Time trends of variability in disease activity in systemic lupus erythematosus

Abstract

Objective

Disease activity both between and within patients with SLE is highly variable, yet factors driving this variability remain unclear. This study aimed to identify predictors of variability in SLE disease activity over time.

Methods

We analysed data from 2930 patients with SLE across 13 countries, collected over 38 754 clinic visits between 2013 and 2020. Clinic visit records were converted to panel data with 1-year intervals. The time-adjusted mean disease activity, termed AMS, was calculated. The yearly change in ${\displaystyle AMS}$, denoted as ${\displaystyle \mathrm{\Delta }AM{S}_t}$, was regressed onto ${\displaystyle AM{S}_{t-1}}$ and other potential predictors using random-effects models. Some variables were split into a person-mean component to assess between-patient differences and a demeaned component to assess within-patient variability.

Results

Overall, variability in SLE disease activity exhibited stabilisation over time. A significant inverse relationship emerged between a patient’s disease activity in a given year and variability in disease activity in the subsequent year: a 1-point increase in person-mean disease activity was associated with a 0.27-point decrease (95% CI −0.29 to –0.26, p<0.001) in subsequent variability. Additionally, a 1-point increase in within-patient disease activity variability was associated with a 0.56-point decrease (95% CI −0.57 to –0.55, p<0.001) in the subsequent year. Furthermore, each 1-point increase in the annual average time-adjusted mean Physician Global Assessment was associated with a 0.08-point decrease (90% CI −0.13 to –0.03, p=0.002) in disease activity variability for the following year. Prednisolone dose and the duration of activity in specific organ systems exhibited negative and positive associations, respectively, with disease activity variability in the subsequent year. Patients from less affluent countries displayed greater disease activity variability compared with those from wealthier nations.

Conclusion

Disease activity tends to be less variable among patients with higher or more variable disease activity in the previous year. Within-patient variability in disease activity has a stronger impact on subsequent fluctuations than differences between individual patients.

Correspondence to Dr Ning Li; [email protected]

WHAT IS ALREADY KNOWN ON THIS TOPIC

• SLE disease activity varies between patients and over time within patients. Understanding these variations is crucial for targeting interventions. Previous research has largely focused on either correlating fixed disease patterns with patient characteristics or examining internal transitions within individuals. Most studies are based on North American and European cohorts, highlighting the need for confirmation in diverse ethnic and geographical contexts.

WHAT THIS STUDY ADDS

• This study reveals that past disease activity negatively predicts future variability in disease activity at the individual patient level—more severe activity is followed by less variability. Additionally, greater within-patient variation predicts reduced future variability, with this effect surpassing between-patient differences. These findings underscore the importance of individual patient data tracking for prognostication in SLE.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The study confirms that high disease activity or wide fluctuations in disease activity often precede more stable periods, suggesting a homeostasis-like dynamic in SLE. This stabilisation may enable prediction of a patient’s total disease activity using limited follow-up data, allowing for lifetime healthcare cost estimations. These findings could enhance resource allocation and inform targeted SLE management strategies, maximising benefits for both individuals and communities.

Introduction

SLE is characterised by significant variability in disease activity both over time and between patients. Disease activity typically fluctuates between periods of high and low intensities, with variable durations. Long-term trajectories, however, have been less well studied. Understanding the factors contributing to these fluctuations could enhance the ability to tailor treatments to patients who would benefit the most and improve the extrapolation of disease activity in the economic assessment of SLE treatments.^1–4

Predicting changes in SLE disease activity over time is challenging. Prior research has used time-adjusted mean measurements to summarise disease activity over a specific time frame⁵ or identify patterns in disease activity and explain the pattern by baseline characteristics.^{6 7} Other researchers have studied the within-person transitions in disease activity and the factors affecting these transitions. One such study, conducted by Watson and colleagues based on the Hopkins Lupus cohort, proposed a random-effects (RE) model. This model conceptualised the change in yearly mean disease activity as a function of the mean disease activity in the previous year, alongside other covariates.⁸ More recently, when applied to the Systemic Lupus International Collaborating Clinics (SLICC) inception cohort,⁹ this model yielded results suggesting that disease activity inversely affects subsequent disease variability. However, the applicability of this type of models in other ethnic and healthcare system contexts is unknown.

The Asia Pacific Lupus Collaboration¹⁰ (APLC) is a large, multicentre, multinational longitudinal cohort of patients with SLE across Asia Pacific countries, where inherent disease severity and healthcare systems differ from those in North America and Europe. In our exploration, we tested the construct of Watson et al but found that a straightforward application of the model could not adequately fit our data. Thus, we refined the model by ‘splitting’ the variability in a patient’s disease activity into two independent parts: one reflecting between-patient variation and the other reflecting within-patient variation. The enhanced model yielded a new understanding on how past disease activity can predict future changes.

Methods

Study cohort

The APLC cohort comprises patients from 25 sites in Australia, Taiwan, Thailand, Singapore, Malaysia, Hong Kong, Indonesia, China, the Philippines, Japan, South Korea, New Zealand and Sri Lanka, listed in the order of total person-years. All patients met either the 1997 American College of Rheumatology Modified Classification Criteria¹¹ or the SLICC 2012 Classification Criteria¹² for SLE. The current analysis focused on patients enrolled between 2013 and 2020 with at least 2 years of longitudinal data; all data were collected longitudinally using specifically designed data collection templates.

Patient and public involvement

There was no direct involvement of patients or the public in the design or conduct of this study.

Data

Data collection

Patient demographics, including age, gender, ethnicity, dates of SLE onset and diagnosis, and education level, were collected at enrolment. The SLE Disease Activity Index 2000 (SLEDAI-2K),^{13 14} the SELENA-SLEDAI Flare Index,¹⁵ the Physician Global Assessment (PGA)¹⁶ and medication usage were collected during routine clinically indicated visits, generally at intervals of 3–6 months. Disease activity indicators for specific organ domains, including the central nervous system (CNS), vasculitis, musculoskeletal (MSK), renal, mucocutaneous, serositis, serological, haematological and constitutional domains, were recorded in binary form (present or absent) without weighting.

Transforming irregular intervisit intervals to equally spaced data

To minimise potential confounding effects from irregular visit intervals, the raw data were converted into a panel data format with equal time intervals. Each patient’s observation period was divided into 1-year intervals, and the time-adjusted mean SLEDAI-2K scores from all visits within the tth 1-year interval were averaged, denoted as AMSt (see online supplemental table A for illustration). The variable ${\displaystyle AM{S}_t}$ served to portray disease activity for year t, collectively referred to as ${\displaystyle AMS}$. Similarly, the annual average time-adjusted mean PGA and the annual average time-adjusted mean prednisolone (PNL) dose were calculated and denoted.

Outcome variable

The difference in ${\displaystyle AMS}$ between annual periods, calculated as ${\displaystyle \mathrm{\Delta }AM{S}_t=AM{S}_t-AM{S}_{t-1}}$, was used as the outcome variable. This difference quantifies the fluctuation, that is, change, in disease activity within individual patients from year to year. A small value of ${\displaystyle \mathrm{\Delta }AMS}$ indicates that disease activity is less variable between the years, while a large value of ${\displaystyle \mathrm{\Delta }AMS}$ signifies greater variability and a wider range of changes in disease activity.

While the primary goal is to understand how ${\displaystyle AM{S}_t}$ changes over time, using $\mathrm{\Delta }AM{S}_t$ as the outcome variable helps eliminate time-invariant individual heterogeneity, providing a better opportunity to reveal patterns in disease activity.

Independent variables

The primary independent variable was the mean disease activity in the previous year, ${\displaystyle AM{S}_{t-1}}$. An equation of the form ${\displaystyle \mathrm{\Delta }AM{S}_t=\alpha +\beta AM{S}_{t-1}}$ links ${\displaystyle AM{S}_t}$ to ${\displaystyle AM{S}_{t-1}}$ via ${\displaystyle AM{S}_t=\alpha +(1+\beta )AM{S}_{t-1}}$. This link allows for examination of the transition in ${\displaystyle AMS}$ over time. Recursive substitution for ${\displaystyle t=1,2,\cdots }$ expresses the disease activity in a given year as a function of the disease activity in prior years.

Other independent variables included age and disease duration at enrolment, age at diagnosis, gender, ethnicity (Asian or non-Asian), tertiary education (yes or no), gross domestic product (GDP) of home country at purchasing power parity per capita (<$20 000, $20 000–$35 000, $35 000–$50 000, >$50 000), and organ-specific indicators for CNS, vasculitis, MSK, renal, cutaneous, serositis, serological, and haematological activity, glucocorticoid (PNL) use and dose, and antimalarial (AM) and immunosuppressant (IS) use. We also considered the potential predictive power of PGA for the variability in disease activity that may not be fully captured by SLEDAI-2K.

Statistical analysis

Replicating the Watson model

We regressed ${\displaystyle \mathrm{\Delta }AM{S}_t}$ on the same independent variables, $x_{it}$, as in the published analyses,⁸ assuming that the individual-specific effect $u_i$ follows a normal distribution ${\displaystyle u_i\sim N(0,{\sigma }_u^2)}$ and is uncorrelated with the regressors (${\displaystyle Cov(u_i|x_{it})=0}$).¹⁷ The model estimation was assessed using residual-versus-fitted value plots and the Hausman specification test¹⁸ that evaluates the null hypothesis H₀: the fitted random-effects model is statistically valid.

Distinguishing between-person and within-person variations

In our analysis, the independent variables ${\displaystyle AM{S}_{t-1}}$ and organ-specific involvement indicators were each separated into a person-mean variable and a deviation-from-person-mean variable. The person-mean variable characterises a patient’s overall disease activity, while the deviation variable is a demeaned trajectory of disease activity, obtained by shifting the original trajectory downward by an amount of the person-mean. This demeaned trajectory therefore has a mean of zero and is identical in shape to the original trajectory, thereby preserving all fluctuations in the original data as a measure of pure variability rather than total activity.

As illustrated in figure 1, the person-mean variable for ${\displaystyle AM{S}_t}$, denoted as ${\displaystyle AM{S}_t^m}$, is the typical ${\displaystyle AMS}$ value for a patient over the entire observation period. This variable remains constant for a patient throughout the observation period, with differences arising only between patients. The person-mean variable facilitates the analysis of between-patient differences (see also online supplemental material C).

Figure 1. Illustration of a patient’s disease activity trajectory and its split components. Hollow squares depict the original variable ( A M S t ), the line represents the person-mean variable ( A M S t m ) and solid squares denote the demeaned deviation variable ( A M S t d ). SLEDAI-2K, SLE Disease Activity Index 2000.

In contrast, the deviation-from-person-mean variable, calculated as ${\displaystyle AM{S}_t^d=AM{S}_t-AM{S}_t^m}$, captures the pure temporal fluctuations within each individual. Importantly, ${\displaystyle AM{S}_t^d}$ and ${\displaystyle AM{S}_t^m}$ are independent of each other. This separation allows for a comprehensive assessment of how between-person and within-person variations impact the outcome variable without requiring extra data.

Improved models

Univariate associations between ${\displaystyle \mathrm{\Delta }AM{S}_t}$ and candidate predictors were examined using RE regression of ${\displaystyle \mathrm{\Delta }AM{S}_t}$ on each variable. Candidate predictors with a p value <0.1 were included in an initial multivariate model, which was then refined using stepwise variable selection. This process continued until the model passed the Hausman test. During the model derivation, certain variables were replaced by their corresponding split components where necessary.

Results

Patient characteristics

Among the 2930 patients analysed, 87% were Asian, 92% were female, nearly half had tertiary education. While more than half of the patients were from countries with a GDP>$50 000 per capita, 14% were from countries with a GDP<$20 000, and 25% and 11% were from countries with a GDP between $20 000–$35 000 and $35 000–$50 000, respectively (table 1). The median ages were 29 (IQR 22–40) years at SLE diagnosis and 39 (IQR 30–50) years at enrolment. The most prevalent organ-specific disease activity was serological (83%), followed by renal (47%) and cutaneous (41%). PNL, AM and IS were used by 86%, 78% and 73% of patients, respectively. The daily PNL dose decreased over time, at an average rate of 0.4 mg/year, with mean doses of 10.8 and 5.2 mg at the first and last visits, respectively.

Table 1

Cohort properties and patient characteristics

	Total number of patients=2930 Total number of visits=38 754
Cohort properties
Duration of follow-up (years), mean (SD), median (IQR)	4.05 (2.2)	3.76 (2.0, 5.9)
Days between clinic visits, mean (SD), median (IQR)	121.0 (81.6)	98 (84, 154)
Intervals between clinic visits >122 days, n (%)	11 425 (29.5)
Intervals between clinic visits >365 days, n (%)	648 (1.7)
Demographics
Age at enrolment (years), mean (SD), median (IQR)	40.91 (13.5)	39 (30, 50)
Age at diagnosis (years), mean (SD), median (IQR)	31.36 (12.7)	29 (22, 40)
Disease duration at enrolment (years), mean (SD)	10.00 (8.6)	8 (3, 15)
Disease duration <1 year at cohort enrolment, n (%)	237 (8.1)
Female, n (%)	2710 (92.5)
Asian ethnicity, n (%)	2557 (87.3)
Tertiary education, n (%)	1396 (47.7)
Country based on GDP (PPP) per capita (Int$), n (%)
<$20 000	401 (13.7)
≥$20 000 ≤$35 000	742 (25.3)
>$35 000 ≤$50 000	311 (10.6)
>$50 000	1476 (50.4)
Disease activity indicators
Flare at visit in the study, n (%)	5597 (14.4)
Flare during the entire study period, n (%)	1806 (61.6)
PGA at enrolment, mean (SD), median (IQR)	0.54 (0.5)	0.4 (0.2, 0.8)
PGA at last visit of the study, mean (SD), median (IQR)	0.48 (0.5)	0.3 (0.1, 0.7)
SLEDAI-2K during the study, mean (SD), median (IQR)	3.62 (3.7)	2 (0, 5)
SLEDAI-2K at enrolment of the study, mean (SD), median (IQR)	4.22 (4.3)	4 (2, 6)
SLEDAI-2K at last visit of the study, mean (SD), median (IQR)	2.88 (3.2)	2 (0, 4)
AMS during the study, mean (SD), median (IQR)	3.58 (2.8)	3.20 (1.6, 5.0)
AMS in the first year of the study, mean (SD), median (IQR)	3.78 (3.2)	3.25 (1.6, 5.2)
AMS in the last year of the study, mean (SD), median (IQR)	3.35 (2.6)	2.97 (1.4, 4.7)
${\displaystyle \mathrm{\Delta }AM{S}_t}$during the study, mean (SD), median (IQR)	−0.13 (0.9)	−0.02 (−0.4, 0.2)
${\displaystyle \mathrm{\Delta }AM{S}_t}$in the second year of the study, mean (SD), median (IQR)	−0.28 (1.3)	0.00 (−0.7, 0.3)
$\mathrm{\Delta }AM{S}_t$in the last year of the study, mean (SD), median (IQR)	−0.13 (0.8)	−0.03 (−0.3, 0.1)
Organ-specific disease activity at least once, n (%)
Central nervous system	79 (2.7)
Vasculitis	117 (4.0)
Musculoskeletal	646 (22.1)
Renal	1366 (46.6)
Cutaneous	1204 (41.1)
Serositis	117 (4.0)
Serological	2426 (82.8)
Haematological	650 (22.2)
Organ-specific disease activity at visit, n (%)
Central nervous system	145 (0.4)
Vasculitis	229 (0.6)
Musculoskeletal	1625 (4.2)
Renal	8985 (23.2)
Cutaneous	4399 (11.4)
Serositis	230 (0.6)
Serological	24 006 (61.9)
Haematological	2382 (6.2)
Medication use
Used PNL during the study period, n (%)	2525 (86.2)
Proportion of patients receiving PNL at 1st visit of the study, n (%)	2319 (79.2)
PNL dose at first visit of the study, mean (SD), median (IQR)	10.77 (11.6)	7 (5, 10)
Proportion of patients receiving PNL at last visit of the study, n (%)	2095 (71.5)
PNL dose at last visit of the study, mean (SD), median (IQR)	5.18 (7.6)	5 (0, 6.5)
Annual average time-adjusted mean PNL dose (mg/day), mean (SD), median (IQR)	6.16 (5.7)	5 (2.3, 8.7)
Change in adjusted mean PNL dose, mean (SD), median (IQR)	−0.41 (2.3)	−0.03 (−0.6, 0.0)
Used antimalarial during the study period, n (%)	2284 (78.0)
Used immunosuppressants during the study period, n (%)	2127 (72.6)

ΔAMS_t = AMS_t − AMS_t−1, where AMS_t is AMS in year t since enrolment.

AMS, annual average of time-adjusted mean SLEDAI-2K scores at visits within 1-year interval; GDP, gross domestic product; Int$, international dollar 2019; PGA, Physician Global Assessment; PNL, prednisolone; PPP, purchasing power parity; SLEDAI-2K, SLE Disease Activity Index 2000.

Disease activity declined over time, with the median SLEDAI-2K score decreasing from 4 (IQR 2–6) at enrolment to 2 (IQR 0–4) at the last visit of this study. Similarly, the median $AMS$ score dropped from 3.3 (IQR 1.6–5.2) in the first year to 3 (IQR 1.4–4.7) in the last year. The median PGA score also showed a decline, decreasing by a quarter over the follow-up period.

Figure 2 visualises the crude ${\displaystyle AMS}$ scores over time for 147 patients (5% of the cohort), providing a detailed view of patient-level variability. However, identifying clear time trends in this figure needs advanced regression models, which we present in later sections. In contrast, figure 3 aggregates disease activity changes for all patients within each year, showing a clear overall decline in ${\displaystyle \mathrm{\Delta }AMS}$ scores (mean ${\displaystyle \mathrm{\Delta }AM{S}_t}$ of −0.3 between the first and second years, and −0.1 between the last 2 years). While informative, this aggregated view does not capture patient-specific time trends, which are the primary focus of this study.

Univariate associations of ΔAMS_t

When examined individually, many variables showed significant associations with the variability of disease activity (table 2). Age and disease duration at enrolment, GDP exceeding $50 000 per capita, flare at year ${\displaystyle (t-1)}$ and between-year changes in the adjusted mean PNL dose each showed a positive correlation with ${\displaystyle \mathrm{\Delta }AM{S}_t}$.

Figure 2. Line graph of time-adjusted mean SLE Disease Activity Index 2000 (SLEDAI-2K) scores over time for a representative, simple random sample of 147 patients (5% of the total cohort).

Figure 3. Yearly changes in A M S over time, that is, [DELTA] A M S 1 , [DELTA] A M S 2 , ⋯ , [DELTA] A M S 8 . AMS , annual average of time-adjusted mean SLEDAI-2K scores at visits within 1-year interval; SLEDAI-2K, SLE Disease Activity Index 2000.

Table 2

Univariate associations of ΔAMS_t

	Coefficient (95% CI)	P value
Demographics
Logarithm of age at enrolment	0.12 (0.03 to 0.21)	0.008
Logarithm of age at SLE diagnosis	−0.02 (−0.09 to 0.05)	0.604
Disease duration at enrolment	0.01 (0.00 to 0.01)	<0.001
Males	0.06 (−0.05 to 0.18)	0.258
Asian ethnicity	0.05 (−0.04 to 0.14)	0.250
Tertiary education	−0.06 (−0.12 to −0.00)	0.046
GDP per capita (Int$)
<$20 000	Ref
$20 000–$35 000	0.05 (−0.05 to 0.15)	0.301
$35 000–$50 000	0.05 (−0.08 to 0.18)	0.480
>$50 000	0.10 (0.01 to 0.20)	0.024
Disease activity at (t−1)
AMSt−1	−0.27 (−0.28 to −0.26)	<0.001
PGAt−1	−0.43 (−0.49 to −0.36)	<0.001
Any flare	0.49 (0.44 to 0.54)	<0.001
Organ-specific disease activity
Central nervous system	1.13 (0.80 to 1.46)	<0.001
Vasculitis	1.25 (0.99 to 1.51)	<0.001
Musculoskeletal disorder	0.44 (0.35 to 0.53)	<0.001
Renal	0.45 (0.40 to 0.49)	<0.001
Cutaneous	0.26 (0.20 to 0.32)	<0.001
Serositis	0.54 (0.30 to 0.79)	<0.001
Serological	0.25 (0.21 to 0.30)	<0.001
Haematological	0.17 (0.09 to 0.25)	<0.001
Medication
PNL use at (t−1)	−0.10 (−0.17 to −0.03)	0.007
Adjusted mean PNL dose at (t−1) (mg/day)	−0.06 (−0.06 to −0.05)	<0.001
Change in adjusted mean PNL dose (mg/day)	0.16 (0.15 to 0.17)	<0.001
Antimalarial use at (t−1)	−0.02 (−0.07 to 0.03)	0.545
Immunosuppressant use at (t−1)	−0.15 (−0.20 to 0.11)	<0.001

ΔAMS_t = AMS_t − AMS_t−1, where AMS_t is AMS in year t since enrolment.

AMS, annual average of time-adjusted mean SLEDAI-2K scores at visits within 1-year interval; GDP, gross domestic product; Int$, international dollar; PGA_t−1, annual average of time-adjusted mean Physician Global Assessment score in year (t−1); PNL, prednisolone; SLEDAI-2K, SLE Disease Activity Index 2000.

In contrast, ${\displaystyle AM{S}_{t-1}}$ was negatively associated with ${\displaystyle \mathrm{\Delta }AM{S}_t,}$ where every 1-point increase in ${\displaystyle AM{S}_{t-1}}$ corresponded to a 0.27-point reduction in ${\displaystyle \mathrm{\Delta }AM{S}_t}$. In parallel, every 1-point increase in the annual average time-adjusted mean PGA at prior year, ${\displaystyle PG{A}_{t-1}}$, corresponded to a 0.84-point reduction in ${\displaystyle \mathrm{\Delta }AM{S}_t}$. Essentially, patients with more severe or more fluctuating disease activity tend to experience a less variable disease course in the following year.

Negative associations were also observed between ${\displaystyle \mathrm{\Delta }AM{S}_t}$ and tertiary education, previous-year PNL use and IS use, and average daily PNL dose. Organ-specific activity in year ${\displaystyle (t-1)}$ showed varying degrees of association with increased ${\displaystyle \mathrm{\Delta }AM{S}_t}$, ranging from 0.17 points in the haematological domain to 1.25 points in vasculitis. Moreover, a 5-mg increase in the mean daily PNL dose in year ${\displaystyle (t-1)}$ correlated with a 0.3-point reduction in ${\displaystyle \mathrm{\Delta }AM{S}_t}$. AM use at ${\displaystyle (t-1)}$ did not show a significant association with ${\displaystyle \mathrm{\Delta }AM{S}_t}$.

Applying the Watson model in APLC data

The regression coefficients obtained by applying the same variables as the Watson model to APLC data were broadly consistent with those reported in the Hopkins and SLICC cohorts^{8 9} (online supplemental table B). First, a negative coefficient for ${\displaystyle AM{S}_{t-1}}$ affirmed prior findings that disease activity inversely affects subsequent variability. Furthermore, the magnitude of estimated effects on ${\displaystyle \mathrm{\Delta }AM{S}_t}$ due to gender, age at enrolment, low complement, haematological involvement and anaemia in the previous year showed overlapping 95% CIs with previous studies. The coefficient of gender was non-significant at the 5% significance level in both APLC and Hopkins Lupus analyses. Previous renal involvement was positively associated with ${\displaystyle \mathrm{\Delta }AM{S}_t}$ (p value <0.001), differing from the Hopkins findings but aligning with the SLICC results. Conversely, previous haematological involvement was associated with increased ${\displaystyle \mathrm{\Delta }AM{S}_t}$, consistent with the Hopkins findings but differing from the SLICC results.

However, a thorough examination revealed that the specific Watson model was not suitable for our data. The Hausman test of the estimated APLC model returned a huge χ²=2316.0 with a p value <0.001, hence rejecting the null hypothesis. Besides, a linear pattern was evident in the residual-versus-fitted value plot (figure 4A), indicating that substantial information from the independent variables predictive of ${\displaystyle \mathrm{\Delta }AM{S}_t}$ was not captured by the model. Moreover, the correlation coefficient ${\displaystyle corr(u_i,Xb)=-0.81}$ suggested that the assumption ${\displaystyle corr\left(u_i,Xb\right)=0}$, required for RE models, did not hold.

Figure 4. Residual versus fitted values for two models. (A) Watson model applied to Asia Pacific Lupus Collaboration (APLC) data. (B) Improved model. In a well-fitted model, residuals should be randomly distributed around zero. Graph A shows a linear pattern, indicating predictive information in the residuals, while graph B demonstrates a better fit with nearly random residuals.

Improved model

We refined Watson model and assessed the validity of the refined version. This process relied on the rationale that an independent variable’s influence is strongest when it closely correlates with the outcome variable. In our data, ${\displaystyle corr(\mathrm{\Delta }AM{S}_t,AM{S}_{t-1})=-0.38,}$ ${\displaystyle corr(\mathrm{\Delta }AM{S}_t,AM{S}_{t-1}^m)=-0.25}$ and ${\displaystyle corr(\mathrm{\Delta }AM{S}_t,AM{S}_{t-1}^d)=-0.50}$. These correlation coefficients suggest that ${\displaystyle AM{S}_{t-1}^m}$ and ${\displaystyle AM{S}_{t-1}^d}$ as two independent variables offer better explanatory power for ${\displaystyle \mathrm{\Delta }AM{S}_t}$ compared with ${\displaystyle AM{S}_{t-1}}$ which is the sum ${\displaystyle AM{S}_{t-1}^m+AM{S}_{t-1}^d=AM{S}_{t-1}}$. Organ-specific activity indicators were also split into within-person and between-person forms to increase statistical power. These split components were found to affect ${\displaystyle \mathrm{\Delta }AM{S}_t}$ at different scales. Without this separation, the unequal effects would have been treated as if equal, masking the true time trends.

The resulting model, presented in table 3, passed the Hausman test with a test statistic χ²=14.72 and a much more probable p value=0.195. The previously observed linear pattern in the residual-versus-fitted value plot nearly disappeared (figure 4B), indicating that the variables in the new model captured most relevant information about ${\displaystyle \mathrm{\Delta }AMS}$. The correlation coefficient between $u_i$ and ${\displaystyle Xb}$ dropped to ${\displaystyle corr(u_i,Xb)=0.02}$. Therefore, the new model was considered statistically valid. Collectively, the variables in this model explained 57% of the variability in ${\displaystyle \mathrm{\Delta }AM{S}_t}$, a high R² value for observational data analysis.

Table 3

New random-effects model of ΔAMS_t as a function of AMS_t−1, analysed in the APLC cohort

	Regression coefficient (95% CI)	P value
AMSt−1—person-mean	−0.27 (−0.29 to −0.26)	<0.001
AMSt−1—deviation from person-mean	−0.56 (−0.57 to −0.55)	<0.001
Time-adjusted mean PGA at (t−1)	−0.08 (−0.13 to −0.03)	0.002
Time-adjusted mean PNL dose at (t−1) (5 mg/day)	−0.05 (−0.06 to −0.04)	<0.001
GDP
<$20 000	Reference
$20 000–$35 000	−0.08 (−0.15 to −0.00)	0.038
$35 000–$50 000	−0.11 (−0.20 to −0.02)	0.019
>$50 000	−0.05 (−0.11 to 0.02)	0.171
Organ-specific disease activity at (t−1)
CNS—person-mean	1.83 (1.34 to 2.32)	<0.001
CNS—deviation from person-mean	1.23 (1.01 to 1.44)	<0.001
Vasculitis—person-mean	1.89 (1.48 to 2.31)	<0.001
Vasculitis—deviation from person-mean	1.32 (1.15 to 1.49)	<0.001
Renal—person-mean	1.49 (1.40 to 1.57)	<0.001
Renal—deviation from person-mean	0.77 (0.74 to 0.81)	<0.001
MSK—person-mean	0.95 (0.80 to 1.11)	<0.001
MSK—deviation from person-mean	0.53 (0.48 to 0.59)	<0.001
Cutaneous—person-mean	0.75 (0.65 to 0.84)	<0.001
Cutaneous—deviation from person-mean	0.33 (0.29 to 0.37)	<0.001
Serositis—person-mean	1.21 (0.72 to 1.69)	<0.001
Serositis—deviation from person-mean	0.34 (0.19 to 0.49)	<0.001
Serological—person-mean	0.85 (0.80, 0.91)	<0.001
Serological—deviation from person-mean	0.37 (0.34, 0.40)	<0.001
Haematological—person-mean	0.42 (0.30, 0.52)	<0.001
Haematological—deviation from person-mean	0.14 (0.09, 0.20)	<0.001
Constant	0.03 (−0.04, 0.10)	0.466
${\displaystyle \widehat{{\sigma }_u}}$	0.51
${\displaystyle \widehat{{\sigma }_e}}$	0.41
corr(ui, Xb)	0.02
Root mean square error	0.41
Hausman test statistics	χ2=14.72, p=0.195
R2—overall	0.57
Number of patients	2870
Number of observations	9943

The annual average time-adjusted mean PGA and PNL dose at (t−1) had overlapping 95% CIs for their split components, so the unsplit variables were included in the model, indicating similar between-person and within-person effects.

${\displaystyle \widehat{{\sigma }_e}}$: Estimated SD of the usual stochastic error term (ε) that varies among patients and over time.

${\displaystyle \widehat{{\sigma }_u}}$: Estimated SD of the patient-specific error term u that varies between patients but remains constant for a given patient.

AMS_t−1, annual average of time-adjusted mean SLEDAI-2K scores at year (t−1); APLC, Asia Pacific Lupus Collaboration; CNS, central nervous system, including seizure, psychosis, organic brain syndrome, visual disturbance, cranial nerve disorder, lupus headache and stroke; corr(u_i, Xb), estimated correlation coefficient between the patient-specific error (u_i) and the population mean (Xb); GDP, gross domestic product; MSK, musculoskeletal disorder; PGA, Physician Global Assessment; PNL, prednisolone; SLEDAI-2K, SLE Disease Activity Index 2000.

Predictors of ΔAMS_t

A significant inverse correlation was observed between disease activity in a given year and subsequent disease activity variability. Specifically, a 1-point increase in the person-mean disease activity for a given year corresponded to a 0.27-point decrease (95% CI −0.29, –0.26, p<0.001) in the subsequent year’s disease activity variability ${\displaystyle (\mathrm{\Delta }AM{S}_t)}$ (table 3). Additionally, a 1-point increase in within-person variability (ie, deviation from the person-mean) was associated with a 0.56-point decrease (95% CI −0.57, –0.55, p<0.001) in the variability of disease activity for the subsequent year. Furthermore, a 1-point increase in the annual average time-adjusted mean PGA was associated with a 0.08-point decrease (90% CI −0.13, –0.03, p=0.002) in disease activity variability for the subsequent year.

The model also identified additional factors such as glucocorticoid use and country of residence as contributing to the variability. Glucocorticoid dose exhibited a small but statistically significant negative correlation with disease activity variability in the subsequent year. Notably, disease activity variability was lower among patients from relatively more affluent countries ($20 000≤GDP≤$50 000) compared with those from the poorest countries (GDP<$20 000), but no discernible association was observed with further level of affluence (GDP>$50 000). Moreover, ethnicity did not significantly correlate with disease activity variability when considered alongside the variables in this model. These findings underscore the complexity of factors influencing variation in SLE disease activity and highlight the need for personalised treatment strategies based on detailed understanding of disease patterns.

We next assessed organ-specific disease activity in relation to prediction of subsequent variability in disease. To interpret the organ-specific disease activity variables, recall the fact that, in regression models including a total score and its component indicators, each indicator’s regression coefficient represents its relative impact on the outcome variable while holding the total score constant. In our case, ${\displaystyle AM{S}_{t-1}}$ was based on the total SLEDAI-2K score derived from the organ-specific domains. With this understanding, the following interpretation is available.

Given ${\displaystyle AMS}$ from the previous year and other variables in the model, individual organ domain activity had highly varying degrees of effect on subsequent changes in ${\displaystyle AMS.}$ These effects ranged from a minimum 0.14-point relative impact in association with haematological activity to a maximum 1.89-point relative impact in association with vasculitis. Specifically, across patients, a 10% increase in the time experiencing vasculitis was correlated with a 0.19-point increase in subsequent disease variability (10% of 1.89). Likewise, a 10% increase in time with haematological involvement correlated with a 0.04-point rise in subsequent disease variability (10% of 0.42). Within individual patients, transitioning from inactive to active vasculitis was associated with a 1.32-point increase in subsequent disease variability. Shifting from inactive to active haematological involvement was correlated with a modest 0.14-point increase in subsequent disease activity variability. The interpretations are similar for other domains (table 3).

Discussion

We identified key predictors of changes in mean SLE disease activity by revising the RE model initially proposed by Watson et al⁸ and later replicated by Clarke et al.⁹ The predictors in our study, jointly explaining 57% of the variation in disease activity changes, comprised disease activity indices in the previous year, annual average daily glucocorticoid dose over the previous year and GDP of the patient’s residence country. The disease activity indices included both the annual average time-adjusted mean SLEDAI-2K and the annual average time-adjusted mean PGA, with the later capturing subtler fluctuations in disease activity that were not fully captured by the aggregated SLEDAI-2K domains. The revised model, which distinguished within-patient changes from between-patient differences, revealed previously unseen patterns.

An important finding from our study was that higher overall disease activity (greater person-mean disease activity) is associated with reduced variability in subsequent disease activity. This suggests that heightened disease activity tends to persist, requiring more effort to attenuate. For instance, if two patients are identical except for a 1-point difference in typical disease activity, an additional 0.27 units of medication would be needed to reduce the fluctuation in the patient with higher disease activity to the same level as the other.

Additionally, we revealed that increased disease variability in 1 year (greater deviation from the person-mean) is associated with reduced variability in the following year. Clinically, this suggests that a period of elevated disease activity variability is often followed by a period of less variable, or more stable, disease activity. Similar to the example above, to reduce disease activity fluctuation by 1 SLEDAI-2K point, an additional 0.56 units of medication would be needed if disease activity increased by 1 point during the previous period, compared with no increase.

The patterns revealed in our study suggest long-term stabilisation in SLE disease activity, following the similar arguments in a recently retested model.¹⁹ The stability is not static: disease activity still varies, but over a smaller range. Such dynamic stability is akin to homeostasis, where the body maintains stable internal conditions. However, unlike well-understood self-regulatory mechanisms such as blood sugar regulation via insulin, understanding of the biological process responsible for long-term SLE disease stabilisation is limited.

Our new insight that the predictive influence on ${\displaystyle \mathrm{\Delta }AMS}$ predominantly arises from within-patient variation was obtained from a novel approach that distinguishes within-person influence from between-person influence on disease activity trends, a consideration overlooked in previous studies. In longitudinal data, these two dimensions are often intertwined, and when their effects on an outcome differ, statistical modelling may require additional refinement to identify the optimal lens for a comprehensive understanding of the phenomenon. The predominance of within-patient variation indicates the importance of individualised disease tracking in managing SLE. Our findings also offer the feasibility of estimating a patient’s total disease activity over a protracted period using limited follow-up data. This opens the opportunity to predict lifetime healthcare costs and to guide interventions aimed at minimising future disease activity.

While our study supported previous findings^{8 9} and deepened the understanding of SLE disease activity dynamics, several assumptions in our analysis may limit the generalisability of our results. The many short panels generated by the short follow-up period (median 3.8 years) posed challenges for estimating the time trend, partly a reason for the splitting. While RE models effectively capture individual heterogeneity, a longer study duration with a more homogeneous sample would enhance the robustness of our findings. Moreover, we assumed a negligible effect of date on disease activity trends, as no new SLE treatments emerged during data collection. Transforming visit records into panel data helped control the confounding effects of irregular visit intervals, but this required the assumption that the visit frequency did not inherently reflect disease activity. Lastly, our analysis mainly relied on SLEDAI-2K, an overall measure of disease activity with acknowledged limitations,²⁰ such as large differences in weighting between organ domains. Future studies should consider these limitations for improvement.

In brief, through rigorous validation, we have uncovered an inverse relationship between disease activity and subsequent variability, with within-person differences being the dominant factor. These findings may help forecast disease activity trends in cohorts, assist with health economic analyses and guide interventions to reduce future disease activity risk per patient using historical longitudinal data.

We are grateful for Dr Julie Monk’s invaluable assistance with the first draft of the manuscript. We also sincerely thank the APLC patient cohort and the data collectors for their essential contributions, as well as the anonymous reviewers for their insightful suggestions, which have improved our understanding and presentation.

Data availability statement

Data are available upon reasonable request. Access to APLC pooled data is subject to the specific guidelines outlined in the APLC Data Access Policy (available upon request). The APLC welcomes requests for aggregate data or analyses of new research questions. Interested parties should submit requests to the APLC steering committee via the APLC project manager.

Ethics statements

Patient consent for publication

Not required.

Ethics approval

Each participating site secured local ethics approval for APLC research activities. The Monash University Human Research Ethics Committee (MUHREC) approved the storage of the central dataset and the analysis of pooled data under Project ID 18778. Written informed consent was obtained from all patients.

Footnote

X @n/a

Contributors All authors reviewed and approved both the initial and final manuscripts. NL recognised the need for and developed the data-splitting method, conducted analysis, interpretation, and wrote the manuscripts. EM, as the guarantor, was responsible for the final decision to publish, oversaw funding acquisition, human resources management, interpretation, and provided extensive editing of the initial and final manuscripts. AH, WL, JCho, SN, SO, LH, FG, TT, YT and EM contributed to critical questioning, interpretation, data collection and review and editing. S-FL, Y-JJW, VG, SS, ALat, SO’N, CSL, MN, YH, MC, ZL, LZ, YK, MH, S-CB, ZZ, JK, KN, NT, NO, Y-HC, BMDBB, ALaw, SK, CT and MLT contributed to data collection, reviewed the manuscripts and provided approval. JChoi contributed to critical questioning and review, was involved in project administration and approved the manuscripts. RK-R was the primary person responsible for funding acquisition, project administration, human resources management, data centralisation, and contributed to questioning and review and editing.

Funding This study was supported by Bristol Myers Squibb (BMS) under the protocol ‘Validation of a new conceptual lupus long-term outcomes prediction model, Observational Study (Protocol IM011-164)’. BMS supplied the Watson model and requested its validation, thus contributing to the study design. BMS also approved the manuscript’s dissemination and coordinated its submission through OpenHealth. The Asia Pacific Lupus Collaboration (APLC) received support from AstraZeneca, Bristol Myers Squibb, EMD Serono, GSK, Janssen, Eli Lilly and UCB. Additionally, S-CB's research was funded by the National Research Foundation of Korea (NRF) through the Ministry of Education (NRF-2021R1A6A1A03038899).

Competing interests NL’s salary from Monash University was partly supported by a research grant from BMS. AH has received a research grant from AstraZeneca, consulting fees from EUSA Pharma (UK), GSK and UCB Australia and speaker fees/honoraria from AbbVie, Eli Lilly, Janssen, Limbic, Moose Republic and Novartis. SS has received consulting fees from Pfizer, AstraZeneca and ZP Therapeutics. MN has received an investigator grant from the National Health and Medical Research Council of Australia (NHMRC GNT1176538), research grants from Boehringer Ingelheim and Janssen, consulting fees from AstraZeneca and GSK, honoraria for presentations from AstraZeneca, Boehringer Ingelheim and GSK and support for conference attendance from Boehringer Ingelheim. SO has received speaker fees/honoraria from AstraZeneca and Limbic. ZL has received consulting fees from Pfizer, Roche, Janssen, Abbott, AbbVie, Bristol Myers Squibb, MSD, Celgene, Eli Lilly, GSK, Novartis and UCB Pharma, and holds royalties with these companies. SN has received consulting and lecture/speaker fees from AstraZeneca, Biogen and Boehringer Ingelheim and is a non-paid member of Viatris (Idorsia) Advisory Board. YK has received payment/honoraria from GlaxoSmithKline KK, AstraZeneca KK, Sanofi KK, Pfizer Japan, Janssen Pharmaceutical KK, Chugai Pharmaceutical, Asahi Kasei Pharma, Astellas Pharma and Mitsubishi Tanabe Pharma. MH has received payment for postmarketing surveillance from GlaxoSmithKline KK, a research grant from Novartis Pharma, and honoraria for lectures from GlaxoSmithKline KK, AstraZeneca KK and Astellas Pharma. FG was a Director on the Board of the Australian Rheumatology Association at the time of the study. ZZ has received payment/honoraria from AbbVie, AstraZeneca KK, Boehringer Ingelheim, Bristol Myers Squibb, Eli Lilly, GSK, Novartis, Pfizer, Roche, Sanofi, Janssen and UCB Pharma, and has participated in advisory boards for BeiGene. TT has received consultancy fees from AstraZeneca, Kowa and Mitsubishi Tanabe. KN has received speaker fees from Novartis. YT has received speaker fees and/or honoraria from AbbVie, Eisai, Chugai, Eli Lilly, Boehringer Ingelheim, GlaxoSmithKline, Taisho, AstraZeneca, Daiichi-Sankyo, Gilead, Pfizer, UCB, Asahi Kasei and Astellas, and received research grants from Boehringer Ingelheim, Taisho and Chugai. Y-HC has received advisory board fees and honoraria from Pfizer, Novartis, AbbVie, Johnson & Johnson, BMS, Roche, Lilly, GSK, AstraZeneca, Sanofi, MSD, Guigai, Astellas, Inova Diagnostics, UCB, Agnitio Science Technology, United Biopharma, Thermo Fisher, Gilead, Eisai and CSL Behring, as well as research grants from the Taiwan Ministry of Science and Technology, Taiwan Department of Health, Taichung Veterans General Hospital, National Yang-Ming University, GSK, Pfizer, BMS, Roche and AstraZeneca, and Medigen Vaccine Biologics. JChoi is an employee of BMS. RK-R has received grants from GSK and Novartis. EM has received consulting fees from AbbVie, AstraZeneca, Biogen, Bristol Myers Squibb, Eli Lilly, EMD Serono, Genentech, Gilead, Janssen, Novartis, Takeda and UCB.

Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.

Provenance and peer review Not commissioned; externally peer reviewed.

Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.

References

¹ Arnaud L, Tektonidou MG. Long-term outcomes in systemic lupus erythematosus: trends over time and major contributors. Rheumatology (Oxford) 2020; 59: v29–38. doi:10.1093/rheumatology/keaa382

² Ibañez D, Gladman DD, Urowitz MB. Adjusted mean Systemic Lupus Erythematosus Disease Activity Index-2K is a predictor of outcome in SLE. J Rheumatol 2005; 32: 824–7.

³ Gavan S, Bruce I, Payne K. Generating evidence to inform health technology assessment of treatments for SLE: a systematic review of decision-analytic model-based economic evaluations. Lupus Sci Med 2020; 7: e000350. doi:10.1136/lupus-2019-000350

⁴ Bojke L, Manca A, Asaria M, et al. How to Appropriately Extrapolate Costs and Utilities in Cost-Effectiveness Analysis. Pharmacoeconomics 2017; 35: 767–76. doi:10.1007/s40273-017-0512-6

⁵ Ibañez D, Urowitz MB, Gladman DD. Summarizing disease features over time: I. Adjusted mean SLEDAI derivation and application to an index of disease activity in lupus. J Rheumatol 2003; 30: 1977–82.

⁶ Györi N, Giannakou I, Chatzidionysiou K, et al. Disease activity patterns over time in patients with SLE: analysis of the Hopkins Lupus Cohort. Lupus Sci Med 2017; 4: e000192. doi:10.1136/lupus-2016-000192

⁷ Choi MY, Chen I, Clarke A, et al. 1704 Identifying clusters of longitudinal autoantibody profiles associated with systemic lupus erythematosus disease outcomes. Lupus Sci Med 2021; 8: A70. doi:10.1136/lupus-2021-lupus21century.99

⁸ Watson P, Brennan A, Birch H, et al. An integrated extrapolation of long-term outcomes in systemic lupus erythematosus: analysis and simulation of the Hopkins lupus cohort. Rheumatology (Oxford) 2015; 54: 623–32. doi:10.1093/rheumatology/keu375

⁹ Clarke AE, St-Pierre Y, Paly V, et al. POS0734 EXTRAPOLATION OF LONG-TERM OUTCOMES IN SYSTEMIC LUPUS ERYTHEMATOSUS: REPLICATING A HOPKINS LUPUS COHORT ANALYSIS WITH THE SYSTEMIC LUPUS INTERNATIONAL COLLABORATING CLINICS (SLICC) INCEPTION COHORT. Ann Rheum Dis 2021; 80 (Suppl 1): 617–618. doi:10.1136/annrheumdis-2021-eular.1790

¹⁰ Kandane-Rathnayake R, Golder V, Louthrenoo W, et al. Development of the Asia Pacific Lupus Collaboration cohort. Int J Rheum Dis 2019; 22: 425–33. doi:10.1111/1756-185X.13431

¹¹ Hochberg MC. Updating the American College of Rheumatology revised criteria for the classification of systemic lupus erythematosus. Arthritis Rheum 1997; 40: 1725. doi:10.1002/art.1780400928

¹² Petri M, Orbai A-M, Alarcón GS, et al. Derivation and validation of the Systemic Lupus International Collaborating Clinics classification criteria for systemic lupus erythematosus. Arthritis Rheum 2012; 64: 2677–86. doi:10.1002/art.34473

¹³ Gladman D, Ginzler E, Goldsmith C, et al. The development and initial validation of the Systemic Lupus International Collaborating Clinics/American College of Rheumatology damage index for systemic lupus erythematosus. Arthritis Rheum 1996; 39: 363–9. doi:10.1002/art.1780390303

¹⁴ Gladman DD, Ibañez D, Urowitz MB. Systemic lupus erythematosus disease activity index 2000. J Rheumatol 2002; 29: 288–91.

¹⁵ Isenberg DA, Allen E, Farewell V, et al. An assessment of disease flare in patients with systemic lupus erythematosus: a comparison of BILAG 2004 and the flare version of SELENA. Ann Rheum Dis 2011; 70: 54–9. doi:10.1136/ard.2010.132068

¹⁶ Chessa E, Piga M, Floris A, et al. Use of Physician Global Assessment in systemic lupus erythematosus: a systematic review of its psychometric properties. Rheumatology (Oxford) 2020; 59: 3622–32. doi:10.1093/rheumatology/keaa383

¹⁷ Wooldridge JM. Introductory Econometrics: A Modern Approach. 7th edn. Boston: Cengag, 2020.

¹⁸ Hausman JA. Specification Tests in Econometrics. Econometrica 1978; 46: 1251–1271. doi:10.2307/1913827

¹⁹ Li N. SLE disease activity stabilizes over time: a hypothesis stimulated by a model reported a few years ago and recently re-tested. Rheumatology (Oxford) 2023; 62: 2338–40. doi:10.1093/rheumatology/kead029

²⁰ Connelly K, Golder V, Kandane-Rathnayake R, et al. Clinician-reported outcome measures in lupus trials: a problem worth solving. Lancet Rheumatol 2021; 3: e595–603. doi:10.1016/S2665-9913(21)00119-3