (see

The effects of the interventions observed when we examine them individually build on each other when multiple interventions are implemented in tandem (see

Combinations of interventions. Probability of an outbreak (top), fraction of the population dying (middle) and time until peak symptomatic cases (bottom) for different combination of interventions. For combinations of interventions including a safety zone, we distinguish between the population living in the green zone, in the orange zone and the whole population. Evac = evacuation of severely symptomatic; lock = lockdown of the buffer zone; safety = safety zone; self = self-distancing; tents = number of available self-isolation tents.

In this study, we propose a number of interventions of immediate applicability to informal settlements. We focused on IDP settlements in NW Syria, taking into account the interventions' feasibility, cultural acceptance and their need for low cost. When confronted with different possible scenarios, we generally considered the worst-cases, highlighting the interventions that are most effective in the direst conditions, but possibly resulting in an overestimate of mortality. This potential overestimation does not change the qualitative picture of the results, which is built on the relative comparison of outcomes between different combinations of interventions, or the lack thereof.

Our results align with previous simulation studies of potential COVID-19 interventions in similarly densely populated, low-resource settings where informal settlements are present, such as urban areas of sub-Saharan Africa. In these settings, social distancing is demonstrated to be an effective intervention, and even small changes are estimated to have large effects on outbreaks,

Self-distancing proves to be an effective measure in our models as well; reducing contacts by 50% has the greatest effect across most outcomes of interest in any of the interventions we examined. However, the difficulty of achieving a reduction of this magnitude cannot be overlooked, especially considering the large proportion of the population composed of children, a group with an already high contact rate that may prove difficult to control.

We also propose self-isolation using individual tents which can be located in a dedicated zone or next to the tents of relatives, where contact with non-isolated individuals is mediated by a buffer zone. This intervention is effective in preventing an outbreak in the camp with even a small number of isolation tents, as low as 5–10 tents per 1000 camp residents. But it requires at least 125 tents per 1000 camp residents to substantially reduce mortality. After conversations with camp managers, we found that this intervention is more likely to be accepted in NW Syria than evacuation to community-based isolation centres. Community-based isolation not only poses cultural challenges; the capacity required to implement it has hardly been met,

Setting up a safety zone has two positive effects that most stand out: a reduction in the probability of an outbreak in the vulnerable population, and an increase in the time until the number of symptomatic cases peaks. Much of the success or failure of the safety zone intervention hinges on the functioning of the buffer zone. The number of interzone contacts per week, the implementation of health checks, and potential lockdowns all have notable effects. Also important is the portion of the population that is protected; protecting only the vulnerable may have the most beneficial effects, but it is precisely these vulnerable individuals, older adults, and people with comorbidities, who may most need family members to care for them. While safety zone scenarios with more family members accompanying their vulnerable relatives may confer greater epidemiological risk, they may also engender greater well-being and social cohesion.

Despite these benefits, we do not observe a clear decrease in IFR with this intervention, although it is possible that our model may overestimate mortality from an outbreak in the green zone in the few instances when there is one. Since it is unlikely that the camps have the economic means to increase the number of tents when implementing this intervention, we assumed that individuals do not reduce their contacts when moved to the green zone since household sizes will not decrease, which implies an increase in the number of contacts between vulnerable individuals. Despite this increase in contacts, we do not observe an increase in mortality in the vulnerable population when the safety zone is implemented. These results address concerns raised around this type of intervention from previous experiences with large numbers of fatalities registered in nursing-homes in developed countries.

An instrumental consideration for our models is the fraction of the population recovered from COVID-19 after a steady state is reached. Although the duration for which SARS-CoV-2 infection confers immunity is uncertain, the proportion of the population recovered after an outbreak should play a role in its protection against future ones. For all interventions except self-distancing >30%, we observed that the fraction of the population recovered meets or exceeds 50%.

Other important considerations for interpreting our results are the modelling assumptions we made. One important parameter in our model is the relative transmissibility of the different infectious stages, whose specific values still have large margins of variability.

It is also important to acknowledge the benefits and limitations of different possible computational implementations. For instance, there are interventions that do not have a natural implementation within our framework, such as those requiring interventions targeting very specific interactions between individuals (as opposed to large groups of individuals), or the reproduction of empirically observed residence times. An example might be the isolation of an individual and his/her family, as proposed by Gilman

A key limitation of our approach is that it simulates an outbreak started by one infectious individual in a single camp with a closed population. We acknowledge that this approach does not fully capture the complexities of the NWS region, where IDPs live interspersed throughout the region in several hundred camps. The dynamics of an outbreak in the region are undoubtedly influenced by inter-community contacts, and the dynamics of an outbreak in a single camp by these region-wide dynamics, as it has been demonstrated in other countries.

Other unaccounted for social and cultural dynamics will undeniably complicate the feasibility of our proposed interventions. One example we have not addressed here is the unlikeliness of children under 13 self-isolating. Although the number of challenges to implementing our proposed interventions are potentially endless, the community-based nature of our approach may make it more robust to such challenges than approaches relying on healthcare system, which often depend on complex political decisions and may take years to build the required capacity for an effective response. If the dynamics of the virus are well understood by local communities and at least some of the interventions we propose are implemented, the impacts of COVID-19 can be mitigated even in an environment as challenging as NW Syria.

Given a rapidly changing environment and slow responses of local and international authorities in conflict regions where political control is disputed, with international authorities often leaving these communities aside in their priorities,

We thank Crowdfight (www.crowdfight.org) for organizing this collaboration upon request from CS. We thank Judith Boumann for valuable contributions. We thank Peter Ashcroft, Juan Poyatos, Noreen Goldman, Burcu Tepekule and members of Sebastian Bonhoeffer's and Bryan Grenfell's groups for useful discussions. We thank two anonymous reviewers who provided thoughtful and insightful comments that greatly improved our article.

Seye Abimbola

@J_D_Klein, @ecam85

JV and EC-F contributed equally.

All authors contributed to the conceptualisation. Design of the methodology: APG, EC-F, JV, JDK and CS. Formal analysis: APG, EC-F and JDK. Code development APG, EC-F and JDK. Conducted research: APG, EC-F, JV and JDK. Validate results: APG, EC-F, JDK, JV and CS. Contributed resources: APG, CS, EC-F, JDK and JV. Data curation: APG, JDK and EC-F. Visualisation: APG, EC-F, JDK and JV. Writing (original draft) APG, EC-F, and JDK. All authors contributed to the final version of the manuscript, and APG supervised the research.

ECF's research is supported by Wellcome Trust grant 204833/Z/16/Z. APG research is supported by the Simons Collaboration: Principles of Microbial Ecosystems (PriME), award number 542381.

These organisations had no role in study design, data collection, data analysis, data interpretation, or writing of the article.

APG is a Board Member of crowdfight COVID19, an initiative from the scientific community to put all available resources at service of the fight against COVID-19. CS (coauthor) is a Board Member of the Pax Syriana Foundation, a non-profit organisation set up for social and philanthropic purposes including promoting and providing support and assistance to civilian aid projects in the fields of education, health,emergency assistance, psychological assistance and humanitarian aid for people affected by wars or humanitarian crises.

Not commissioned; externally peer reviewed.

This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.

Data are available in a public, open access repository. All the code and results are freely available at the url

Not required.