Examining the accessibility levels to healthcare can help policymakers improve public health and promote quality of life  Urban geographers and planners have examined the geographical distribution of supply and demand and the role of traffic conditions in that distribution [2, 3] to better evaluate accessibility to healthcare. Although the three factors (i.e., demand, supply, and traffic condition) are temporally not fixed [4–8], the temporal variations have been given relatively less attention in healthcare accessibility research. Furthermore, given that an accessibility model that does not consider the temporal variations is more likely to overestimate (or underestimate) accessibility levels , which mislead the wise policy decision of healthcare planners, measuring the realistic spatial accessibility of healthcare is crucial for monitoring the effectiveness of a healthcare system and ascertaining policy measures that minimize spatial disparities .
As an effort to measure reasonable place-based accessibility of healthcare, the development of the two-step catchment area (2SFCA) method has facilitated the application of 2SFCA-based methods to healthcare accessibility research. Many studies have further attempted to enhance its ability of demand, supply, and traffic conditions. For example, the three-step FCA  and the conditional logit model-based FCA  accounted for the demand side, the modified 2SFCA , Seoul enhanced 2SFCA , hierarchical 2SFCA [15, 16], and grid-to-level 2SFCA [15, 16] tackled supply dimension, and the enhanced 2SFCA  and multi-modal 2SFCA  considered traffic conditions in the models.
While studies that use person-based accessibility have considered the temporal variations [19, 20], it is only recently that studies on place-based accessibility have attempted to include the temporal variations in accessibility measures. This is because measuring place-based spatiotemporal accessibility with the variations requires large-scale data on demand, supply, and traffic conditions between locations. Recent advancements in big data and enhanced computing power have made the examination of place-based spatiotemporal accessibility more feasible  with various applications, such as, jobs [e.g., 21,22], grocery stores [e.g., 23,24], electric vehicle charges [e.g., 25], and healthcare [e.g., 7–9,26,27]. However, there are very few studies that consider the three factors simultaneously [23, 25].
In healthcare accessibility research that uses the 2SFCA-based method, considering the temporal variations of all the three factors is critical as the accessibility outcome achieved from the 2SFCA-method is determined by demand, supply, and catchment area by distance/time. Furthermore, in reality, the geographical distribution of population varies owing to shifts in activities ; the availability of healthcare services further depends on opening hours [7, 23], and varying traffic conditions by congestions [4–6]. In this regard, the static accessibility model that ignores the temporal variations of demand, supply, and traffic conditions is likely to fail to estimate the temporal changes in geographical accessibility patterns.
However, to the best our knowledge, there is no study that considers all the three factors in the place-based accessibility of healthcare. Moreover, previous studies on healthcare accessibility have considered one or two of the three factors. For example, Xia et al.  evaluated real-time population distribution using GPS trajectory data but used travel time based on the Euclidean distance and constant travel speed. Hu et al.  examined the effect of traffic congestion on the spatial accessibility of emergency medical services in China without considering temporal changes in demand and supply. Zhang et al.  measured the walk accessibility of older adults to general practitioners in London, United Kingdom, and evaluated vertical equity between older adults and other age groups. However, they only considered a temporal variation of opening hours. Regarding the nature of the 2SFCA-based method, researchers need to consider the three factors simultaneously. Moreover, given that accessibility levels can be overestimated (or underestimated) when ignoring temporal variations [4, 9], measuring accessibility more realistically is crucial for helping healthcare planners make wise decisions.
Accordingly, this study aims to (1) explore the areas with high or low levels of time-varying accessibility and (2) identify spatial discrepancy between time-varying and static accessibility. We further use the generalized 2SFCA (G2SFCA) method with de facto population from hourly-collected mobile-based data, available healthcare facilities, and actual traffic conditions in Seoul, South Korea, to evaluate the time-varying accessibility of primary healthcare services. To achieve these goals, time-varying accessibility patterns are visualized in relative manners. Areas in which the two accessibility models differ significantly are identified using bivariate Local Moran’s I statistic. Therefore, this study contributes to capturing spatiotemporal variations in the accessibility to primary healthcare and better knowledge for the decision making of healthcare planners.
This paper proceeds as follows. The next section describes the study area, data collection, and methodology, including an accessibility measure and bivariate Local Moran’s I. We further present and compare the accessibility results of the static and time-varying models. In conclusion, we discuss our findings and suggest policy implications and limitations to be addressed in future research.
Data And Study Settings
We conducted this study in Seoul, Korea’s capital city, to demonstrate the temporal fluctuations and effects of three factors on the spatial accessibility of primary healthcare services. Seoul is the most densely populated city worldwide (9.64 million persons and 605 km² in 2019) and a commuting and shopping destination. In 2021, the daily average population influx from other cities was approximately 1.6 times the number of residents. This influx means that the demand for healthcare services can change over time even within the same zone. Furthermore, compared with other cities, Seoul has significant traffic congestion, with an average commuting time of 44.7 minutes in 2021.
Owing to variations in de facto population and traffic congestion, Seoul was the ideal location for this case study because (1) data about the three factors (i.e., de facto population, operation schedules of primary healthcare facilities, and travel time based on traffic congestion) are available, (2) a temporal variation in spatial accessibility by location is easy to capture owing to the high density of primary healthcare services, and (3) the greenbelt (i.e., restricted development zones) restrains expansion around Seoul, reducing the edge effect for spatial accessibility of primary healthcare facilities on the outskirts of the city (see Fig. 1).
We collected data from weekdays in April 2021, a month with no national holidays or national events, to minimize fluctuations in demand, supply, and traffic conditions from one day to the next. Additionally, compared with other months, April 2021 had less rain; this can affect traffic conditions. Weekdays included Monday through Thursday; we excluded Fridays owing to the possibility of distinct commuting and traffic patterns.
We collected data about each factor from several sources. We calculated the de facto population of Seoul by the hour using location information from mobile phones (Seoul Open Data Plaza), based on a spatial unit known as a jipgyegu (hereafter, a census block), which is the equivalent of a census block in the United States. On average, the de facto population was lowest at 3 a.m. (10,214,039 people, M = 533.29, SD = 541.23) and highest at 3 p.m. (10,827,023 people, M = 565.30, SD = 1446.73), including commuters or shoppers arriving from surrounding cities. Moreover, some areas were more concentrated than others.
We used Kakao and Naver, web-based map services in Korea, to collect supply data. This study defines “primary healthcare services,” in consideration of the context of the healthcare delivery system in Korea. In Korea, there are many medical specialists who do not belong to large hospitals and can easily open their own private hospitals and clinics and public health centers. Similar to primary healthcare services provided by general practitioners in the United Kingdom or family physicians in the United States, these private hospitals and clinics provide services for disease prevention and treatment of mild diseases (e.g., colds). Unlike in the United Kingdom, where patients make an appointment to see their doctor beforehand, patients in Korea can visit nearby hospitals for mild treatment or vaccination at any time without booking an advance appointment. In this context, it is important to measure time-varying access to healthcare services based on the de facto population rather than the home-based population.
We therefore considered clinics or hospitals that had 30 or fewer beds and offered medical services in any of the following areas: internal medicine, pediatrics, otolaryngology, and family medicine. Based on the definition, we split Seoul into grids and sought facilities within 1,000 meters of the centroid of each grid cell by iterating all grids using the Kakao map API, to further collect information about primary healthcare facilities (e.g., X and Y coordinates, name, and address). However, because the Kakao map API does not return operating schedules for facilities, we collected schedule information from Naver map, using the name and address of information gathered from Kakao map API; this procedure yielded 2,827 primary healthcare facilities.
We used the road speeds provided by Seoul’s Transport Operation & Information Services (TOPIS), a traffic control agency, to estimate traffic conditions. TOPIS collects traffic information from many agencies involved in transportation in Seoul, including the bus management system, the public transit card system, and the Korea Expressway Corporation. We integrated TOPIS data that include daily and hourly speed information with road network data from the Korea Transport Database (KTDB). For roads missing speed information, we inputted the average speed of roads in Seoul into the roads, based on each time period, then calculated travel time (i.e., length of the road link divided by speed and multiplied by 60). To calculate travel times between census blocks and primary healthcare facilities, we used the OD cost matrix function of the network analyst in ArcMap 10.7.1 via Python’s arcpy module.
Because the extent of a change in accessibility may not be significantly large by an hour, we considered five time periods (i.e., 8 a.m., 10 a.m., 3 p.m., 6 p.m., and 8 p.m.) based on variations in population, traffic condition, and hospitals between 8 a.m. and 9 p.m., as primary healthcare services in Seoul do not operate 24 hours (Fig. 2). At 8 a.m., there exists 555 persons per census block on average (SD = 1,057), approximately 2.37 min/km of the average travel time (SD = 0.44), and 446 hospitals, which account for 15.8% of the total. At 10 a.m., there are 562 persons per census block (SD = 1,338), approximately 2.67 min/km of the average travel time (SD = 0.54), and 2,799 hospitals, which account for 99% of the total and is similar at 3 p.m. At 6 p.m., that is the afternoon rush hour, the average population decreases to 557 persons (SD = 1,120), the travel time is as high as 2.9 min/km (SD = 0.63), and operating hospitals also decrease to 48% of the total. Lastly, at 8 p.m., there are 546 persons per census block (SD = 856), 2.78 min/km travel time on average (SD = 0.57), and 73 hospitals, which account for 2.6% of the total.
The three components spatially vary by time as well. For example, the population density increases around CBD and GBD from morning to afternoon, then decreases after 6 p.m. The traffic congestion is worse in the afternoon rush hour than in the morning rush hour. Hospitals are less open at 8 a.m. and 8 p.m., while almost 99% of hospitals operate at 10 a.m. and 3 p.m.