Real time surveillance of COVID-19 space and time clusters during the summer 2020 in Spain

Spain is composed of 17 Autonomous Regions and 2 Autonomous Cities. These are subdivided into 52 provinces (shown in Map 1). The smallest territorial administrations are municipalities. As of January 1, 2020, there are 8131 of them. Their number and extension are heterogeneous among territories. We used 2019 population data for each Spanish municipality from INEbase, the National Statistics Institute of Spain official database [18].

COVID-19 cases, recorded by Autonomous Regions and Autonomous Cities as part of the National Epidemiological Surveillance Network (RENAVE in Spanish), are stored in the Spanish Surveillance System electronic platform (SiViES in Spanish), and managed by the National Centre for Epidemiology. This database contains information on demographic, epidemiological, clinical and laboratory aspects. The official protocols implemented during this period aimed at the early detection of symptomatic cases and, if possible, the active search for cases (both symptomatic and asymptomatic) among the contacts of the cases [5]. A COVID-19 case is considered confirmed, and therefore is notified to RENAVE if there is a positive polymerase chain reaction (PCR) test, or an ELISA-based serological test (IgM) in patients with suitable symptoms and negative PCR [5].

For the analysis, the date assigned to each case was computed as the imputation date. The date of symptom onset and the date of diagnosis were used for symptomatic and asymptomatic cases, respectively. When date of symptom onset was missing, the imputation date was calculated as diagnosis date minus 3 days (median time from symptom to diagnosis in our study) [5]. Cases with none of these dates available were excluded.

A prospective version of Poisson STSS was employed to detect active clusters at municipal level in Spain. STSS is characterized by a cylindrical window where the base is the spatial scanning window and the height corresponds to time. The method scans through space and time analyzing each possible geographic area and time range. Thus, we obtain an extensive number of cylinders for the entire area of analysis.

As we assume the number of COVID-19 cases follow a Poisson distribution, under the null hypothesis (Ho) the risk within the cylinder is constant while under the alternative hypothesis (H₁) the risk inside the cylinder differs from outside. As stated by Kulldorf [10] and Desjardins [15], the equation below (Eq. 1) is used to calculate, under the null hypothesis, the expected number of cases (μ) where p is the population in cylinder base area, C the number of COVID-19 cases in Spain and P the total Spanish population.

Then, a likelihood ratio test (Eq. 2) [8, 15] for each cylinder is calculated taking into account the number of cases and the population at risk, the higher the value the less likely that the detected cluster occurred by chance.

The ratio is defined by the quotient between the likelihood of the cylinder Z, L(Z), and, L₀, the likelihood under the null hypothesis. The variables are: n_z, the number of COVID-19 cases within the cylinder area; μ(Z), the expected cases calculated by the Eq. 1; N, the total number of cases in Spain across the time range and μ(T) the expected cases in the area across the time range.

The significance was evaluated for each detected cluster using the Monte Carlo test. The obtained likelihood ratio was compared to likelihood distribution calculated with the permutations of the data. For each analysis (day), clusters are presented in order of likelihood ratio of occurrence.

To easily compare the relative risk between the detected clusters, a relative risk for each cluster is calculated with the formula below (Eq. 3) [13, 15]: where c and e are the observed and expected cases in a cluster, respectively, and C is the total number of cases in Spain.

SiViES data was retrieved during September 2020, and a daily prospective analysis was emulated for the study period. To this end, only cases recorded up to the analyzed day were taken in consideration, and only active clusters were reported, not accounting for outcomes of following days. Table 1 summarizes parameters used in the STSS analysis. The cylindrical window maximum radius for our analysis was 25 km (mean distance between municipalities in Spain) whereas the maximum time period of aggregated analysis was from 2 to 7 days (to include median incubation period of 5 days).

STSS was performed using the SaTScan™ v9.6 software (https://​www.​satscan.​org/​). The automatization of the process was carried out with R Software 4.0.2 and the package rsatscan [19]. Additional graphics and maps were obtained with ggplot2 [20] and ArcMap 10.3. A dynamic online viewer was developed using the packages Leaflet [21] and shiny [22] and uploaded to a server to be consulted by the public.

From June 21st to August 31st, a total of 257,881 COVID-19 cases were registered through RENAVE. Figure 1a shows the epidemic curve, where an increasing trend of daily new cases can be appreciated beginning in July, shortly after the full lifting of control measures.

Cluster evolution followed a similar trend (Fig. 1b). Its spatio-temporal evolution can be consulted in a Shiny web application, available online at the following link: (https://​coviddifusion.​isciii.​es/​SpSumClus/​). As can be observed in the viewer, at the first day of our study, June 21th, a total of 7 active clusters were active in Spain, mostly located in the north-eastern regions of Aragón, Cataluña and País Vasco, which encompass a geographical region called Ebro’s valley. Since then, and under our analysis parameters, a constant increase in the number of clusters was noted, reaching around 50 active clusters by late-July, and 100 by the second week of August. It must be noted that the orange-colored cluster represents the most likely in terms of likelihood.

In July, these clusters were grouped into two areas: Ebro’s valley and Madrid. The Mediterranean coast was also affected, reaching Andalucía. As the time passed, cluster distribution became more homogeneous throughout the national territory. In August, clusters had extended from Madrid to neighboring areas of Castilla-La Mancha and Castilla y León. By August 31st, cluster number had grown to 129, extending over almost every region of Spain with notable exceptions such as Asturias.

The location of the most likely cluster has changed throughout the study period. Initially, it was placed in some municipalities of Huesca and Lleida, and later, around the city of Zaragoza. Later on, it was displaced to the Barcelona metropolitan area by July 20th, and from August 8th to the surroundings of the capital city of Madrid.

A more detailed insight on the study outcomes is discussed below, focusing on two specific days (June 25th and August 1st) as they help describe the spread of the disease in the Spanish territory (Table 2 and Map 2).

The scan analysis for the 25th of June detected seven statistically significant emerging space-time clusters of COVID-19 are described at Table 2 and shown in Map 2a. The most likely cluster (cluster 1) was located between two administrative regions: Huesca (Aragón) and Lleida (Cataluña). It presented a relative risk (RR) of 28.8, with 122 observed cases compared to 4.26 expected, and it was active from June 23rd to the date of analysis. Although most cases were located in 3 municipalities, 15 out of the remaining 25 included municipalities had a RR > 1.

Other small clusters are found in the south, cluster number 2 and 4 (Málaga and Huelva) and cluster 5 in the north of Spain (Navarra), each including a single municipality. They involved 62, 9 and 4 cases, respectively, with a RR of 8.23, 24.95 and 300.36. The start date of the active cluster varies: June 24th (cluster 2 and 4) and June 23rd (cluster 5); while the following analysis date was the same (June 25th).

Cluster 3 contains 18 municipalities, one of them shows the highest individual RR for the day of analysis (478.07). The global cluster RR was 35.44, with only 3 municipalities with a RR > 1.

By August 1st, 2020, 62 clusters of COVID-19 were active in Spain, present over the whole territory but two Autonomous Regions and the Autonomous Cities. They are described in Table 2 and shown in Map 2b. The number of cases ranged from 7 in the smallest aggregation to 4359 in the largest one. Cluster duration had increased and more than half of them had started 5 days before (on June 27th).

The Mediterranean coast of Cataluña was also affected during this period. Cluster 2 is located in Barcelona and its surroundings. A total of 4359 cases were reported, affecting 86 municipalities (72 with a RR > 1) and with a RR of 9.87 for the complete aggregation.

Cluster 3 is located in the Madrid Autonomous Region and is formed by 2928 cases (RR: 5.31). It encompasses Madrid and its metropolitan area (30 of 31 municipalities with RR > 1). Around this cluster, there are 4 less-likely ones, at the rural areas of this administrative region. The remaining significant clusters can be consulted at Map 2b.

By that day, we can observe that practically most of north-eastern Spain was affected by COVID-19 clusters as well as Madrid and the Mediterranean coast, especially the south-eastern regions of Murcia, Almeria and Alicante.

Cluster radius and number of municipalities per cluster allow us to assess another quantitative dimension of pandemic evolution (Fig. 2). Assuming an initial scenario of low transmission during the first 2–3 weeks of our study period, a grouping of small clusters can be observed, with a mean radius of 10 to 14 km that includes 10 to 15 municipalities (Fig. 2a and b). Around mid-July an increase is noted in both parameters, with a mean radius ranging from 18 to 20 km and a daily mean of 30 municipalities per cluster, this value will remain up to August 31st (Fig. 2a and b). The third included parameter (Fig. 2c), mean time duration of active clusters, presents the same evolution but in a more progressive trend, as a linear increase can be appreciated. During the first days of the study, clusters only were active for a mean of 2–3 days, but as epidemic evolved this duration was progressively increased. By mid-August its mean value almost reached the 7 days limit imposed on our model.

We have applied STSS to detect emerging outbreaks of COVID-19 in Spain after de-escalation. By June 21th, 9 clusters were active, located in specific regions. They represent some of the very first outbreaks that took place since then. From there, transmission spread geographically to neighboring regions during July, making the total daily clusters rise to 59 by July 31st, mostly in Ebro’s Valley but a trace can also be observed along the Mediterranean coast and Madrid. In August, when Madrid became the most likely cluster, diffusion in the regions adjacent to the capital can be appreciated, leading to a heterogeneous cluster prevalence by the end of the study period.

The main and first cluster in our analysis, that took place in the cities of Lleida and neighboring municipalities of Huesca, was strongly linked to vulnerable collectives, both meat-processing plants and agricultural industry/seasonal workers [24]. Occupational-related outbreaks were the most commonly observed up to July [6]. They have been widely reported in Europe during the 2020 summer, mainly related to food packaging and processing sectors, factories and manufacturing, office settings and health workers [25]. In addition, COVID-19 spread has been more intense between vulnerable collectives and low-income populations [26].

From mid-July, social and family outbreaks largely outnumbered occupational ones, with 40% out of the total, as transmission began to grow [6] after June 21st and the return of mobility in Spain. This increase during July coincides at the time that cluster number, mean cluster radius and mean locations per cluster reached upper bounds according to parameter limitation. As transmission grew, initial outbreak detection was no longer informative since community transmission was suspected to be occurring, which has implications in surveillance.

By August (https://​coviddifusion.​isciii.​es/​SpSumClus/​), we observed the appearance of multiple neighboring clusters. Cluster mean radius and number of included municipalities by cluster had reached a maximum, suggesting a probable widespread transmission. STSS could become a valuable tool to respond to an important question in any epidemic outbreak: when and how to determine if a community transmission might be happening. Using prospective STSS, Masrur et al. suggest that a COVID-19 community transmission might had happened in Bangladesh in March [17]. This scenario is of enormous importance for public health authorities and decision-makers, and we believe that spatial aggregation saturation of clusters over an area, once reached maximum established parameters (cluster radius fundamentally), can be understood as a relevant indicator to consider that community level transmission could be happening.

Analysis options and parameters may be adjusted depending on desired surveillance objectives. Cluster diameter may be changed to reach specific surveillance goals: in an outbreak control strategy, a smaller radius can help detect single municipalities whose RR is increasing rapidly. However, once community transmission has been detected, it may be more useful to use a larger diameter to detect an aggravation of the situation at wider regions. Moreover, time range and a minimum of included cases could be set in order to look for aggregation of different sizes. A 5 cases/7 days ratio may be a good option to detect family outbreaks in a low-density area, while 300 cases/ 7 days could be a useful tool in a decision-making panel.

Figure 2 represents, to the best of our knowledge, the first attempt to evaluate timely performance of STSS surveillance under established parameters. By mid-July, a window widening might be necessary to better capture emerging behavior since parameters already show a saturation (i.e., they reached their upper limit for the entire period). Other research teams chose to group clusters by population-at-risk included, with a minimum of 10% of the total population needed, resulting in less but wider clusters [15‐17]. Population in Spain is unequally distributed, mainly within the coastal regions and Madrid. If we had chosen this population-at-risk option, emerging clusters affecting single small municipalities would have never been detected at the beginning of the study period.

The use of surveillance records is subject to reporting bias. During low prevalence time periods, data can be expected to arrive daily in complete form. But as epidemic progresses, notification delays become present and data quality might decline, which may complicate a real-time assessment. Usual weekend bias is also accounted for.