Abstract
Objective.
SLE is marked by immune dysregulation linked to varied clinical disease activity. Using a unique longitudinal cohort of SLE patients, this study seeks to identify optimal immune mediators informing an empirically refined Flare Risk Index (FRI) reflecting altered immunity prior to clinical disease flare.
Methods.
SLE-associated plasma mediators (n=37) were evaluated by microfluidic immunoassay in 46 Pre-Flare and 53 Pre-Nonflare samples selected from a unique longitudinal cohort of 106 patients with classified SLE (meeting ACR and SLICC criteria). Autoantibody specificities, hybrid SLEDAI (hSLEDAI) scores, clinical features, and medication usage were also compared at Pre-Flare (111±47 days prior to Flare) vs. Pre-Nonflare (99±21 days prior to Nonflare) time points. Variable importance was determined by random forest with logistic regression subsequently applied to determine the optimal number and type of analytes informing a refined FRI.
Results.
Pre-Flare vs. Pre-Nonflare differences were not associated with demographics, autoantibody specificities, hSLEDAI scores, clinical features, nor medication usage. Forward selection and backward elimination of mediators ranked by variable importance drew up to 17 candidates differentiating Pre-Flare from Pre-Nonflare. A final combination of 11 mediators best informed a newly refined FRI, achieving a maximum of 97% sensitivity and 98% specificity after applying decision curve analysis to define low, medium, and high FRI scores.
Conclusion.
We verified altered immune mediators associated with imminent disease flare; a subset improved the FRI to identify SLE patients at risk of imminent flare. This molecularly-informed proactive management approach could be critical in prospective clinical trials and the clinical management of lupus.
Keywords: SLE, disease flare, cytokines
INTRODUCTION
Systemic lupus erythematosus (SLE) is a chronic, often debilitating, autoimmune disease characterized by immune dysregulation with a varied, waxing and waning clinical disease course. Despite validated clinical disease activity instruments and improved treatment strategies, SLE patients still suffer end-organ damage (1). In addition to patients with waxing/waning and clinically active disease, SLE patients with quiescent disease are also at risk of clinical disease flare (2). Early recognition and treatment of disease flares are necessary to avoid the need for major immunosuppressants that carry significant side effects, and to prevent irreversible tissue and organ damage associated with poor SLE outcomes (3). Requiring considerable medical resources, clinical disease flares result in average healthcare costs per year in excess of $20,000 (inpatient, outpatient, and pharmacy) per patient (4).
Novel tools are needed to propagate proactive, steroid-sparing strategies to prevent or decrease the severity of disease flares, end-organ damage, and the pathogenic and socioeconomic burdens of SLE. Currently, no flare prediction test exists; traditional serological markers of disease activity are ineffective at predicting flare (reviewed in (5)). We have previously demonstrated in retrospective, cross-sectional SLE samples the alteration of plasma immune mediators from multiple pathways within 12 weeks of clinical disease flare that can inform a flare risk index (FRI) capable of differentiating Pre-Flare from comparable Pre-Nonflare visits. (6, 7). The purpose of the current study is to refine the FRI for clinical use, informed by prospectively collected plasma samples from an ongoing unique longitudinal SLE cohort assessed on an automated microfluidic immunoassay platform compatible with near-patient use.
METHODS
Study population, clinical data, and sample collection
Experiments were performed in accordance with the Helsinki Declaration and approved by Institutional Review Boards of the Oklahoma Medical Research Foundation (OMRF) and the Mayo Clinic. Study participants meeting American College of Rheumatology and Systemic Lupus International Collaborating Clinics classification for SLE (8) were enrolled in an ongoing prospective observational study after written informed consent (Table S1); patients were assessed at baseline and quarterly thereafter for one year. Demographic and clinical information were collected, including disease activity as defined by the hybrid Systemic Lupus Erythematosus Disease Activity Index (hSLEDAI) (9). Flare was defined by the SELENA-SLEDAI Flare Index (SFI) (10), as described in the Supplementary Methods.
Pre-Flare visits were designated as the quarterly visit immediately preceding the visit where an SFI-defined flare was recorded. Pre-Nonflare/nonflare visits were similarly designated during a corresponding period of nonflare. Blood was collected at each clinical visit; plasma and sera were isolated, aliquoted, and stored at −80°C until assessment (6, 7).
Plasma immune mediator determination
Fully automated microfluidic immunoassays for 37 immune mediators (Table S2) were performed by the CAP certified, CLIA approved laboratory at Progentec Diagnostics on the Protein Simple/Bio-techne (San Jose, CA) Ella platform according to manufacturer instructions (11). Briefly, freshly thawed, centrifuged (10,000*g for five minutes) human plasma aliquots were diluted in sample diluent, with equal distribution across cartridges by cohort, race, age, sex, and clinical disease activity/flare status to minimize batch effects. Diluted samples, quality controls (QCs), and buffer were loaded into each cartridge. High range QCs, generated, aliquoted, and frozen according to manufacturer instructions, were thawed and diluted appropriately to generate medium and low range QCs at time of cartridge run. All QCs were processed in parallel with plasma aliquots to minimize handling effects and performed within manufacturer-designated concentration ranges. Triplicate immunoassay reactions were performed for every analyte/sample. Raw (background-subtracted) signal levels were automatically back-fit to barcode-embedded standard curves, then multiplied by user-defined dilution factors to provide mean calculated concentrations (pg/ml) for each analyte.
Serum autoantibody specificity determination
Serum samples were assessed for autoantibody specificities, including SLE-associated specificities toward dsDNA, chromatin, Ro/SSA, La/SSB, Sm, SmRNP complex, and RNP by xMAP BioPlex 2200 (Bio-Rad Technologies, Hercules, CA) (12). Semi-quantitative values for anti-dsDNA were reported as IU/mL (positive ≥10 IU/mL). All other specificities were reported in autoantibody index (AI) units (range 0–8) based on the fluorescence intensity per manufacturer-specified cutoff (positive ≥1 AI). Factor XIIIb levels served as both a serum confirmation and an indicator of sample integrity.
Statistical Analysis
Power analyses.
To achieve > 80% statistical power with α=0.05 and prevalence=0.5 (50% Pre-Flare, 50% Pre-Nonflare) (13), ≥ 34 visits/samples (17 Pre-Flare and 17 Pre-Nonflare) were required for the current study to inform a single-entity FRI score based on FRI scores calculated in previous retrospective studies, with effect size = 2.18 (6, 7). To ensure > 80% statistical power while comparing Pre-flare and Pre-nonflare visits across a maximum of 37 individual soluble mediator FRI subscores (Bonferroni corrected α=0.0014), ≥ 68 visits/samples (34 Pre-Flare and 34 Pre-Nonflare) were required for the current study.
Univariate analyses.
Associations between categorical variables were analyzed by generalized estimating equations (GEE) to account for repeated measures. Binomial distributions and logit link functions were employed for binary outcomes and multinomial distributions with generalized logit link functions for outcomes with more than two categories. Linear mixed models with subject specific random intercepts were used for comparisons of two groups of continuous data (e.g. Pre-Flare vs. Pre-Nonflare; Flare vs. Nonflare). Kruskal-Wallis test with Dunn’s multiple comparison correction was used for comparisons of three or more groups (e.g. Mild/Moderate vs. Severe vs. Nonflare). Correlations between FRI scores and hSLEDAI scores were determined by Spearman correlation. Receiver operating characteristic (ROC) analysis was used to determine area under the curve (AUC) and flare prediction significance. Effect size was determined using Cohen’s d (14), the mean difference between Pre-Nonflare (M1) and Pre-Flare (M2) groups divided by the pooled standard deviation ([M2-M1]/SDpooled, where SDpooled = √[{SD12+SD22}/2]). Multiple comparison p-values were corrected using Bonferroni correction, dividing an alpha of 0.05 (minimum significant p-value) by number of statistical tests for a given category of variables.
Multivariate analyses.
Random forest partition tree classification based on Genuer et al. (15), able to incorporate repeated measures (16), was implemented to rank variables in their ability to differentiate Pre-Flare from Pre-Nonflare visits. Settings used included maximum number of trees = 50 and maximum number of variables to consider for splitting a node = 100, with 5-fold cross validation conducted across 2,000 iterations. Variables were ordered from most to least informative based on variable importance. Variables were then applied to the FRI by adding one at a time (starting with most informative variable; forward selection) or removing one at a time from the total list of variables (starting with least informative variable; backward elimination) (15) to optimally predict flare outcome. Spearman correlation and Principal Component Analysis were used to assess correlation among variables.
Flare Risk Index (FRI).
The overall level of inflammation in Pre-Flare vs. Pre-Nonflare SLE clinic visits in relationship to disease activity at time of future concurrent flare or non-flare was compared using the FRI, derived by the cumulative contribution of n (1–37) number of log-transformed and standardized plasma mediators assessed at time of Pre-Flare/Pre-Nonflare weighted in relationship to hSLEDAI disease activity at flare/nonflare, as previously described (6, 7). Step 1: log-transformed plasma analytes were standardized = (observed value) - (mean value of informative samples [pre-flare and pre-nonflare])/(standard deviation of informative samples [pre-flare and pre-nonflare]). Step 2: Spearman correlation coefficients were calculated for each pre-flare/pre-nonflare analyte vs. flare/nonflare hSLEDAI disease activity. Step 3: log-transformed, standardized analyte levels were weighted by the respective Spearman coefficients, resulting in analyte FRI subscores. Step 4: analyte subscores informing the FRI were summed (total FRI score) (6, 7).
Logistic regression was performed to determine threshold probabilities for flare risk based on FRI scores in Pre-Flare vs. Pre-Nonflare samples. Low/medium and medium/high risk FRI score cutoffs were further determined using decision curve analysis, as previously described (17). Briefly, the threshold probability of flare risk for each Pre-Flare/Pre-Nonflare FRI score was compared to net benefit, such that “net benefit” = True positive count/n – (False positive count/n * [pt/1-pt]), where n is the total number of patients in the study and pt is the (predictive) threshold probability for any given FRI score. To create the decision curve, the threshold probability (pt) is varied to cover the range of threshold probabilities associated with the Pre-Flare and Pre-Nonflare FRI scores included in the analysis. For each pt: 1. Define a Pre-flare visit as true positive if Pre-flare pt ≥ selected pt; 2. Define a Pre-Nonflare visit as false positive if Pre-Nonflare pt ≥ selected pt; 3. Calculate the number of true and false positives for a given pt; 4. Calculate net benefit True positive count/n – (False positive count/n * [pt/1-pt]) (17). In addition, the trajectory of the FRI-11 and informing mediators were visualized over time in representative flare and nonflare SLE patients with sufficient accrued visits.
GEE and linear mixed model analyses were performed using SAS, Version 9.4, while other univariate analyses, logistic regression, and decision curve graphing were performed using GraphPad Prism 8.4.3 (GraphPad Software, San Diego, CA). Multivariate random forest was performed using JMP® Genomics, Version 9. SAS Institute Inc., Cary, NC, 1989–2021. Power analyses were performed using MedCalc 20.027 (MedCalc Software Ltd, Ostend, Belgium).
RESULTS
Select Immune Mediators Inform a Lupus Flare Risk Index
SLE patients from a unique cohort were followed quarterly at the Mayo Clinic (n=30) or OMRF (n=76) and evaluated for evidence of hSLEDAI-defined disease flare, Table S1. Forty-six samples from 42 subjects were available within 12 weeks of impending flare (Pre-Flare) and 53 samples from 39 subjects available within a comparable non-flare window (Pre-Nonflare), Table 1 and Figure 1, with subjects providing multiple visits differing by disease activity and clinical features at each visit. As shown in Table 1, neither demographics, SLE-associated AutoAb specificities, medication, nor hSLEDAI scores and organ system manifestations were significantly different between Pre-Flare and Pre-Nonflare visits after adjusting for multiple comparison. This includes the immunologic manifestations of increased anti-dsDNA binding and decreased complement levels.
Table 1.
Demographic and Clinical Features of Baseline Pre-Flare/Pre-Nonflare SLE Patient visits
Pre-Flare/Pre-Nonflare SLE Patients | Pre-Flare (n=42) | Pre-Nonflare (n=39) | p-value a |
| |||
Female (n, %) | 39 (93%) | 35 (90%) | 0.4729 |
Race (n, %) | 0.6607 | ||
European American | 28 (67%) | 26 (67%) | Reference |
African American | 5 (12%) | 6 (15%) | 0.6607 |
Native American | 5 (12%) | 4 (10%) | 0.8117 |
Otherb | 4 (9%) | 3 (8%) | 0.7523 |
| |||
Number of Visits | Pre-Flare (n=46) | Pre-Nonflare (n=53) | p-value a |
| |||
Age at Visit (mean, SD) | 46 (15) | 45 (16) | 0.9475 |
Days to Follow-up (mean, SD) | 111 (47) | 99 (21) | 0.1135 |
# AutoAb Specificitiesc (mean, SD) | 1.9 (2.2) | 1.5 (1.9) | 0.6948 |
AutoAb positivityc (n, %) | |||
Chromatin | 16 (35%) | 11 (21%) | 0.1604 |
Ro/SSA | 19 (41%) | 17 (32%) | 0.7894 |
La/SSB | 5 (11%) | 3 (6%) | 0.4472 |
Sm | 10 (22%) | 9 (17%) | 0.8598 |
SmRNP | 14 (30%) | 17 (32%) | 0.7715 |
RNP | 13 (28%) | 17 (32%) | 0.5582 |
Medicationsd (n, %) | |||
Belimumab | 4 (9%) | 7 (13%) | 0.3173 |
Steroid | 24 (52%) | 34 (64%) | 0376.6937 |
Hydroxychloroquine | 32 (70%) | 41 (77%) | 0.3425 |
Immunosuppressants | 26 (56%) | 32 (60%) | 0.5455 |
| |||
hSLEDAIe (mean/SD/range) | 2.5 (2.6 [0–10]) | 2.8 (3.4 [0–16]) | 0.1187 |
Organ System Manifestations (n, %) f | |||
Arthritis | 3 (7%) | 10 (19%) | 0.0376 |
Renal | 2 (4%) | 3 (6%) | 0.5339 |
Urinary Casts | 0 | 0 | -- |
Hematuria | 0 | 1 (2%) | -- |
Proteinuria | 2 (4%) | 2 (4%) | 0.9565 |
Pyuria | 0 | 0 | -- |
Mucocutaneous | 15 (33%) | 16 (30%) | 0.7698 |
Rash | 11 (24%) | 9 (17%) | 0.5328 |
Alopecia | 6 (13%) | 5 (9%) | 0.5783 |
Mucosal Ulcers | 3 (7%) | 6 (11%) | 0.5938 |
Serositis | 0 | 2 (4%) | -- |
Pleurisy | 0 | 2 (4%) | -- |
Pericarditis | 0 | 0 | -- |
Immunologic | 19 (41%) | 17 (32%) | 0.6150 |
Low Complement | 11 (24%) | 8 (15%) | 0.1887 |
C3 levels (mg/dL, mean, SD) | 111 (27) | 110 (26) | 0.9759 |
C4 levels (mg/dL, mean, SD) | 22 (9) | 22 (8) | 0.8494 |
Increased DNA Binding | 14 (30%) | 12 (23%) | 0.8392 |
anti-dsDNA levels (IU/mL, mean, SD) | 27 (74) | 8 (19) | 0.0730 |
Hematologic | 4 (9%) | 5 (9%) | 0.7734 |
Thrombocytopenia | 0 | 2 (4%) | -- |
Leukopenia | 4 (9%) | 3 (6%) | 0.3995 |
Continuous variables significance by linear mixed models; Categorical variable significance by GEE (unadjusted significant p≤0.05)
Other = Asian, Hispanic, Multi-race, and Unknown
# of AutoAb specificities include anti-dsDNA, chromatin, Ro/SSA, La/SSB, Sm, SmRNP, and RNP (Bonferroni corrected significant p-value for multiple comparison =0.0071)
Immunosuppressants = azathioprine, methotrexate, sirolimus, tacrolimus, mycophenolate mofetil, cyclophosphamide
hSLEDAI = SELENA-SLEDAI + proteinuria defined by SLEDAI-2K
Bonferroni corrected significant p-value for multiple comparison (Arthritis, Renal, Mucocutaneous, Serositis, Immunologic, and Hematologic organ systems manifestations) = 0.0083
Figure 1. Immune mediators informing a Flare Risk Index (FRI).
(A) Variable Importance of top 17 immune mediators differentiating Pre-Flare from Pre-Nonflare as determined by random forest (top 12 mediators in blue and remaining top 5 mediators in red as determined by forward selection and backward elimination, respectively); ranking of all immune mediators and demographic/clinical features can be found in Table S2); (B) FRI (Mean ± 95% CI) in Pre-Flare (top panel) vs. Pre-Nonflare (bottom panel) samples informed by top number of mediators in variable importance order from (A); (C) Performance of FRI informed by top 12 (FRI-12) vs 17 (FRI-17) mediators listed in (A); graphs presented as box plots ± maximum and minimum scores. **p<0.01; ***p<0.001, ****p<0.0001 Mann-Whitney test
These findings were confirmed in random forest multivariate analysis given their relatively low level of importance alongside the plasma levels of 37 immune mediators, Table S2. Thus, only soluble mediators were included in further flare predictive analyses. Random forest variable importance (Figure 1A and Table S2) revealed a combination of 12–17 mediators that best differentiated Pre-Flare from Pre-Nonflare samples, with additional variables adding minimal predictive value. These mediators were applied to the FRI, informed by log-transformed and standardized plasma mediators, weighted by the hSLEDAI scores at time of the subsequent, quarterly visit where the associated Flare/Nonflare event occurred, Figure 1B and Table S3. Both the FRI-12 and FRI-17, informed by the top 12 or 17 mediators respectively, significantly differentiated Pre-Flare from Pre-Nonflare samples, with effect size of 0.941 (FRI-12) vs. 0.817 (FRI-17) and AUC = 0.756 (FRI-12, p<0.0001) vs. 0.727 (FRI-17, p=0.0001). In addition, both FRI-12 and FRI-17 correlated with the hSLEDAI score at time of future concurrent flare/non-flare, with Spearman r = 0.440 (FRI-12) vs. 0.435 (FRI-17, p<0.0001), Figure 1C, greater than any individual contributing soluble mediator, including the most informative mediator, OPN (Spearman r = 0.212, p=0.0065), Table S3.
Refinement of Flare Risk Index for Clinical Application
FRI performance combined with technical feasibility and cost-effectiveness of running the laboratory-based test was subsequently considered to refine which mediators inform a refined FRI for clinical application. A number of analyte combinations were considered, the top five outlined in Figure S1, with an 11-analyte combination striking a balance between technical feasibility, cost-effectiveness, and performance (Figure 2). The FRI-11 configuration, requiring two microfluidic immunoassay cartridges (at 1:10 or 1:2 sample dilutions, Figure 2A), maximized the 32-sample inlet x 4 single-plex (1:10) or multi-plex (1:2) channel configuration of the Ella immunoassay cartridges. Although MMP-9 ranked ninth among the analytes by random forest analysis, Figure 1A, we dropped this analyte as the only 1:100 dilution candidate, with no 32-sample inlet cartridges available to accommodate MMP-9 as a single 1:100 dilution analyte. Despite its lower random forest ranking, BLyS maximized the 1:10 cartridge containing OPN, TNFRI, and TNFRII, as the highest ranking/performing 1:10 dilution analyte (Table S2 and Figure S1) with this substitution having no difference in overall model performance.
Figure 2. Flare Risk Index (FRI) designed for clinical/commercial use.
(A) Immune mediators (n=11) informing FRI-11 (bold mediators determined by 1:10 sample dilution, all others determined by 1:2 dilution on a separate Ella microfluidic immunoassay cartridge); (B) FRI-11 differentiation of Pre-Flare vs. Pre-Nonflare samples from Fig. 1 (****p<0.0001 Mann-Whitney test, graph presented as box plot ± maximum, minimum); (C) FRI-11 differentiation of Pre-Flare (Severe vs. Mild/Moderate [Mild/Mod]) vs. Pre-Nonflare samples in (B) (**p<0.01, ****p<0.0001 Kruskal-Wallis test with Dunn’s multiple comparison, graph presented as mean ± 95% CI); (D) Decision Curve Analysis plotting Threshold Probability vs. calculated Net Benefit; Low (24%) and High (76%) Probability marked with dashed lines; overlaid decision curves: black = Pre-Flare + Pre-Nonflare, blue = Pre-Nonflare, red = Pre-Flare (E) Predictive probability of future disease flare determined by logistic regression; Low (−6.4) and High (9.0) FRI-11 cut-off levels (associated with 24% and 76% Threshold Probability, respectively) marked with dashed lines; (F) Performance characteristics of FRI-11 (AUC = Area Under the Curve [receiver operating characteristic curve analysis]), CI = confidence interval, PPV = positive predictive value, NPV = negative predictive value
Of utmost importance, FRI-11 was able to differentiate Pre-Flare vs. Pre-Nonflare visits (p<0.0001, Figure 2B), including those of European American (Pre-Flare=3.53 [mean; 1.17, 5.88 95% CI]; Pre-Nonflare=−1.5 [−3.83, 0.83], p=0.0025) or non-European American ancestry (Pre-Flare = 2.26 [−1.77, 6.30]; Pre-Nonflare = −5.33 [−7.28, −3.38], p=0.0005). SLE patients at risk of imminent severe or mild/moderate flare were also distinguished, Figure 2C. To improve clinical utility, decision curve analysis (Figure 2D) was applied in the context of logistic regression (Log Odds of Flare = −0.13 + 1.53*Pre-Flare; Figure 2E) to determine low/medium (FRI-11 = −6.4; 24% probability of flare) and medium/high (FRI-11 = 9.0; 76% probability of flare) risk cutoffs for imminent flare. Decision curve analysis determines the net benefit of clinical intervention in a lupus patient population (Figure 2D, y-axis) along the threshold probability (Figure 2D, x-axis) of imminent clinical disease flare. The lower FRI-11 limit for Pre-Flare (24% probability of flare; equivalent to an FRI-11 score of −6.4, Figure 2E) provided the greatest net benefit (0.47) and maximum sensitivity (97%, with NPV = 94%, Figure 2F), while the upper FRI-11 limit for Pre-Flare (76%; equivalent to an FRI-11 score of 9.0, Figure 2E) provided a net benefit of 0.07 with maximum specificity (98%, with PPV = 88%; Figure 2F). A positive/negative FRI-11 cutoff of 0 was indicative of 47% risk (threshold probability) of imminent disease flare, with a net benefit of 0.30.
Overall, the FRI-11 performed well, with a large (>0.8 (18)) effect size of 0.907, AUC = 0.755 ([0.660, 0.849], p<0.0001) and correlation with hSLEDAI scores at the future concurrent Flare/Nonflare visit (Spearman r = 0.434 [0.253, 0.585], p<0.0001), Figure 2F. Using a positive/negative cut-off, SLE patients were five times more likely to have an imminent flare with a positive FRI-11 score. While the FRI-11 contained correlated analytes (Table S4) that were not necessarily significant by univariate analysis (Table S5), the 11 mediators informing the refined FRI were necessary to account for the immunologic heterogeneity noted in the SLE patients assessed (Table S6) and contributed to varied parameters of FRI-11 performance (Figure S2).
Performance of Flare Risk Index at Time of Flare and in Longitudinal Samples
In addition to differentiating Pre-Flare/Pre-Nonflare, we also assessed the ability of the FRI to differentiate concurrent Flare/Nonflare in 52 and 120 available samples from 47 and 57 subjects, respectively, Table 2 and Figure 3, with multiple sample subjects differing by disease activity and clinical features at each visit. Similar to Pre-Flare/Pre-Nonflare visits (Table 1), neither demographic, SLE-associated AutoAb specificities, nor medication variables distinguished SLE patients at time of concurrent disease flare vs. nonflare. As expected, visits where SLE patients experienced flare had higher hSLEDAI scores (p<0.0001), significant for arthritis (p<0.0001) and mucocutaneous (p=0.0007) manifestations, compared to nonflare visits, Table 2. Immunologic manifestations, including increased anti-dsDNA binding and low complement, were similar between concurrent flare and nonflare. At the time of flare/nonflare, the FRI-11 determined significant differences between Severe and Mild/Moderate flare visits and Nonflare visits, Figure 3A, with Mild/Moderate flares most notable for arthritis (63%), cutaneous (24%), and nasopharyngeal (20%) features, while severe flares were most noted for qualifying increase in PGA (67%), prednisone doubling + nephritis (33%), and new SLE medication or hospitalization (33%), Table 2. Of interest, the FRI-11 distinguished Pre-Flare/Pre-Nonflare or Flare/Nonflare irrespective of organ system involvement (Figure 3B–G), including differences between flare and nonflare in SLE patients with or without arthritis (Figure 3B), renal (Figure 3C), mucocutaneous (Figure 3D), serositis (Figure 3E), immunologic (Figure 3F), or hematologic (Figure 3G) manifestations.
Table 2.
Demographic and Clinical Features of Follow-up Flare/Nonflare SLE Patient visits
Flare/Nonflare SLE Patients | Flare (n=47) | Nonflare (n=57) | p-value a |
| |||
Female (n, %) | 43 (91%) | 51 (89%) | 0.3391 |
Race (n, %) | |||
European American | 31 (66%) | 36 (63%) | Reference |
African American | 6 (13%) | 11 (19%) | 0.2959 |
Native American | 6 (13%) | 4 (7%) | 0.2892 |
Otherb | 4 (8%) | 6 (11%) | 0.6250 |
| |||
Number of Visits | Flare (n=52) | Nonflare (n=120) | p-value a |
| |||
Age at Visit (mean, SD) | 47 (15) | 45 (15) | 0.1822 |
Days from Baseline (mean, SD) | 113 (47) | 100 (24) | 0.0351 |
# AutoAb Specificitiesc (mean, SD) | 2.0 (2.2) | 1.6 (1.9) | 0.1460 |
AutoAb positivityc (n, %) | |||
Chromatin | 21 (40%) | 30 (25%) | 0.0653 |
Ro/SSA | 21 (40%) | 41 (34%) | 0.7551 |
La/SSB | 5 (10%) | 7 (6%) | 0.6655 |
Sm | 10 (19%) | 19 (16%) | 0.6305 |
SmRNP | 18 (35%) | 36 (30%) | 0.3110 |
RNP | 17 (35%) | 34 (28%) | 0.2879 |
Medicationsd (n, %) | |||
Belimumab | 2 (4%) | 10 (8%) | 0.9168 |
Steroid | 28 (54%) | 67 (56%) | 0.1455 |
Hydroxychloroquine | 36 (69%) | 94 (78%) | 0.4675 |
Immunosuppressants | 28 (54%) | 66 (55%) | 0.7658 |
| |||
hSLEDAIe (mean/SD/range) | 6.8 (3.6 [2–24]) | 3.2 (3.3 [0–16]) | <0.0001 |
Organ System Manifestations (n, %) f | |||
Arthritis | 35 (67%) | 30 (25%) | <0.0001 |
Renal | 8 (15%) | 9 (7%) | 0.0864 |
Urinary Casts | 0 | 0 | -- |
Hematuria | 2 (4%) | 2 (2%) | 0.4582 |
Proteinuria | 6 (12%) | 7 (6%) | 0.1476 |
Pyuria | 1 (2%) | 0 | -- |
Mucocutaneous | 31 (60%) | 36 (30%) | 0.0007 |
Rash | 17 (33%) | 23 (19%) | 0.1945 |
Alopecia | 12 (23%) | 12 (10%) | 0.0660 |
Mucosal Ulcers | 11 (21%) | 10 (8%) | 0.0426 |
Serositis | 3 (6%) | 8 (7%) | 0.8002 |
Pleurisy | 3 (6%) | 7 (6%) | 0.9822 |
Pericarditis | 0 | 1 (1%) | -- |
Immunologic | 21 (40%) | 37 (31%) | 0.2320 |
Low Complement | 16 (31%) | 21 (18%) | 0.1289 |
C3 levels (mg/dL, mean, SD) | 113 (31) | 112 (26) | 0.2141 |
C4 levels (mg/dL, mean, SD) | 21 (9) | 22 (9) | 0.7863 |
Increased DNA Binding | 14 (27%) | 26 (22%) | 0.9603 |
anti-dsDNA levels (IU/mL, mean, SD) | 18 (54) | 12 (40) | 0.3816 |
Hematologic | 7 (13%) | 12 (10%) | 0.2432 |
Thrombocytopenia | 1 (2%) | 2 (2%) | 0.8454 |
Leukopenia | 7 (13%) | 10 (8%) | 0.1412 |
| |||
M/M Flare (n=46) | Severe Flare (n=6) | All Flare (n=52) | |
| |||
Mild to Moderate (M/M) Flare Features (n, %) | |||
Cutaneous | 11 (24%) | N/A | 11 (21%) |
Nasopharyngeal | 9 (20%) | N/A | 9 (17%) |
Pleuritis | 3 (7%) | N/A | 3 (6%) |
Arthritis | 29 (63%) | N/A | 29 (56%) |
Fever | 1 (2%) | N/A | 1 (2%) |
Prednisone increase (not to >0.5 mg/kg/day) | 5 (11%) | N/A | 5 (10%) |
PGA increase (>1, not more than 2.5) | 6 (13%) | N/A | 6 (12%) |
| |||
Severe Flare Features (n, %) | |||
Prednisone doubling + Nephritis | N/A | 2 (33%) | 2 (4%) |
Prednisone doubling + Thrombocytopenia | N/A | 1 (17%) | 1 (2%) |
Prednisone increase (>0.5 mg/kg/day) | N/A | 1 (17%) | 1 (2%) |
New SLE Medication or Hospitalization | N/A | 2 (33%) | 2 (4%) |
PGA Increase (>2.5) | N/A | 4 (67%) | 4 (8%) |
Continuous variables significance by linear mixed model analysis; Categorical variable significance by GEE analysis (unadjusted significant p≤0.05)
Other = Asian, Hispanic, Multi-race, and Unknown
# of AutoAb specificities include anti-dsDNA, chromatin, Ro/SSA, La/SSB, Sm, SmRNP, and RNP (Bonferroni corrected significant p-value for multiple comparison =0.0071)
Immunosuppressants = azathioprine, methotrexate, sirolimus, tacrolimus, mycophenolate mofetil, cyclophosphamide
hSLEDAI = SELENA-SLEDAI + proteinuria defined by SLEDAI-2K
Bonferroni corrected significant p-value for multiple comparison (Arthritis, Renal, Mucocutaneous, Serositis, Immunologic, and Hematologic organ systems manifestations) = 0.0083
Figure 3. Flare Risk Index (FRI) at time of concurrent Flare vs. Nonflare by organ system manifestations.
(A) FRI-11 differentiation of Flare (Severe vs. Mild/Moderate [Mild/Mod]) vs. Nonflare samples; (B-G) FRI-11 differentiation of Flare vs. Nonflare (NF) by Arthritis (B), Renal (C), mucocutaneous (Mucocut, D), Serositis (E), Immunologic (F), or Hematologic (G) organ system manifestations (*p≤0.05, **p<0.01, ***p<0.001, ****p<0.0001 Kruskal-Wallis test with Dunn’s multiple comparison, graphs presented as mean ± 95% CI).
Because each SLE patient has a unique disease course, we evaluated the trajectory of the FRI-11 over time in SLE patients with sufficient accrued visits in our longitudinal cohort (representative patients displayed in Figure S3). SLE patients with quiescent disease were more likely to have “low” range FRI-11 scores (Figure S3A, upper panel), with informing mediators near or below 0 (Figure S3A, middle and lower panels). A non-flaring patient with non-quiescent disease may have slightly positive, “medium” range, FRI-11 scores at/near 0 (Figure S3B, upper panel, visits 0–3), while increased FRI-11 score accompanied by an increase in multiple informing mediators advocates the risk of future flare (Figure S3B, visit 4). SLE patients with a single flare in 12 months of follow-up were also likely to have “medium” FRI-11 scores positive for multiple informing mediators with imminent flare (Figure S3C). However, SLE patients with recurrent flares were more likely to have “high” FRI-11 scores that did not decrease as the patient went in/out of flare (Figure S3D), reflected by a number of positive informing mediators that changed minimally over time (Figure S3D, lower panel).
DISCUSSION
A pro-active approach is needed to manage immune dysregulation and improve outcomes in SLE. Delays in treating SLE flares may potentiate chronic inflammation and end-organ damage, with increased morbidity and early mortality. Conversely, patients with long-standing quiescent disease may benefit from reduction in immunosuppressant medication associated with significant side effects and potential organ damage. Validated disease activity instruments, including the hSLEDAI, assess and weigh clinical organ system changes (5). However, disease flares are only detected after immune system changes have occurred that result in uncontrolled inflammation and the accrual of tissue and end-organ damage (6, 7, 19).
No serologic prognostic tool currently exists to identify SLE patients at risk of imminent disease flare; the lack of an immune mechanism-informed disease management test in SLE stems from no individual immune pathway-informed biomarker acting as a universal surrogate (5). Classical serologic markers of disease activity are not necessarily sufficient biologic signals nor are they necessarily altered prior to impending disease flare. Used alone or in combination, anti-dsDNA, complement, complement split products, and inflammatory markers (ESR and CRP) have been relatively limited in their ability to predict flare due to their heterogeneous presence (e.g. only 40–70% of SLE patients are anti-dsDNA positive) and inconsistent correlation with disease activity (reviewed in (5)). Our current and previous studies (6, 7) showed a similar pattern, with limited and insignificant difference in the hSLEDAI immunologic manifestations of increased anti-dsDNA binding and low complement between Pre-Flare and Pre-Nonflare visits or Flare vs. Nonflare visits, whether assessed as categorical or continuous variables.
The 11-analyte combination (FRI-11) chosen for clinical and commercial application struck a balance between performance, technical feasibility, and cost-effectiveness. Unlike our previous retrospective studies (6, 7) that utilized samples assessed using hypothesis-generating, research and development (R&D) xMAP and ELISA immunoassay platforms, standard of care management of lupus necessitates a prospective approach. Additionally, R&D immunoassay platforms suffer from inter-lot, inter-vendor, and inter-platform variability (20). Despite the challenge of selecting visits/samples from a prospectively assessed cohort where patients were not followed in lock-step nor flare rates predicted to have an equal number of pre-flare and pre-nonflare visits, we confirmed that immune changes occurred prior to clinical disease flare, whether assessed en masse or on an individual patient level, and this information was harnessed to inform a refined FRI. The microfluidic, automated Ella immunoassay platform, designed for commercial use with minimal inter-user, inter-site, and inter-lot batch effects (11, 21–23), allowed for design of the refined FRI for clinical application.
Narrowing the list of top performing analytes to a combination that maximized the use of one each 1:10 and 1:2 dilution microfluidic immunoassay cartridges, nine of the eleven mediators informing the FRI-11 were previously shown to distinguish Pre-Flare from Pre-Nonflare samples in European American and/or African American SLE patients (6, 7). Of particular interest are our findings regarding osteopontin (OPN), ranking highest in random forest variable importance analysis and univariately distinguishing Pre-Flare from Pre-Nonflare samples in the current study. In addition to being readily detected in plasma in ng/ml quantities, OPN is associated with SLE genetic risk (24), increased clinical disease activity (25), and promoted activation and migration of antigen presenting cells, including macrophages and dendritic cells, as well as differentiation of multiple T-helper cell pathways, including Th1, Th17, and Tfh (reviewed in (5)). Conversely, the TNF superfamily member, BLyS, is less consistent in its ability to differentiate Pre-Flare from Pre-Nonflare samples in our past (6, 7) and current studies. A high-value therapeutic target in SLE, a subset of patients with increased BLyS levels demonstrating strong interferon (IFN) gene signatures respond well to belimumab with delayed clinical disease flare (26). BLyS is driven by both type I and type II IFNs (reviewed in (5)). In the current study, in addition to correlating with the IFN-inducing chemokine MCP-1, BLyS levels also correlated with other IFN-associated mediators, including IFN-γ, IP-10, and MIG (data not shown). BLyS ranked higher in multivariate analysis than most of these IFN-associated mediators, possibly due to its downstream upregulation that may capture more Pre-Flare SLE patients.
Moving from reactively managing clinical disease flare, defined by the SFI (10) or the BILAG Index (27), to a proactive, interventional approach informed by fluctuating immune profiles, even within the same patient, is a paradigm shift that requires consideration. The refined FRI-11 distinguished Pre-Flare from Pre-Nonflare visits by both univariate and ROC analysis (p<0.0001), with a large (>0.8) (18) Cohen d effect size of 0.907. While these statistical parameters suggest utility for anticipating imminent disease flare, they are not able to conclusively identify where along the FRI spectrum to proceed with flare prevention and/or early intervention measures.
Decision curve analysis focuses on the net benefit for intervention (28, 29) given a particular threshold probability of, in the current study, imminent clinical disease flare based on the FRI. Introduced in 2006 (30), net benefit as a critical component of decision curve analysis is widely supported (28, 29, 31, 32) and now recommended by the TRIPOD guidelines for prediction models (29, 33). The highest net benefit, capturing the greatest number of SLE patients for early intervention, was at 24% threshold probability, with a net benefit of 0.47 at an FRI score of −6.4. Yet having a negative FRI score may not meet an intervention threshold by providers. On the other end of the spectrum, utilizing a very conservative approach with far fewer SLE patients captured for early intervention, a threshold probability of 76% gave a net benefit of 0.07 at an FRI score of 9.0. Yet, it is these patients that would be at greatest risk for imminent severe flare, as noted in the results of the current study. A “middle” approach might be early intervention coinciding with a positive FRI score (cutoff = 0), whereby there is a 47% threshold probability and a net benefit of 0.30. Ultimately, the threshold probability for early intervention reflects the risk of imminent flare at which providers are indifferent about changing treatment (34).
Limitations in the current study include a minimal number of SLE patients who experienced severe flares and a cohort enriched for non-organ threatening disease, resulting in fewer SLE patients with renal and serositis-associated organ system manifestations. Expanding our cohort in future studies to include an increased number of ethnic minorities and SLE patients with a history of organ-threatening disease would allow for greater likelihood of encompassing both severe flares and renal and serositis manifestations. Future longitudinal crossover studies to include standard of care vs. alternative intervention strategies would also likely improve the utility of the FRI for providers. These could include early intervention approaches based on an observed increase in FRI scores or the effect of steroid withdrawal/use of steroid sparing treatments on FRI scores and flare rates over time. This would be in addition to prospective, longitudinal assessment of patients with varied clinical features to validate our findings ((6, 7) and current study) that clinical and serological features do not contribute to flare risk within the same patient. These future studies would allow for further enhancement and validation of the FRI using the TRIPOD checklist (33).
The ultimate goal of the FRI is to provide a tool to guide early intervention and prevent accrual of permanent organ damage associated with clinical disease flares (35), decrease the need for high dose steroids (36), and decrease medical costs for SLE patients and their providers (37). Providers must balance risk of flare vs. risk of infection (35) and other detrimental side effects and/or toxicities from lupus treatments (35). Of interest, one of the most prominently published risks of disease flare is lack of treatment compliance (35, 36, 38–40). With the availability of hydroxychloroquine level testing and other digital and health coach resources (40, 41), compliance monitoring may be a beneficial early intervention without escalating treatment, particularly for those patients with a positive FRI below 76% threshold probability for imminent flare.
In this way, the ability to identify patients at risk of impending flare could lead to improved adherence of current treatment regimens, as well as discern the timing of disease suppression therapy and contribute to more effective and efficient clinical trial designs. This could improve patient outcomes and reduce the pathogenic and socioeconomic burdens of SLE (42). An advantage of calculating a patient’s FRI is that it does not require cut-offs for each soluble mediator to establish positivity, nor does it require a priori knowledge of which inflammatory mediators may contribute to flare in a particular patient. Data from our current study, combined with our previous reports (6, 7), suggest the refinement of the FRI calculation to a single, clinically actionable, algorithm is applicable across organ system manifestations, correlating with increased disease activity at future time of concurrent flare, with higher FRI scores associated with increased probability for impending disease flare.
Supplementary Material
ACKNOWLEDGEMENTS
We would like to thank the OMRF Rheumatology Research Center of Excellence personnel, the OMRF Biorepository personnel, Mayo Clinic Rochester Rheumatology department personnel, and all of the study participants for their time and commitment to the study.
Grant support: This study was supported by the National Institute of Arthritis and Musculoskeletal and Skin Diseases, the National Institute of Allergy and Infectious Diseases, and the National Institute of General Medicinal Sciences of the National Institutes of Health under award numbers R01AR077518 (HZ), R44AI142967 (MEM), P30AR073750 (JAJ), UM1AI144292 (JAJ), and U54GM104938 (JAJ), as well as Oklahoma Center for the Advancement of Science and Technology under award numbers AR16–014 (MEM) and AR18–019 (EJ).The content of this publication is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Disclosure: Salary (MEM and DB) and research support (MEM, UT, and JAJ) have been received from Progentec Diagnostics, Inc. DD, MP, and EJ are full-time employees with Progentec Diagnostics, Inc.
REFERENCES
- 1.Bruce IN, O’Keeffe AG, Farewell V, Hanly JG, Manzi S, Su L, et al. Factors associated with damage accrual in patients with systemic lupus erythematosus: results from the Systemic Lupus International Collaborating Clinics (SLICC) Inception Cohort. Ann Rheum Dis. 2015;74(9):1706–13. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Steiman AJ, Gladman DD, Ibanez D, Urowitz MB. Prolonged serologically active clinically quiescent systemic lupus erythematosus: frequency and outcome. J Rheumatol. 2010;37(9):1822–7. [DOI] [PubMed] [Google Scholar]
- 3.Koelmeyer R, Nim HT, Nikpour M, Sun YB, Kao A, Guenther O, et al. High disease activity status suggests more severe disease and damage accrual in systemic lupus erythematosus. Lupus science & medicine. 2020;7(1). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Kan HJ, Song X, Johnson BH, Bechtel B, O’Sullivan D, Molta CT. Healthcare utilization and costs of systemic lupus erythematosus in Medicaid. BioMed research international. 2013;2013:808391. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Thanou A, Jupe E, Purushothaman M, Niewold TB, Munroe ME. Clinical disease activity and flare in SLE: Current concepts and novel biomarkers. J Autoimmun. 2021;119:102615. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Munroe ME, Vista ES, Guthridge JM, Thompson LF, Merrill JT, James JA. Pro-inflammatory adaptive cytokines and shed tumor necrosis factor receptors are elevated preceding systemic lupus erythematosus disease flare. Arthritis & rheumatology. 2014;66(7):1888–99. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Munroe ME, Vista ES, Merrill JT, Guthridge JM, Roberts VC, James JA. Pathways of impending disease flare in African-American systemic lupus erythematosus patients. J Autoimmun. 2017;78:70–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Ines L, Silva C, Galindo M, Lopez-Longo FJ, Terroso G, Romao VC, et al. Classification of Systemic Lupus Erythematosus: Systemic Lupus International Collaborating Clinics Versus American College of Rheumatology Criteria. A Comparative Study of 2,055 Patients From a Real-Life, International Systemic Lupus Erythematosus Cohort. Arthritis care & research. 2015;67(8):1180–5. [DOI] [PubMed] [Google Scholar]
- 9.Thanou A, Stavrakis S, Dyer JW, Munroe ME, James JA, Merrill JT. Impact of heart rate variability, a marker for cardiac health, on lupus disease activity. Arthritis Res Ther. 2016;18:197. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Buyon JP, Petri MA, Kim MY, Kalunian KC, Grossman J, Hahn BH, et al. The effect of combined estrogen and progesterone hormone replacement therapy on disease activity in systemic lupus erythematosus: a randomized trial. Ann Intern Med. 2005;142(12 Pt 1):953–62. [DOI] [PubMed] [Google Scholar]
- 11.Dysinger M, Marusov G, Fraser S. Quantitative analysis of four protein biomarkers: An automated microfluidic cartridge-based method and its comparison to colorimetric ELISA. J Immunol Methods. 2017;451:1–10. [DOI] [PubMed] [Google Scholar]
- 12.Bruner BF, Guthridge JM, Lu R, Vidal G, Kelly JA, Robertson JM, et al. Comparison of autoantibody specificities between traditional and bead-based assays in a large, diverse collection of patients with systemic lupus erythematosus and family members. Arthritis Rheum. 2012;64(11):3677–86. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Rosner B Fundamentals of biostatistics. 7th ed. Boston, MA: Brooks/Cole, Cengage Learning; 2011. [Google Scholar]
- 14.Cohen J A power primer. Psychological bulletin. 1992;112(1):155–9. [DOI] [PubMed] [Google Scholar]
- 15.Genuer R, Poggi JM, Tuleau-Malot C. Variable selection using random forests. Pattern Recogn Lett. 2010;31(14):2225–36. [Google Scholar]
- 16.Capitaine L, Genuer R, Thiebaut R. Random forests for high-dimensional longitudinal data. Stat Methods Med Res. 2021;30(1):166–84. [DOI] [PubMed] [Google Scholar]
- 17.Vickers AJ, Cronin AM, Elkin EB, Gonen M. Extensions to decision curve analysis, a novel method for evaluating diagnostic tests, prediction models and molecular markers. BMC Med Inform Decis Mak. 2008;8:53. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Cohen J Statistical Power Analysis for the Behavioral Sciences. 2nd ed: Lawrence Erlbaum Associates; 1988. [Google Scholar]
- 19.Lu R, Guthridge JM, Chen H, Bourn RL, Kamp S, Munroe ME, et al. Immunologic findings precede rapid lupus flare after transient steroid therapy. Scientific reports. 2019;9(1):8590. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Leng SX, McElhaney JE, Walston JD, Xie D, Fedarko NS, Kuchel GA. ELISA and multiplex technologies for cytokine measurement in inflammation and aging research. J Gerontol A Biol Sci Med Sci. 2008;63(8):879–84. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Aldo P, Marusov G, Svancara D, David J, Mor G. Simple Plex() : A Novel Multi-Analyte, Automated Microfluidic Immunoassay Platform for the Detection of Human and Mouse Cytokines and Chemokines. American journal of reproductive immunology. 2016;75(6):678–93. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Purushothama S, Dysinger M, Chen Y, Osterlund K, Mora J, Chunyk AG, et al. Emerging technologies for biotherapeutic bioanalysis from a high-throughput and multiplexing perspective: insights from an AAPS emerging technology action program committee. Bioanalysis. 2018;10(3):181–94. [DOI] [PubMed] [Google Scholar]
- 23.Cao J, Seegmiller J, Hanson NQ, Zaun C, Li D. A microfluidic multiplex proteomic immunoassay device for translational research. Clin Proteomics. 2015;12:28. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Han S, Guthridge JM, Harley IT, Sestak AL, Kim-Howard X, Kaufman KM, et al. Osteopontin and systemic lupus erythematosus association: a probable gene-gender interaction. PLoS One. 2008;3(3):e0001757. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Wirestam L, Enocsson H, Skogh T, Padyukov L, Jonsen A, Urowitz MB, et al. Osteopontin and Disease Activity in Patients with Recent-onset Systemic Lupus Erythematosus: Results from the SLICC Inception Cohort. J Rheumatol. 2019;46(5):492–500. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Wilkinson C, Henderson RB, Jones-Leone AR, Flint SM, Lennon M, Levy RA, et al. The role of baseline BLyS levels and type 1 interferon-inducible gene signature status in determining belimumab response in systemic lupus erythematosus: a post hoc meta-analysis. Arthritis Res Ther. 2020;22(1):102. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Gordon C, Sutcliffe N, Skan J, Stoll T, Isenberg DA. Definition and treatment of lupus flares measured by the BILAG index. Rheumatology (Oxford). 2003;42(11):1372–9. [DOI] [PubMed] [Google Scholar]
- 28.Vickers AJ, Van Calster B, Steyerberg EW. Net benefit approaches to the evaluation of prediction models, molecular markers, and diagnostic tests. Bmj. 2016;352:i6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Van Calster B, Wynants L, Verbeek JFM, Verbakel JY, Christodoulou E, Vickers AJ, et al. Reporting and Interpreting Decision Curve Analysis: A Guide for Investigators. Eur Urol. 2018;74(6):796–804. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26(6):565–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Rousson V, Zumbrunn T. Decision curve analysis revisited: overall net benefit, relationships to ROC curve analysis, and application to case-control studies. BMC Med Inform Decis Mak. 2011;11:45. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Steyerberg EW, Vickers AJ. Decision curve analysis: a discussion. Med Decis Making. 2008;28(1):146–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. Bmj. 2015;350:g7594. [DOI] [PubMed] [Google Scholar]
- 34.Pauker SG, Kassirer JP. Therapeutic decision making: a cost-benefit analysis. N Engl J Med. 1975;293(5):229–34. [DOI] [PubMed] [Google Scholar]
- 35.Adamichou C, Bertsias G. Flares in systemic lupus erythematosus: diagnosis, risk factors and preventive strategies. Mediterranean Journal of Rheumatology. 2017;28(1):4–12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Arnaud L, Tektonidou MG. Long-term outcomes in systemic lupus erythematosus: trends over time and major contributors. Rheumatology (Oxford). 2020;59(Suppl5):v29–v38. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Hammond ER, Desta B, Near AM, Wang X, Jiang M. Frequency, severity and costs of flares increase with disease severity in newly diagnosed systemic lupus erythematosus: a real-world cohort study, United States, 2004–2015. Lupus science & medicine. 2021;8(1). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Zeng X, Zheng L, Rui H, Kang R, Chen J, Chen H, et al. Risk factors for the flare of systemic lupus erythematosus and its influence on prognosis: a single-center retrospective analysis. Adv Rheumatol. 2021;61(1):43. [DOI] [PubMed] [Google Scholar]
- 39.Costedoat-Chalumeau N, Houssiau F, Izmirly P, Guern VL, Navarra S, Jolly M, et al. A Prospective International Study on Adherence to Treatment in 305 Patients With Flaring SLE: Assessment by Drug Levels and Self-Administered Questionnaires. Clinical pharmacology and therapeutics. 2019;106(2):374–82. [DOI] [PubMed] [Google Scholar]
- 40.Hsu CY, Lin YS, Cheng TT, Syu YJ, Lin MS, Lin HF, et al. Adherence to hydroxychloroquine improves long-term survival of patients with systemic lupus erythematosus. Rheumatology (Oxford). 2018;57(10):1743–51. [DOI] [PubMed] [Google Scholar]
- 41.Marengo MF, Waimann CA, de Achaval S, Zhang H, Garcia-Gonzalez A, Richardson MN, et al. Measuring therapeutic adherence in systemic lupus erythematosus with electronic monitoring. Lupus. 2012;21(11):1158–65. [DOI] [PubMed] [Google Scholar]
- 42.Lau CS, Mak A. The socioeconomic burden of SLE. Nat Rev Rheumatol. 2009;5(7):400–4. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.