A spatial-temporal analysis of organ-specific lupus flares in relation to atmospheric variables and fine particulate matter pollution

Abstract

Objectives:

We performed a spatial-time cluster analysis of the Hopkins Lupus Cohort with the goal of identifying potential clusters of SLE organ specific flares and their relation to temperature changes and fine particulate matter pollution (PM2.5).

Methods:

1261 patients who fulfilled the SLICC classification criteria for SLE and who had recorded home addresses were included in the analysis. Disease activity was expressed as the Lupus Activity Index. Assessment of rash, joint involvement, serositis, neurologic, pulmonary, renal, and hematologic activity was quantified on a 0–3 visual analogue scale (VAS). An organ specific flare was defined as an increase in VAS of 1 point or more compared to the previous visit. Spatiotemporal cluster detection was conducted using the SaTScan software. Regression models were used for cluster adjustment and included individual, county-level, as well as environmental variables.

Results:

Three statistically significant (p<0.05) clusters unadjusted for environmental variables were identified for joint flares, four rash flare clusters (p<0.05), three hematologic flare clusters (p<0.05), two neurologic flare clusters (p<0.05), four renal flare clusters (p<0.001), two serositis clusters (p<0.001), and two pulmonary flare clusters (p<0.001). The majority of the identified clusters changed in significance, temporal, or spatial extent after adjusting for environmental variables.

Conclusions:

We describe the first space-time clusters of lupus organ-specific disease activity. Seasonal, as well as multi-year cluster patterns were identified, differing in extent and location for the various organ-specific flare types. Further study focusing on each individual lupus organ-specific activity will be required to better understand the driving forces behind these observed changes.

INTRODUCTION

Systemic lupus erythematosus (SLE) is a complex multisystem autoimmune disease with strong epidemiologic evidence of association with several environmental factors, including crystalline silica exposure¹, cigarette smoking², and exogenous estrogens (oral contraceptives and postmenopausal hormones)³ as well as potential associations between other exogenous factors such as mercury⁴, ultraviolet radiation, solvents, and pesticides⁵.

When it comes to atmospheric impact, we previously described significant seasonal variation in SLE disease activity with more arthritis activity in the spring and summer months, and an increase in renal activity in winter months, significantly higher anti-dsDNA antibody titers in the fall, and a significant variation of global disease activity as measured by SELENA -SLEDAI through the year⁶. In a cohort of 2802 SLE patients from China, the absolute number of patients with active SLE (SLEDAI>12) in a month was positively correlated with the amount of precipitation and wind speed⁷. There was no significant correlation between average temperature, average humidity, and average percentage of sunshine hours and SLE activity.

Fine particulate matter pollution (PM2.5) averaged for up to 10 days prior to the patient visit was associated with anti-dsDNA and cellular casts but not with global disease activity in a Montreal lupus cohort⁸. A population-based cohort study from Taiwan showed a positive association between a 10.2 μg/m³ increase in fine particulate matter concentration and new diagnoses of SLE⁹. Similarly, population based studies from Alberta and Quebec showed that PM2.5 exposure may be associated with an increased risk of systemic autoimmune diseases, including SLE¹⁰.

We pursued the development of spatial-temporal analytical models of lupus flares with the goal of identifying potential flare clusters and their relationship to fine particulate matter pollution and temperature changes. These spatial-temporal models serve as the foundation for a novel approach to the study of environmental factors in systemic lupus.

METHODS

Patients and Activity Indices.

As previously described¹¹, the Hopkins Lupus Cohort is a prospective cohort study of predictors of lupus flare, atherosclerosis, and health status in SLE. The study cohort includes all patients at the Hopkins Lupus Center who have a clinical diagnosis of SLE and meet classification criteria for SLE¹². All have given informed written consent to participate in the study. Subjects enrolled in the cohort are followed quarterly or more frequently if clinically necessary. The clinical features, laboratory testing, and damage accrual data are recorded at the time of entry into the cohort and are updated at subsequent visits. The Hopkins Lupus Cohort has been approved by the Johns Hopkins University School of Medicine Institutional Review Board and complies with the Health Insurance Portability and Accountability Act.

The Hopkins Lupus Cohort included 2486 patients with the vast majority of patients living in Maryland and the surrounding states. A 350-kilometer radial buffer around the Johns Hopkins Lupus Center was therefore considered as the study area, since it included a high and consistent density of patients necessary for cluster detection. This area included most of Maryland, Delaware, and District of Columbia, as well as parts of Pennsylvania, New Jersey, Virginia, and West Virginia (Figure 1).

Figure 1.

Johns Hopkins Lupus Cohort Study Area (1999–2017)

Patients who had home addresses within the study area between 1999 and 2017 were included in the study. A 1999 cutoff was selected since consistent PM2.5 data was available only starting from that year. All address changes during study follow up were recorded and were considered. Disease activity was expressed as the Lupus Activity Index¹³. Assessment of rash, joint involvement, serositis, neurologic, pulmonary, renal, and hematologic activity was quantified on a 0–3 visual analogue scale (VAS). An organ specific flare was defined as an increase in VAS of 1 point or more compared to the previous visit. For the increase to count as a flare, the previous visit must have been within 110 days of the current visit. Patients were generally asked to check in at least every three months, however, this number varied, and 92 day, 100 day, and 110 day cutoffs were considered. A 110 day cutoff was considered the most appropriate, since it allowed retention of 70% of the records while still excluding patients that visited too rarely to determine whether a flare had occurred. After applying all the criteria described above, 1628 patients making a total of 29,677 visits were included in the study. Individual variables considered in the analysis included patient sex, age, ethnicity, smoking status, household income, years of education, and urban vs. rural living environment. County level variables included median income and proportion of black population. All address changes during follow up were taken into account. Flare types included rash, joint, serositis, neurologic, renal, pulmonary, and hematologic. Available data are further summarized in Table 1.

Table 1:

Patient and county-level variables as well as lupus flare outcomes summarized by patient gender

	Females (%)	Males (%)	Total (%)
Patients	1504	124	1628
Clinic Visits	27376	2301	29677
Average Age	38.8	43.2	39.1
Race (Black)	618 (41.1)	41 (33.1)	659 (40.5)
Race (White)	761 (50.6)	74 (59.7)	835 (51.3)
Race (Other)	125 (8.3)	9 (7.3)	134 (8.2)
Years of Education	14.3	14.2	14.3
Mean Personal Income	65491.2	77642.5	66322.2
Smoking Status	157 (10.4)	9 (7.3)	166 (10.2)
Urban Living	1281 (85.2)	101 (81.5)	1382 (84.9)
Average County Income	63655.7	65802.7	63823.7
Average County Proportion of Black Population	29.9	27	29.7
Rash Flares	1146 (4.2)	61 (2.7)	1207 (4.1)
Joint Flares	1665 (6.1)	102 (4.4)	1767 (6)
Serologic Flares	470 (1.7)	25 (1.1)	495 (1.7)
Neurologic Flares	292 (1.1)	22 (1)	314 (1.1)
Renal Flares	1696 (6.2)	188 (8.2)	1884 (6.3)
Pulmonary Flares	58 (0.2)	3 (0.1)	61 (0.2)
Hematological Flares	446 (1.6)	62 (2.7)	508 (1.7)

Daily fine particulate matter pollution (PM2.5) data measured in micrograms per cubic meter, temperature measured in degrees Fahrenheit, ozone concentration measured in parts per million (ppm), resultant wind measured in miles per hour, relative humidity expressed as a percentage, and barometric pressure expressed as millimeters of Mercury (mmHg) were collected at various monitoring stations in the eastern United States and obtained from the Environmental Protection Agency. Ordinary Kriging was used via the “gstat” R package¹⁴ to predict the 10-day average level of the environmental variables for each patient prior to each visit date. Ordinary Kriging¹⁵ is a widely used statistical interpolation method that allows to predict the value of a variable in any geographic location based on known variable measures at other locations, while taking into account the spatial dependence in the distribution of the measured variable. This method allowed to predict environmental exposures of each patient based on appropriately weighted exposure measures at the surrounding monitoring stations.

Both univariate and multivariate Generalized Estimating Equations (GEE) logistic regression models with an exchangeable correlation structure were built to study the association of individual (age, sex, race, smoking status, household income, years of education, urban living status), county (county income, proportion of black population), and environmental (PM2.5, temperature, ozone, resultant wind, barometric pressure, relative humidity) variables with the seven different outlined types of lupus disease activity. GEE regression was used in order to account for the repeated measures each patient was subject to during their multiple clinic visits, thus violating the independence assumption necessary for ordinary regression.

Spatiotemporal cluster detection of flares was conducted using SaTScan v.9.4.4 software. In order to detect such clusters, SaTScan utilizes a moving window of variable size, that centers at each data point location, and considers all possible time intervals, recording the number of observed and expected cases inside and outside the window. For each window location, size, and time, the observed and expected cases are compared, the likelihood function is maximized, identifying the window that is least likely to occur by chance, and this process is repeated 999 times through Monte Carlo hypothesis testing in order to obtain a p-value, and identify the window as a statistically significant cluster. The method used to calculate expected cases depends on the selected statistical model. Patient data were aggregated to 66 counties that spanned the study area and considered as the spatial units for the analysis. A discrete Poisson SaTScan model was utilized, where the total number of each type of lupus flares per day, in each county, is considered to be the case, and the total number of patient visits per day, in each county, is considered as the population, if the cluster is unadjusted. In order to determine whether individual, county, or environmental covariates help explain the identified clusters, the expected number of flares determined from the GEE logistic regressions discussed above can be used in place of the population within the SaTScan Poisson model. In this study, two adjusted spatiotemporal cluster analyses were conducted. In the first, only adjustments for individual and county level socioeconomic factors were made. In the second, in addition to the individual and county variables, adjustments based on environmental variables were also included. One month long minimum time intervals were considered for this analysis, and spatially overlapping clusters were allowed as long as the overlapping cluster did not contain the centroid of the cluster that was already there.

Data availability.

The data that support the findings of this study are available on request from the corresponding author G.S. The data are not publicly available as they contain information that could compromise research participant privacy.

RESULTS

In a univariate regression analysis, rash (OR = 1.029) and joints (OR = 1.026) flares were found to be positively associated (p<0.05) with PM2.5 exposure, while serositis (OR = 1.024) was found to be marginally (p=0.053) associated. Rash (OR = 1.065), joints (OR = 1.047), and hematologic (OR = 1.095) flares were found to be significantly positively associated with temperature, while renal flares (OR = 0.960) were found to be marginally (p=0.072) negatively associated with temperature. Ozone concentration was associated (p<0.05) with rash (OR=1.013), and negatively associated with renal flares (OR=0.992). Resultant wind was positively associated (p<0.05) with joint (OR=1.039), neurologic (OR=1.099), renal (OR=1.028), and pulmonary (OR=1.135) flares. Relative humidity was associated (p<0.05) with joints (OR=1.163), and marginally (p=0.077) associated with neurologic (OR=1.099) flares. No significant associations were found for barometric pressure. Furthermore, rash, serositis, and renal flares were found to be negatively associated with age (p<0.05). Rash and serositis symptoms were found to be more likely in African American rather than white populations, serositis being marginally so (p=0.092), and joint flare odds were higher in African American rather than in other non-white ethnicity categories. Rash flares were found to be more likely in smokers (OR = 1.862) and be negatively associated (p<0.05) with years of education (OR = 0.943) along with renal flares (OR=0.943). Finally, joint (OR = 0.969) and renal (OR = 0.964) flares were found to be negatively associated (p<0.05) with median county income. These findings are summarized in Table 2.

Table 2:

Odds ratios from univariate regressions of the outcome variables with individual, county, and environmental covariates. Statistically significant values (p<0.05) are bolded, marginally significant values (p<0.1) are underlined. The p-values presented are Bonferroni-adjusted.

Variable	Rash		Joints		Serositis		Neurologic		Renal		Pulmonary		Hematologic
	Odds	P-value	Odds	P-value	Odds	P-value	Odds	P-value	Odds	P-value	Odds	P-value	Odds	P-value
	Ratio		Ratio		Ratio		Ratio		Ratio		Ratio		Ratio
Age	0.989	0.021	1.000	> 0.9	0.968	< 0.001	0.997	> 0.9	0.976	< 0.001	0.993	> 0.9	0.991	> 0.9
Sex	0.620	0.5115	0.721	0.7095	0.683	> 0.9	1.127	> 0.9	1.341	0.3225	0.640	> 0.9	1.822	0.3705
Race (Other)	0.671	> 0.9	0.571	0.0075	0.764	> 0.9	0.448	0.6885	0.924	> 0.9	1.450	> 0.9	1.037	> 0.9
Race (White)	0.668	< 0.001	0.875	0.7965	0.658	0.0915	0.734	> 0.9	0.536	< 0.001	0.574	> 0.9	0.986	> 0.9
Smoking Status	1.862	< 0.001	1.201	0.8745	0.499	0.8475	0.990	> 0.9	0.957	> 0.9	1.166	> 0.9	1.023	> 0.9
Personal Income**	0.993	> 0.9	0.989	0.153	0.995	> 0.9	0.998	> 0.9	0.998	> 0.9	1.009	> 0.9	1.004	> 0.9
Years of Education	0.943	0.012	0.972	0.4605	0.943	0.498	1.007	> 0.9	0.943	< 0.001	0.980	> 0.9	1.007	> 0.9
Urban Living	0.975	> 0.9	0.929	> 0.9	0.954	> 0.9	0.812	> 0.9	0.840	0.9225	0.936	> 0.9	0.968	> 0.9
County Proportion of Black Population*	1.044	0.438	1.006	> 0.9	0.959	> 0.9	1.005	> 0.9	1.034	0.6795	0.989	> 0.9	1.025	> 0.9
Median County Income**	0.980	> 0.9	0.969	0.0015	0.994	> 0.9	0.949	0.144	0.964	0.0015	0.908	0.153	0.985	> 0.9
PM2.5	1.029	< 0.001	1.026	< 0.001	1.028	0.0525	1.026	0.399	1.004	> 0.9	1.043	0.8625	1.025	0.138
Temperature*	1.065	0.033	1.047	0.0285	1.061	0.6135	1.068	> 0.9	0.960	0.072	1.108	> 0.9	1.095	0.027
Ozone	1.013	< 0.001	1.004	> 0.9	1.011	0.3255	1.002	> 0.9	0.992	0.0105	1.010	> 0.9	1.005	> 0.9
Resultant Wind*	1.009	> 0.9	1.039	< 0.001	1.007	> 0.9	1.099	< 0.001	1.028	0.0165	1.135	0.0045	1.046	0.147
Barometric Pressure*	0.993	0.5655	1.004	> 0.9	1.006	> 0.9	0.989	> 0.9	0.998	> 0.9	1.053	0.6585	1.007	> 0.9
Relative Humidity*	0.991	> 0.9	1.163	< 0.001	1.125	> 0.9	1.297	0.0765	1.045	> 0.9	1.293	> 0.9	1.159	0.1365

^*indicates measures per 10 percent/units of measurement

^**indicates measures per $5,000

For most outcomes, identified clusters changed spatially or temporally after considering environmental variables in addition to county and individual ones. The general interpretation of cluster behavior after adjusting for environmental covariate can be classified into three categories:

1. Clusters that remain unchanged temporally or spatially after adjusting for environmental covariates indicate areas where no association between environmental covariates and flare occurrence was identified.

2. Clusters that disappear or decrease spatially or temporally after adjusting for environmental covariates indicate areas where there were high levels of covariates along with a general positive association of the covariates with the number of flares, or low levels of covariates with a general negative association.

3. Clusters that emerge or increase spatially or temporally after adjusting for environmental covariates indicate areas where there were low levels of covariates along with a general positive association of the covariates with the number of flares, or high levels of covariates with a general negative association.

Three statistically significant (p<0.05) environmentally unadjusted clusters were identified for the joint flares. One encompassed most of Maryland’s eastern shore and Delaware, ranging from July 2001 to July 2005, the second one centered around the City of Baltimore and Baltimore County ranging from April 2001 to July 2009, and the third one to the southwest, including Washington DC and parts of Virginia and Maryland, ranging from January 2002 to September 2005. After adjusting for environmental variables, cluster 1 located on the Eastern Shore decreased temporally, cluster 2 in the west remained unchanged, and cluster 3 decreased spatially and shifted temporally. A new cluster (cluster 4) appeared in the north-east after adjustment (Figure 2).

Figure 2.

Top – Joint flare clusters adjusted for individual variables (age, sex, race, smoking status, personal income, years of education, urban living status) and county variables (county income, proportion of black population). Bottom – Joint flare clusters adjusted for individual and county variables as well as environmental variables (PM2.5, temperature, ozone, resultant wind, barometric pressure, relative humidity)

After adjusting for individual and county variables, four statistically significant (p<0.05) rash flare clusters were identified in eastern Maryland, south-eastern Pennsylvania, and Delaware. Clusters 1 and 2 located in the Eastern Shore and north of Baltimore City decreased spatially after adjustment, while clusters 3 and 4, located in the south-west and north-east respectively increase spatially after adjustment (Figure 3).

Figure 3.

Top - Rash flare clusters adjusted for individual (age, sex, race, smoking status, personal income, years of education, urban living status) and county variables (county income, proportion of black population). Bottom - rash flare clusters adjusted for individual and county variables, as well as environmental variables (PM2.5, temperature, ozone, resultant wind, barometric pressure, relative humidity).

Three statistically significant environmentally unadjusted (p<0.05) hematologic clusters were identified around New Jersey (September 2000 – February 2002) and western Maryland and Virginia (December 2002 – December 2005). None of these clusters remained significant after adjusting for atmospheric variables and PM2.5 (Figure 4).

Figure 4.

Hematologic flare clusters adjusted for individual (age, sex, race, smoking status, personal income, years of education, urban living status) and county variables (county income, proportion of black population). No significant (p<0.05) clusters were identified after adjusting for all those, as well as environmental variables (PM2.5, temperature, ozone, resultant wind, barometric pressure, relative humidity).

Two large neurologic significant (p<0.05) clusters (Supplementary Figure 1) were identified, in the eastern and western halves of the study area. After adjusting for atmospheric variables and PM2.5, spatially and temporally smaller clusters during time periods different from the original clusters appeared Two significant (p<0.001) environmentally unadjusted serositis clusters (Supplementary Figure 2) were identified, and both remained temporally and spatially unchanged after the adjustment.

Four highly significant (p<0.001) renal clusters were identified (Supplementary Figure 3), spanning between 2001, 2002, or 2004 and 2006, taking up much of the map except for some counties in the west and south. These clusters remained largely unchanged spatially after adjusting for atmospheric variables and PM2.5, while clusters 2 and 3 became slightly temporally smaller after adjustment

Two significant (p<0.001) pulmonary clusters (Supplementary Figure 4) in the north-west were identified spanning between January and July of 1999. Both clusters became spatially smaller after adjustment for atmospheric variables and PM2.5, and a third cluster appeared in the north-east after adjustment.

DISCUSSION

Cluster detection, the identification of spatial units adjacent in space that are associated with distinctive patterns of data of interest relative to background variation, is an important tool in disciplines such as spatial epidemiology and disease surveillance¹⁶. Clusters have distinctive risks of an event of interest, typically elevated, but possibly reduced, relative to background variation¹⁶.

We performed a spatial-time cluster analysis of the Hopkins Lupus Cohort and detected the first space-time lupus flare clusters. Seasonal, as well as multi-year cluster patterns were identified, differing in extent and location for the various organ-specific flare types. The large-scale, multi-year clusters we defined did not conform to any known pattern of infectious disease or environmental exposure.

After adjusting for individual or environmental covariates, all outcomes experienced at least slight changes in temporal or spatial extents, or significance, of the identified clusters, with serositis clusters being the only ones that remained almost entirely unchanged. Generally, a decrease in temporal extent, as seen in some of the joints (Figure 2) or renal clusters (Supplementary Figure 3), or spatial extent, for example in the neurologic clusters (Supplementary Figures 1 and 4) after adjustment for covariates indicates that the adjusted covariates were partially driving the occurrence of that cluster at that time and location. Some clusters saw an increase in spatial extent after adjustment, such as the rash cluster 3, or temporal extent such as rash cluster 1 (Figure 2). An increase in cluster size, temporal extent, or significance after adjustment for covariates might suggest an area where flare activity is high, despite the presence of individual and environmental covariates that are associated with lower flare activity, and thus this is an area of particular interest for further research.

GEE based regressions were used to quantify effects of individual, county and environmental variables on the odds of flare outcomes. While these results were needed to properly adjust the cluster detection analysis, a more in-depth analysis was outside the scope of this paper. An inferenced based approach not only identifying and quantifying these effects but further investigations into effect modification, separating residual variability at the individual and county level (multi-level modeling) and deriving the most parsimonious models for each flare outcome is of interest and will be the primary focus of our future work.

The potential mechanisms underlying the effect of environmental factors on lupus flares is an interesting subject to speculate on. Elevated temperature has profound effects on the immune system, particularly by increasing T-cell proliferation rates, interleukin 1 (IL-1)-driven secretion of IL-2, and primary antibody responses to T-dependent antigens^17,18 but whether changes in environmental temperature affect the immune response is unknown. Rodo et al. described a causal relationship between large-scale wind currents originating in northeastern China with the major epidemics of Kawasaki disease in Japan, Hawaii, and San Diego¹⁹. Candida sp. were the dominant fungal species (54% of all fungal DNA clones) isolated in the aerosol samples from these wind currents²⁰ underlining the potential of aerosols transported by wind currents over long distances to trigger human disease. PM2.5 has been shown to alter innate immunity by affecting TLR signaling, inflammasome activation, and oxidative stress^21–23. One could speculate that similar mechanisms could underlie the effect of these environmental factors on lupus flares.

The shortcoming of cluster analysis as an epidemiological method is the low likelihood of establishing a definitive cause-and-effect relationship between the health event and an exposure. Clusters are useful for generating hypotheses but may not be as useful for testing hypotheses. The issues raised by a cluster cannot be definitively answered by the investigation per se; as they require an alternative epidemiologic approach. This is also true regarding the interpretation of the cluster changes after adjustment. While a change in clusters after adjustment for individual or environmental covariates suggests that these covariates in part drive the formation, location, or temporal extent of these clusters, the exact interpretation of every change can be difficult and require further study.

The difficulty in interpretation is partially driven by SaTScan software itself, which produces a ranked list of clusters based on significance. To make reporting and interpretability easier, by default, SaTScan only reports clusters that do not spatially overlap, meaning that there can be multiple identified statistically significant overlapping clusters in an area, but only the most significant of those would be reported. A slight change in the significance of the clusters could lead to a change in ranking order, and ultimately change what clusters are mapped and reported, even if the change in significance is minimal. By loosening the default restrictions and allowing SaTScan to report spatially overlaying clusters as long as they do not contain the centroids of more significant clusters, we allowed more significant clusters to be reported, without making the plots overwhelming.

We describe the first space-time clusters of lupus organ-specific disease activity. Seasonal, as well as multi-year cluster patterns were identified, differing in extent and location for the various organ-specific flare types. Many of the identified clusters changed in significance, temporal, or spatial extent after adjusting for environmental or individual covariates. Further study focusing on each individual lupus organ-specific activity will be required to better understand the driving forces behind these observed changes.