Original data collection was conducted as part of WA DOH COVID-19 surveillance. As such, per the Office for Human Research Protections guidelines, institutional review board approval was not required [32]. As part of public health surveillance activities authorized by a public health authority, the secondary analysis of the data described in this study was excluded from the definition of research provided by the revised Common Rule [33] and did not receive approval or an exemption from an institutional review board. The researchers were granted access to the data as part of their official duties within the WA DOH.

During pilot testing, the statewide model identified clusters that were large in radius and size, which did not meet the goal of actionable clusters. Independent models for each LHJ with the same parameters produced smaller clusters, which did meet that goal. Therefore, our final design included LHJ-specific models that were combined to provide statewide results.

Next, we compared daily prospective and weekly retrospective models. There were 12 significant clusters identified by the daily prospective model; 10 (83%) of which were also identified by the weekly retrospective model. While the daily prospective used more recent data, the model required more resources to produce and review. Given the overlap between the model output and resource constraints, we selected the weekly retrospective analysis for the final model.

In comparing 1-, 2-, and 3-day lag periods, we found about half (47/90, 52%) identified similar clusters regardless of lag. The remainder identified similar clusters across the 2- and 3-day lag. We chose a 2-day lag to maximize timeliness.

The initial model, which aggregated cases to census tracts, produced clusters that were too large geographically for public health action. Using the smaller unit of census block group was favorable based on heuristic criteria.

Lastly, we tested various maximum cluster radii. The 5-km model identified the smallest clusters (median radius 3.3, range 0-5, IQR 1.5-4.5 km), followed by the 10-km model (median radius 5.2, range 0-10, IQR 2.3-8.3 km), and then the 20-km model (median radius 6.1, range 0-19.5, IQR 2.9-11.3 km). The 20-km radius was determined to capture both smaller clusters in urban areas and larger clusters in rural areas, and thus was selected for the final model.

From July 17 to December 17, 2021, there were 341,505 COVID-19 cases reported among Washington State residents. Of these, 316,642 (92.72%) had a geocoded residential address. A median of 41,217 cases was included in each weekly SaTScan analysis (range 24,996-61,069, IQR 30,863-52,842); a total of 4659 clusters were identified (weekly median 236, range 197-275, IQR 218-248). Clusters included 98,172 unique cases (28.74% of reported COVID-19 cases); most (2874/4659, 61.68%) were new clusters. The remaining analyses are limited to new clusters.

Among the 2874 new clusters, 887 (30.86%) occurred in large LHJs, 890 (30.96%) in medium LHJs, 734 (25.54%) in small LHJs, and 363 (12.63%) in rural LHJs (Tables 1 and 2). Overall, the median cluster size was 15 cases (range 2-1045, IQR 8-29) and significantly differed across population groups (large P<.001, medium P<.001, small P<.001, and rural P<.001). Rural LHJs had the smallest median cluster size (4 cases, range 2-28, IQR 3-6) and large LHJs the largest (27 cases, range 5-1045, IQR 16-49). The median number of weekly clusters per LHJ also differed significantly: rural LHJs had the least (2, range 1-6, IQR 1-3; P<.001), followed by small (3, range 1-9, IQR 2-4; P<.001), medium (5, range 1-12, IQR 3-6; P<.001), and large LHJs (9, range 2-21, IQR 6-11; P<.001). Overall, the median cluster radius was 2.4 (IQR 0.9-6.88) km. The radius significantly differed for medium (2.5 km; P=.045), small (3.8 km; P<.001), and rural LHJs (0 km; P<.001). Of the 2874 clusters, 486 (16.91%) were statistically significant in the weekly analysis. Rural LHJs had a significantly smaller percentage of significant clusters (37/363, 10.2%; P<.001). There was no significant difference in duration (overall 4 days, range 1-7) or timeliness (overall 1674/2874, 58.24%) by the LHJ population group.

The SPI was significantly positively correlated with the weekly number of cluster cases among large (r=.82; P<.001), medium (r=0.76; P=.01), and small (r=0.73; P=.01) LHJs and with cluster size among large (r=0.73; P=.01) and medium LHJs (r=0.82; P<.001; Figure 1). There were few nonsignificant moderate correlations between SPI and median cluster size, number of weekly clusters, and percentage of reported cases associated with clusters. There was no correlation between the timeliness of clusters and SPI (large r=–0.19, medium r=–0.22, small r=–0.12, and rural r=–0.42).

There were 2874 reported outbreaks from July 17 to December 17, 2021, and 363 (12.63%) matched to ≥1 SaTScan cluster (Table 3). Among reported outbreaks, the most common settings were schools and youth programs (825/2874, 28.71%), workplaces (617/2874, 21.47%), and long-term care facilities (541/2874, 18.82%). Among the 363 matched outbreaks, the most common settings were also schools and youth programs (108/363, 29.8%), long-term care facilities (99/363, 27.3%), and workplaces (56/363, 15.4%), as well as congregate housing (44/363, 12.1%). Of the 2874 reported outbreaks, the settings with the largest percentage of reported outbreaks that matched clusters were community settings (16/72, 22%), congregate housing (44/212, 20.8%), long-term care facilities (99/541, 18.3%), and school and youth programs (108/825, 13.1%). Settings with <10% of matched outbreaks were workplaces (56/617, 9.1%), food and beverage establishments (10/113, 8.8%), inpatient (6/77, 8%), outpatient (8/104, 7.7%), colleges (1/16, 6%), and retail settings (14/287, 4.9%).

The median duration among all reported outbreaks was 8 days (range 1-151, IQR 4-14) and 10 days (range 1-151, IQR 5-16) among matched outbreaks. Among all reported outbreaks, the median size was 4 cases (range 2-256, IQR 2-7) and among matched outbreaks, it was 6 cases (range 2-232, IQR 3-14).

Outbreaks in community settings (16/72, 22%; P=.01), congregate housing (44/212, 20.8%; P<.001), and long-term care facilities (99/541, 18.3%; P<.001) were more likely to match to clusters than other settings, while outbreaks in workplaces (56/617, 9.1%; P=.003) and retail settings (14/287, 4.9%; P<.001) were less likely to match. For one-third (119/363, 32.8%) of matched outbreaks, all cases were associated with the outbreak event in the surveillance system after the SaTScan cluster was identified. This resulted in a “report date” later than the “analysis date.” Of the 363 matched outbreaks, those in workplaces (24/56, 43%; P<.001) and schools and youth program settings (42/108, 38.9%; P<.001) were more likely to have report dates later than analysis dates, while outbreaks in community settings (4/16, 25%; P<.001), long-term care facilities (20/99, 20%; P<.001), and congregate housing (8/44, 18.2%; P<.001) were less likely (Table 3).

A reported outbreak could match with more than one cluster and vice versa. The 363 outbreaks matched 349 SaTScan clusters, which resulted in 384 unique outbreak-cluster pairs. Matched outbreaks had a median size of 6 (IQR 3-14) cases and a duration of 10 (IQR 5-16) days. Matched clusters had a median size of 17 (IQR 12-34) cases and a duration of 5 (IQR 3-6) days. The median number of linked cases (those common to the outbreak and cluster in a matched pair) was 3 (range 1-58, IQR 1-6). This represents a median of 40% (IQR 26.4%-60%) of outbreak cases that were part of the cluster and 12.5% (IQR 8.3%-25%) of cluster cases that were part of the outbreak. The median JS was 10% (IQR 7.2%-17.8%; Tables 4 and 5). For an illustrative example, an 8-case workplace outbreak matched a 29-case cluster by 8 linked cases, with 100% (8/8) of outbreak cases in the cluster and 28% (8/29) of cluster cases in the outbreak, resulting in JS was 28%.

During a time when LHJs were heavily burdened with case and outbreak investigations, WA DOH initiated systematic statewide cluster detection for COVID-19 using SaTScan. SaTScan parameters can be adjusted based on study objectives, data availability, and disease and population characteristics [26,34]. Washington State required a model tailored to its overtaxed information systems and decentralized public health system, which serves almost 8 million residents in 39 diverse counties. We collaborated with LHJs to implement a space-time permutation model that maximized real-time surveillance data and identified opportunities for public health intervention.

We assessed a daily prospective model, which would have provided the most real-time surveillance; however, it did not produce substantially different results than a weekly retrospective model. Considering the time and resource constraints faced by LHJs, we determined the weekly retrospective model was the most appropriate. Despite having parameters unique from other studies [11,22,23], the model worked well in Washington and identified clusters early enough for intervention. It is important to identify outbreaks as quickly as possible (when there are fewer cases) to prevent further transmission and illness. While outbreak size may vary due to population and disease transmission levels, our analysis showed that 58.24% (1674/2874) of clusters were timely regardless of population or incidence. By identifying clusters promptly, LHJs can conduct targeted investigations, implement containment measures quickly, and prioritize resource distribution to areas with the greatest need.

The number of weekly clusters and cluster size were both correlated with population; more and larger clusters were identified in highly populated urban areas which generally have larger health departments with greater epidemiologic capacity [35]. These data support the use of a systematic statewide approach as it benefits LHJs with varying populations and resource capacity, ensuring that all LHJs, regardless of resources, can effectively manage outbreaks by leveraging a standardized, statewide surveillance system.

Space-time scan statistics have successfully identified respiratory disease outbreaks [13,23,36], including COVID-19 [37]. To determine if the model could identify true outbreaks, we compared clusters with reported outbreaks. Generally, we did not expect a high degree of alignment between reported outbreaks and SaTScan clusters, given SaTScan was intended to fill gaps in current outbreak detection methods. Given the use of residential address as the geographic input for the model, outbreaks in congregate living settings (eg, long-term care facilities or corrections) had a high cluster matching rate as expected. Outbreaks that occur further from the home, such as in workplaces, would likely be harder for the model to detect.

During the COVID-19 pandemic, several settings have been a high priority and focus for public health interventions, including workplaces, community settings, and schools. Workplaces, the second most common setting among reported outbreaks in Washington State, play a central role in the COVID-19 community transmission [38-40] and have been subject to safety prevention requirements throughout the pandemic [41] to reduce the spread of the disease. As expected, workplace outbreaks matched fewer clusters than all reported outbreaks. Incorporating workplace addresses may have helped, however occupational and industry data for COVID-19 are largely incomplete and often not standardized, which may create barriers to systematically identifying workplace outbreaks [39,42]. Efforts to collect and standardize these data and improvement of occupational health surveillance may address this [38].

Community settings have also been a high priority for public health interventions. In fall 2021, when SaTScan was first implemented in Washington State, many efforts were made to implement safety precautions, including limiting capacity at events and gatherings [43], masking [44], and vaccination requirements at large events [45]. Outbreaks among community settings made up only 2.51% (72/2874) of all reported outbreaks, yet 22.2% (16/72) of these outbreaks were linked to clusters, suggesting SaTScan might be a useful method for systematically detecting these types of outbreaks.

The model performed moderately well for schools and youth programs, another high priority and focus of public health interventions. For the 2021-2022 school year in Washington State, public health officials collaborated on guidance to allow students to return safely to in-person instruction [46]. Given their residential proximity, we expected a high cluster match rate for educational settings. Schools and youth programs were the most common setting among reported outbreaks as well as matched outbreaks. However, only 13.1% (108/825) of school outbreaks matched a cluster compared to 12.63% (363/2874) overall, arguably leaving room for improvement. The model could be further tailed to capture outbreaks in educational settings by incorporating school district-level models or including age as an input variable.

Lastly, about a third of outbreaks that matched clusters had a report date in the surveillance system that was later than when the cluster was identified by the model. This suggested that the model might be able to identify outbreaks earlier than existing outbreak detection methods. Detecting outbreaks sooner allows for quicker implementation of containment measures, reducing the spread of the disease and minimizing the impact on communities. By automating the detection process, local public health workers can spend more time on the outbreak intervention measures and less on the initial outbreak detection and investigation work.

The findings in this report are subject to several limitations. First, the model does not account for irregular geographics, such as major waterways. Second, residential address was used to determine outbreak detection as this was the most complete and available. However, residential addresses may not be the most representative of where exposure occurs and may limit the identification of clusters where transmissions happened away from home. Third, reported outbreaks are identified through many indicators and sources that vary by LHJ, contain known biases including lags or gaps during surges [6-9], and likely underestimate actual outbreaks [47,48]. Fourth, because the final model was run independently for each LHJ, cross-jurisdictional clusters were not identified and thus not matched with similar outbreaks. Lastly, as expected, relatively few SaTScan clusters matched outbreaks, and without further investigation to identify epidemiologic or genomic linkages, we cannot determine if unmatched clusters are true outbreaks.

Our goal was to implement a statewide systematic cluster detection process to identify COVID-19 clusters. WA DOH successfully developed and implemented a SaTScan space-time permutation model that met these goals within the state’s unique structure and systems. Regardless of the LHJ population, the model identified reasonably sized, timely clusters for investigation and resource prioritization. Cluster size increased with incidence but likely remained actionable even for smaller LHJs. Among reported outbreaks that matched clusters, the model performed well in congregate living settings, schools, and events, which have been high-priority settings throughout the COVID-19 pandemic. There is an opportunity to tailor the model to further improve cluster detection in some settings, such as workplaces and schools. Evidence suggests running a weekly model might identify some outbreaks sooner than existing outbreak detection methods, particularly in workplace outbreaks where data collection remains a challenge. In summary, our SaTScan model was able to identify timely, actionable clusters, especially in high-priority settings, which can serve to support outbreak detection to reduce further COVID-19 transmission.