Spatiotemporal Study of COVID-19 in Fars Province, Iran, October-November 2020: Establishment of Early Warning System

Background Using time series and spatiotemporal analyses, this study aimed to establish an Early Warning System (EWS) for COVID-19 in Fars province Iran. Methods A EWS was conducted on (i) daily basis city-level time series data including 53 554 cases recorded during 18 February–30 September 2020, which were applied to forecast COVID-19 cases during 1 October–14 November 2020, and (ii) the spatiotemporal analysis, which was conducted on the forecasted cases to predict spatiotemporal outbreaks of COVID-19. Results A total of 55 369 cases were forecasted during 1 October–14 November 2020, most of which (26.9%) occurred in Shiraz. In addition, 65.80% and 34.20% of the cases occurred in October and November, respectively. Four significant spatiotemporal outbreaks were predicted, with the Most Likely Cluster (MLC) occurring in ten cities during 2–22 October (P < 0.001 for all). Moreover, subgroup analysis demonstrated that Zarrindasht was the canon of the epidemic on 6 October (P=0.04). As a part of EWS, the epidemic was triggered from Jahrom, involving the MLC districts in the center, west, and south parts of the province. Then, it showed a tendency to move towards Zarrindasht in the south and progress to Lar in the southernmost part. Afterwards, it simultaneously progressed to Fasa and Sepidan in the central and northwestern parts of the province, respectively. Conclusion EWS, which was established based on the current protocol, alarmed policymakers and health managers on the progression of the epidemic and on where and when to implement medical facilities. These findings can be used to tailor province-level policies to servile the ongoing epidemic in the area; however, governmental level effort is needed to control the epidemic at a larger scale in the future.


Introduction
e surveillance systems for infectious diseases are encouraging well-developed techniques that can warn them by detecting the exact time and place of disease occurrence and, subsequently, targeting interventions and resources to more risky areas [1]. Some infections follow space and time trends for which Early Warning Systems (EWS) can be useful tools to find the potential outbreaks [2]. EWS are alert systems to predict epidemic outbreaks in the region. In the 1910s, Captain S.R Christophers from the British army set a system to predict malaria in India. Since then, it has been used as a useful tool to aid infected nations, make health policies, and manage services to the infected area [1]. e epidemic statue of SARS-CoV-2 disease (COVID-19) was officially declared by the World Health Organization in January 2020 [3,4]. What makes it important to set a EWS for ongoing COVID-19 epidemic is that it could early notify the policy makers and health managers on the time and location of the outbreaks, so that they could focus on more needy areas and dedicate more medical facilities. Finally, people are helped with less COVID-19 threat, infections, and death. e question is that, using COVID-19 time series data, if the current study could establish a EWS for future  outbreaks, the surveillance system could act timely, prioritize times and locations for targeting interventions, allocate resources more efficiently, and improve decision-making in the area.
Time series data refer to a sequence of successive naturally temporal data, which are equally spaced in points of time. A time series analysis consists of methods for analyzing time series data and forecasting the future values based on the previously observed values [5]. Spatial and Temporal Scan (SaTScan) modeling discovers more about the space and time of an event occurrence. Space-time permutation scan statistic, which is derived from SaTScan, was first introduced by Martin Kulldorff in 2015 [6]. is model aims at determining whether the cases are distributed uniformly over time and space. e statistic utilizes scanning windows moving systematically across the study area to identify the clusters of cases. is modeling helps surveillance systems decide timely on more at risk areas [7]. Using one model output as another mode input, this study could develop a protocol of EWS in the area by (i) forecasting the COVID-19 cases, (ii) discovering the future space-time outbreaks of COVID-19, (iii) predicting the progression of epidemic, and (iv) having more control on the epidemic.

Study Design, Data Source, and Study Variable.
Daily basis city-level time series data including 53 554 cases recorded in 26 cities of Fars province, Iran, during 18 February-30 September 2020 (225 days) were applied to forecast COVID-19 cases during 1 October-14 November 2020 (45 days). e cases were obtained from 44 COVID-19 sampling centers affiliated to Shiraz University of Medical Sciences (SUMS). All data were reported to SUMS either online, https://www.coronalab.sums.ac.ir, or by manual registration. ere were no exclusion criteria, meaning that all cases were included in the study. e COVID-19 cases were the positive cases confirmed by real-time Reverse Transcription Polymerase Chain Reaction (RT-PCR). Additionally, the geographical coordinates of the locations were obtained through Google-Earth (US Dept. of State Geographer 2021) based on the latitude-longitude coordinate system.

Case Study and Area.
Fars province, the fourth most populated (4 852 274 dwellers) and widespread (122 608 km 2 ) province amongst 31 provinces in Iran, is located at 27°3′ and 31°40′ northern latitude and 50°36′ and 55°35′ eastern longitude in south of Iran. It includes 26 cities, with Shiraz, the fifth most populated city in Iran, being the capital. Marvdasht, Jahrom, Fasa, Kazerun, Darab, Firoozabad, Lar, and Abade, also, are the most populated areas (more than 100 00 dwellers per capita). 67.6% of the province are urban dwellers (urban areas and suburbs), 31% are villagers (small towns and rural areas), and the rest are nomads (no fixed habitation who regularly move to and from the same areas breeding livestock) [8].

Statistical Analysis.
Mean difference and Std. Error of difference were used to describe quantitative variables, and frequency (relative frequency) was used to describe qualitative variables. e distribution of COVID-19 cases across cities and times are tested using independent-sample one way ANOVA and pair wise comparisons (LSD test). Additionally, time series analysis including (i) Akaik Information Criterion (AIC)/Bayesian Information Criterion (BIC), which are the estimators of the relative quality of time series models comparative to other models for a given set of data, (ii) Auto Correlation Function (ACF)/Partial Auto Correlation Function (PACF) sample/model, which assess the conformity of the observed and fitted pattern in the data, and (iii) residual analysis, which evaluates the presumptions of white noise parameter, was used to forecast the COVID-19 cases during 1 October-14 November 2020 (45 days). A set of time series models were applied, and the best ones were selected based on lower AIC/BIC scores, ACF/PACF conformity, and residual analysis.
Descriptions on time series models: A simple additive decomposed time series model is Appling time series forecasted cases in spatiotemporal analysis, high-rate clusters were detected considering the minimum and maximum temporal sizes that were equal to one day and 50% of the study period, respectively. e minimum and maximum spatial sizes were also equal to two cases and 50% of at risk population, respectively. Temporal data were checked to ensure that all the cases were within the specified temporal study period. In addition, geographical data check was done to ensure that all the observations were within the specified geographical area. No geographical overlap and subgroup analysis were used.

Validation Study.
A validation study using real data from 1 October-14 November 2020 was done to assess the accuracy of the results.

Distributions of 53 554 COVID-19 Cases in Fars
COVID-19 cases in February, March, April, and May were significantly lower than those in June, July, August, and   September (P < 0.05 for all, respectively); COVID-19 cases in July were significantly higher than those in August (P < 0.05 for all); COVID-19 cases in August were significantly lower than those in September (P < 0.05 for all).

Forecasting COVID-19 Cases in Fars Province during 1
October-14 November 2020. From the 53 554 observed cases during 18 February-30 September 2020, 55 369 cases were forecasted during 1 October-14 November 2020 using variety of time series models. e frequency of 55 369 forecasted cases alongside the results of time series analysis for the 26 cities has been presented in Table 5. From the 55 369 forecasted cases during 1 October-14 November 2020, the highest proportion was related to Shiraz (26.9%). Additionally, 65.80% and 34.20% of the cases were forecasted to occur in October and November, respectively. e time series results for Jahrom and Shiraz have been presented in Figures 1 and 2, respectively. e interpretation of the results was similar for the 24 other cities.

COVID-19 Time Series Analysis in Jahrom during 1
October-14 November 2020. To reach the stability of variance, Box-Cox transformation was set zero. Besides, SARMA 12 (3,4) was fitted with the observed quadratic trend and seasonality of 12. e estimated COVID-19 cases during 18 February-30 September 2020, the predicted COVID-19 cases during 1 October-14 November 2020 (45 days), and ACF/PACF sample/model for Jahrom have been presented in Figure 1. SARMA 12 (3,4) model was shown to be a good fit that can predict COVID-19 cases in Jahrom during October 1st -November 14th 2020.

Spatiotemporal Analysis.
e results of the spatiotemporal analysis including 55 369 COVID-19 forecasted cases at the city level on a daily basis together with the results of subgroup analysis have been presented in Table 6.
Overall, four outbreaks occurred in Fars province during 1 October-14 November 2020, the largest and longest of which occurred in ten cities during 2-22 October, with Jahrom being the canon (MLC). Outbreaks 2 and 3 occurred in Sepidan and Fasa, respectively, and lasted from 3 to 13 November. Besides, outbreak 4 occurred in Lar and Zarrindasht on 21 October. e results of subgroup analysis demonstrated that Zarrindasht was the canon of the epidemic on 6 October.
As a part of EWS, the progression of COVID-19 epidemic resulting from spatiotemporal outbreaks of the 55 369 COVID-19 cases in Fars province, Iran, during 1 October-14 November 2020 has been shown in Figure 3.
COVID-19 epidemic inclined to trigger from Jahrom on 2 October, involving the MLC districts in the center, west, and south parts of the province; then, it showed a tendency to move towards the south in Zarrindasht on 6 October. Afterwards, it progressed to the southernmost part in Lar on 21 October. Afterwards, it simultaneously progressed to Fasa and Sepidan in the central and northwestern parts of the  is progression alarmed the policy makers and health managers on the potential starting point of ongoing epidemic (Jahrom city on 2 October, 2020) and, also, on when and where to prioritize the medical facilities and staffs; moreover, it could intervene with the transmission chain of COVID-19 in the area.  Table 7.
A total of 12 cities got involved in the outbreaks during 1 October-14 November 2020 using real data, nine of which were predicted in the EWS. e EWS predicted time periods were either the same as the observed periods or occurred earlier. Based on the validation study findings, the results of EWS were shown to be valid enough for our purpose.

Discussion
Covid-19 cases did not distributed uniformly across cities and times during 18 February-30 September 2020 in Fars province, Iran. In this protocol presented as EWS, 55 369 COVID-19 cases were forecasted from the 53 554 daily basis city-level time series data recorded during 18 February-30 September 2020. e forecasted cases resulted in the prediction of four significant spatiotemporal outbreaks of COVID-19 in the area including the MLC and three other clusters. e MLC, which was also the largest and the longest cluster, included ten cities during 2-22 October, with Jahrom being the canon. is finding, schematically presented as the progression of the epidemic during 1 October-14 November 2020, revealed that the epidemic inclined to trigger from Jahrom on 2 October, involving the MLC districts in the center, west, and south parts of the province; then, it showed a tendency to move towards the south in Zarrindasht on 6 October. Afterwards, it progressed to the southernmost part in Lar on 21 October. Later, it simultaneously progressed to Fasa and Sepidan in the central and northwestern parts of the province, respectively on 3 November. To the best of our knowledge, this protocol alarmed the surveillance system to focus on the progression of the epidemic and especially on the starting point, Jahrom, so that proper interventions could be done on the transmission chain of SARS-CoV-2 resulting in lower prevalence of the disease in the area. Moreover, the surveillance system can  dedicate more test sites, medical resources, community quarantine, vaccination coverage, and travel bans on more at risk locations at specified detected times.
Despite being the most important transmission line of the disease, it should be explained why the most populated city, Shiraz, with the maximum number of COVID-19 cases    was not included in the predicted outbreaks. Shiraz was ignored in permutation scan statistics analysis, in which clusters are ranked by their likelihood ratio scores rather than by the population density or the area size. is can be attributed to the fact that, with the capital city of Fars province being with advanced medical services, many patients from nearby cities travel to Shiraz to receive health services, thereby transmitting the infection to other cities of the province. Time series models based on statistical methods were revealed to fit well for forecasting time dependent data. ey presented well-fit results in case of the COVID-19 epidemic as well. For instance, the AR model could forecast the confirmed COVID-19 cases in Iran [10]. Additionally, ARIMA-based time series analysis forecasting near future could assist the government to be prepared for the upcoming emergencies of the current epidemic in India [11][12][13][14]. In another study, ARIMA models were fitted to forecast shortterm COVID-19 cases in the five most affected countries of the world including Italy, Spain, France, the United Kingdom, and the United States of America [15]. Space-time permutation scan statistics could well manage space-time dependent variables through informing the policymakers on the progression of the epidemic in China [16][17][18]. In the United States, the space-time outbreaks of COVID-19 were detected during two periods of time (22 January to 9 March 2020 and 18 January to 27 March 2020) using space-time statistics [19]. Space-time scan statistics also resulted in the detection of space-time outbreaks in Hong Kong during 23 January-14 April 2020 [20]. Furthermore, space-time permutation scan statistics worked better for predicting highly frequent infectious diseases with high transmission rates [8,21,22]. Another study employed almost the same protocol as the one used to establish the EWS in the current study in order to detect the vulnerable districts for COVID-19 in India. Using ARIMA time series model for forecasting short-term COVID-19 cases and performing spatial analysis, they indicated that the western and southern parts of the country were highly vulnerable for COVID-19 infection [23,24].
EWS establishment could effectively warn the surveillance system about the progression of the epidemic in Fars province, Iran. However, it had several limitations, which might have affected the results. Firstly, the surveillance system of infectious diseases in Iran is a passive system, and many SARS-CoV-2 infected people with mild to moderate symptoms did not refer to health centers and were not diagnosed. erefore, the forecasted cases of COVID-19 might have been underestimated. is could affect estimating the required health services and medical facilities. Secondly, permutation scan statistics discover more than one cluster in a given area and time. Hence, interpretation of the results could be somewhat difficult, especially when incorporating main-cluster and subcluster results. Regarding clinical and statistical significance, prioritizing the clusters can also be difficult. Another limitation of scan statistics is that it uses a circular scan window.
us, the precise limits of the detected clusters remain uncertain. In other words, either the site within the detected clusters may not have a high risk, or a high risk site may be out of the cluster. Yet, using subgroup analysis for detecting significant clusters within large clusters could solve the problem. Further investigations are suggested to assess the genetic variation and meteorological risk factors of COVID-19 in the long run.

Conclusion
e EWS presented here would alarm the policy makers and health managers on the potential starting point of ongoing epidemic and, also, on when and where to prioritize the medical facilities and staffs. e MLC, which was also the largest and the longest cluster, included ten cities during 2-22 October, with Jahrom being the canon. is finding revealed that the epidemic inclined to trigger from Jahrom on 2 October, involving the MLC districts in the center, west, and south parts of the province; then, it showed a tendency to move towards the south in Zarrindasht on 6 October. Afterwards, it progressed to the southernmost part in Lar on 21 October. Afterwards, it simultaneously progressed to Fasa and Sepidan in the central and northwestern parts of the province, respectively, on 3 November. To the best of our knowledge, this protocol alarmed the surveillance system to focus on the starting point and progression of the epidemic as well as to dedicate more test sites, medical resources, stricter quarantine, vaccination coverage, and travel bans on more at risk locations at specified times. ese findings can be used to tailor province-level policies to servile the ongoing epidemic in the area; however, governmental level effort is needed to control the epidemic at a larger scale in the future.
Data Availability e health service data used to support the findings of this study are restricted by the Ethics Committee of the Ministry of Health, Treatment, and Medical Education affiliated to the Shiraz University of Medical Sciences, Shiraz, Iran, in order to protect patient privacy. Data are available from Marjan Zare, Central Building of Shiraz University of Medical Sciences, Vice Chancellor for Health, fourth floor, Zand St, Shiraz, Iran, telephone number: 07132122368, E-mail: marjan.zare@gmail.com, for researchers who meet the criteria for access to confidential data.

Ethical Approval
All ethical steps including data collection and analysis as well as reporting the results were in accordance with the standards approved by the Ethic Committee of the Ministry of Health, Treatment, and Medical Education (IR.SUMS.-REC.1399.574). e work processes were anonymous, and the results were reported to the study participants.

Conflicts of Interest
No conflicts of interest were declared by the authors.