SPATIOTEMPORAL EARLY WARNING SYSTEM FOR COVID-19 PANDEMIC

Wuhan, China reported the outbreak of COVID-19 in December 2019. The disease has aggressively spread around the world, including Indonesia. The emergence of COVID-19 has serious implications for public health and socio-economic development worldwide. No country is prepared to face COVID-19. Because of the rapid transmission of COVID-19, the early warning systems (EWS) in each country are not prepared to deal with it. Controlling and preventing COVID-19 transmission in an effective and efficient manner is critical not only for public health, but also for economic sustainability and long-term viability. Consequently, an efficient and effective EWS for COVID-19 is required. The EWS for COVID-19 must be capable of monitoring and forecasting the spatiotemporal transmission of COVID-19. This study demonstrates how an EWS could be a proactive system that would be able to predict the spatiotemporal distribution of COVID-19 and detect its sudden increase in small areas such as cities. Early COVID19 data in Bandung, Indonesia from 17 March 2020 to 22 June 2020 was used to demonstrate the construction of an effective and efficient EWS using the spatiotemporal model. We observed that the relative risk of COVID-19 2 JAYA, ANDRIYANA, TANTULAR, KRISTIANI fluctuates geographically and temporally, gradually increasing throughout the estimate phase (17 March 2020-22 June 2020) and increasing slightly during the prediction period (23 June–06 July 2020). We discovered that human mobility is a major aspect that must be addressed in order to minimize COVID-19 transmission during the early pandemic phase.

. Four elements of an early warning system A failure in any of these elements could cause the system to fail (see [14,15] for detail). The first and second components entail the monitoring and forecasting of the geographic evolution of disease risk in the environment. It has a strong relationship with the statistical model on which we will be concentrating. The EWS for COVID-19 must be capable of monitoring and forecasting the spatiotemporal transmission of COVID-19. It is discussed in this article how an EWS could be a proactive system that would be able to predict the spatiotemporal distribution of COVID-19 and detect its sudden increase.
• Communicate information and early warning • Is the warning information clear and usable?
• Build rational and community response capability • Are people prepared and ready to react to warning?
• Develop hazard monitoring and early warning service • Can accurate and timely warnings be generated • Systematically collect data and undertake risk assessment • Are risk maps and data widely available COVID-19 study is typically conducted in large areas, such as regional areas [16] and countries [17,18]. There has been very little research conducted on such a small region as a city.
COVID-19 will spread more densely in a limited region due to its high level of mobility, which results in a high degree of spatial dependence. Small area characteristics should be taken into account while building an EWS. Additionally, significant risk factors should be identified during the pandemic's early stages. Numerous studies have indicated that human mobility may have an effect on COVID-19 transmission. However, quantifying human movement is fairly difficult [19,20] and some proxies are required.
The fundamental concept in assessing disease risk in EWS is relative risk. The standard method for calculating relative risk is to divide the number of cases by the expected number of case, called as standardized case ratio (SIR) [21][22][23][24][25]. However SIR provides unreliable estimate of the relative risk for small area with small number of case and population at risk [26][27][28][29]. Bayesian smoothing is frequently used to solve the unreliability problem associated with SIR. A hierarchical Bayesian spatiotemporal model is used to smooth, model, forecast, and map disease risk for spatiotemporal data. It provides a reliable estimate of the relative risk over space and time by accounting for spatially and temporally structured and unstructured effects, as well as their interactions [26,30]. The reliable estimate of relative risk is required for effective and efficient EWS. Additionally, the effectiveness of early intervention techniques remains to be determined by an analysis of available surveillance data.
The purpose of this retrospective analysis is to gain a better understanding of the disease patterns associated with COVID-19 in this region and, as a result, to guide future pandemic response.
The remainder of the paper is divided into the following sections. The next section discusses the materials and methods used in this study. Following that is the Application section, which details the COVID-19 EWS in Bandung, Indonesia. The last part discusses, summarizes, and makes recommendations for additional research.

Data sources
We use COVID-19 data in Bandung city, Indonesia to demonstrate the construction of an effective and efficient EWS using the spatiotemporal model for small areas and identify the critical risk factors. The COVID-19 dataset was obtained from a website in Bandung that provides realtime information on COVID-19 pandemic disease outbreaks (https://covid19.bandung.go.id/). The website keeps track of newly confirmed cases in Bandung by district and date. Between 17 March 2020 and 22 June 2020, we extensively observed COVID-19 press announcements and situation reports produced by 30 Bandung district health agencies. Because we are dealing with a small area, confirmed cases are extremely rare on a daily basis. As a result, to avoid a large number of zero cases, we take the cumulative cases every week. It is based on COVID-19's incubation period.
Symptoms may appear 2-14 days after virus exposure [31]. As proxy measures of human mobility, we used population density and poverty rate. The population at risk, population density and the poverty rate on 2019 were extracted from Bandung Data (https://data.go.id/). Table 1 displays detailed data are used in this study for each district.

Moran's I
For an effective monitoring and controlling COVID-19, the prior information on spatiotemporal autocorrelation of COVID-19 risk transmission is required [3]. Spatiotemporal Moran's I statistic on case rate is used. It is defined as [24].

Bayesian hierarchical spatiotemporal-model
Bayesian hierarchical models have been used often to evaluate spatiotemporal disease transmission 8 JAYA, ANDRIYANA, TANTULAR, KRISTIANI extending [33][34][35]. It has been successfully to forecast disease risk and monitor the disease transmission [24] which are crucial components in developing EWS. Here we adopt the one defined by Knorr-Held (2000) [36] in which the spatial, temporal and spatiotemporal interaction components are model as random effects. The interaction components were included in the model to consider the variation of temporal trend over districts. The Bandung city is divided becomes districts over periods. For the district in periode we assume COVID-19 cases follows Poisson distribution 1 [24]: where denotes the expectation and variance of at district and time . In order to study of the relative risk, is defiend as a product of expected number of cases ( ) and the relative risk ( ); = , = 1, … , and = 1, … , . The expected number of confirmed case is defined as [24,25]: Now we will examine the Poisson distribution's mean, which we will decompose using the natural logarithm link function [24,25]: The second component in Eq. (4), = log( ), is the focus of further research. Specifically: where is the intercept representing the overall relative risk ; , , and are the random effect components that are spatially structured and unstructured effects, and temporally structured and unstructured effects, respectively. is a representation of spatiotemporal of interaction. The random effect of region i ( ) is described spatially structured using the intrinsic conditional autoregressive (iCAR) prior [33]: where − refers to the elements in except the th element, = ( ) is queen spatial weight matrix. is the precision parameter of . The random effect of region ( ) is spatially unstructured and follows an exchangeable normal distribution (i.e. a sequence of random variables that are independently and identically normal distributed) [24,25]: where is the precision parameter of . A prior for temporal trend ( ) is a random walk of order one (RW1) [24]: with being the precision parameter. We may use a random walk of order two (RW2) instead of a RW1. This RW2 priors is commonly used if the data has a pronounced linear trend. The temporal trend ( ) of a RW2 is [24,37]: Temporally unstructured effect is assumed to follows exchangeable normal distribution [24,25]: with being the precision parameter of . The interaction effect follows for different structures I-IV which describe the product of spatially and temporally structured and unstructured effects (see [24,36] for detail).
We specify a vague Gaussian prior distribution with zero mean and a large variance 2 = τ −1 for , i.e. ~(0, 10 6 ) and half Cauchy (HC) prior for hyperparameters. We proposed 25 as scale parameter for the HC hyper-prior. It is possible that not all components in the model (5) must be included in the model. For the purpose of evaluating our model using deviance information criterion (DIC). The main objective of this paper is to explore the spatiotemporal distribution of COVID-19 and make a relative risk prediction for two weeks ahead. Forecasting with INLA can be easily implemented by imputation missing value scenarios. We can enter 'Not Available (NA)' for the observations that need to be forecasted [38].
For the visual representation of the geographical distribution of COVID-19 we present choropleth maps.

Bayesian spatiotemporal model
Preparing to run the spatiotemporal model (5) We analyze six different sub-models of model (5)    determined to fit the model selection criterion. It has a low DIC, WAIC, MAE, and RMSE but a high MPL and r. As a result, for the following analysis, we will use this model, which we will refer to as M3. The posterior means of the hyperparameters are presented in Table 3. The spatial autoregressive coefficient ( ) of 0.46 suggests that spatial dependence between adjacent districts is moderate. A random effect's posterior mean of variance quantifies its contribution to the spatiotemporal variation in COVID-19 risk. As shown in Table 3, the posterior means of the random effects' variances range significantly, ranging from 0.082 for the spatially unstructured effect to 1.293 for the temporally unstructured effect's variation. Additionally, Table   3 illustrates the fraction variance of the hyperparameters. The spatiotemporal variance in COVID-19 risk in Bandung is best explained by temporally structured and unstructured effects with total fraction variance is around 86.92%. This shows that there are risk variables that vary over time and play a significant role in COVID-19 transmission, such as human mobility. Additionally, as illustrated in Figure 3, the Cicendo district is consistently classified as a high-risk area. To determine whether or not the high-risk regions constituted a significant hotspot for COVID-19 risk. We next calculated the exceedance probability Pr(̂> 1| ). Figure 4 shows the exceedance probability of estimated and forecasted result of COVID-19 relative risk. In some regions of Bandung, a spatiotemporal relative risk greater than one can be observed with posterior probability more than 0.8, indicating a relatively low level of related uncertainty. In general, the relative risk and associated uncertainty vary significantly by geography.
The biggest relative risk is seen in Bandung's western and southeast districts, including Antapani, Astana Anyar, Bandung Wetan, Cicendo, and Gede Bage.  The exceedance probability Pr(̂> 1| ) is depicted in Figure 4. Districts with Pr(̂> 1| ) more than 0.80 may be classified as high-risk zones or hotspots. Figure 4 demonstrates that during weeks 3, 5, 6, and 12, more than 90% of the districts are classified as high-risk or hotspot. It indicates that during those weeks the number of new cases were significantly increased. Cicendo district has a rather high exceedance probability for the forecast period week 15. During the forecast period of 29 June to 06 July 2020, we anticipate an extra 20 cases. In regard to the early warning system, we attempted to identify risk variables that should be controlled in order to prevent the spread of COVID-19. We consider human movement to be the primary risk factor. However, measuring and obtaining human mobility is fairly challenging. According to studies conducted by [39] social mobility is significantly related to population density and poverty.
These two variables are used to account for human movement.  After obtaining COVID-19 forecasts for the next two weeks, we examine the association between population density and poverty rate and estimated of the COVID-19 risk. We done the analysis separately to avoid spatiotemporal confounding problem [25]. We discovered a positive correlation between population density and COVID-19 risk, but a negative correlation between poverty rate and COVID-19 risk. This finding corroborated multiple earlier studies [40]. Human mobility was previously measured by population density and poverty rate. Economic growth is frequently accompanied by an increase in social mobility [39]. In Bandung areas, densely populated places, human mobility is rather high, while it is quite low in impoverished ones. As a result, it is possible to deduce that human movement is a crucial role in COVID-19 transmission [40].

DISCUSSION
The mapping and forecasting of diseases are inextricably linked to risk assessment and the EWS warning service. Disease risk mapping has existed in public health and epidemiology for a long period of time [41] Bayesian spatiotemporal disease mapping has been effective in identifying areas at risk [42]. Numerous studies on the COVID-19 pandemic focused on confirmed cases rather than risk factors, oblivious to the fact that the population at risk varies over time and space.
It is critical to identify clusters or hotspots of high-risk individuals using the adjusted population at risk. The ratio of observed to expected cases is referred to as the relative risk or excess risk. Not only can spatial and temporal analysis be used to visualize epidemiological data and aid in intuitive disease distribution, but it can also be used to identify spatial and temporal clusters, as well as areas of high and low risk [30]. Using new technologies, GIS analysis, and highly structured mathematical and statistical techniques, the occurrence of infectious diseases can be described and analyzed. Bayesian methods are advantageous for analysing models with complex and flexible structures that accurately represent the characteristics of a particular geographical environment or disease for small area [43].
This is the first study to document the spatiotemporal patterns of the COVID-19 pandemic in Bandung, Indonesia, an area critical for EWS. As a baseline, we used Spatiotemporal Moran's I 19 EARLY WARNING SYSTEM FOR COVID-19 statistics on prevalence rates to assess the variation in disease risk spatial dependence over time.
To gain a better understanding of the COVID-19 pandemic's spatiotemporal pattern, the Bayesian spatiotemporal model was used to smooth the relative risk of COVID-19 [36]. It is useful for identifying high-risk clusters and forecasting the COVID-19 pandemic relative risk over a onemonth period. The Poisson model was found to be superior to the Negative Binomial, Zero inflated Poisson, and Zero inflated Negative Binomial model using DIC, WAIC, MAE, RMSE and Pearson correlation (r). The optimal model included spatially structured and unstructured effects, temporally structured and unstructured effects, and interaction type I. We discovered that the order Due to population mobility, the COVID-19 pandemic spread from Cicendo to other districts [44].
This possibility lends credence to the study's discovery of spatial dependency. Using a Bayesian spatiotemporal model, we discovered clear spatiotemporal transmission of the COVID-19 pandemic. The disease spread from Bandung's west to southeast. COVID-19 is extremely likely to spread between geographically neighboring places, according to the results of the examination utilizing geographical neighbourhoods. This could be due to the fact that residents of nearby regions frequently interact. Additionally, the district of Gede Bage in the southeast region, which has a low number of confirmed cases, was designated as a high-risk region due to the district's small population at risk and a higher than expected number of confirmed cases. When comparing risks across space and time, it is necessary to consider the population at risk, even more so when the population at risk varies significantly across space and time. Additionally, we discovered that human mobility, as defined by population density and poverty rate, is a crucial component in COVID-19 transmission during the early pandemic.
Two additional risk enhancements are visually identified as high-risk clusters in Bandung's western and south-eastern regions on the choropleth map. Due to a physician shortage and a lack of hospital and health center beds, the Bandung government should prioritize clusters at high risk.
This data can be used to develop a highly effective and efficient early warning system for COVID-19 transmission control. Bandung's government should place a greater emphasis on high-risk clusters, allocate additional resources to high-risk areas, and implement health protocols such as mask use, social distancing, hand washing, and avoiding crowded areas with limit the human mobility. Human mobility can be restricted by enforcing local lockdowns in locations regarded as hotspots.

CONCLUSION
The COVID-19 pandemic is the century's most extraordinary health problem and the greatest threat to humanity since the Second World War. It is rapidly spreading throughout the world, as Wuhan, China reported an outbreak on 30 December 2019 [3]. On 1 March 2020, the WHO declared COVID-19 to be a pandemic. As of 5 June 2020, the coronavirus COVID-19 was present in 213 countries and territories worldwide, as well as on two international transports [5].
This study demonstrates the COVID-19 pandemic in Bandung, Indonesia, on a small spatiotemporal scale. In small regions such as Bandung, the COVID-19 pandemic spread rapidly from district to district despite the government's policy of restricting people's mobility. Examining the spatiotemporal spread is critical to preventing the local transmission and second wave from increasing. This is believed to be the first study to examine the virus's spatiotemporal transmission using a combination of Moran's I and Bayesian spatiotemporal models. The forecasting outcome 21 EARLY WARNING SYSTEM FOR COVID -19 can be used to develop a robust early warning system. Future research, such as an examination of the spatiotemporal distribution of this disease based on demography, hereditary disease, and patient age, will aid in its control and prevention.
Through this work, it will be possible to ascertain the factors that influence death and recovery. The Bayesian spatiotemporal model can be used to account for covariates and to control heterogeneity [45,46].