Risk prediction models for incident systemic lupus erythematosus among women in the Nurses' health study cohorts using genetics, family history, and lifestyle and environmental factors.

The characteristics of the 138 women who developed incident SLE and the 1136 controls who did not in these NHS/NHSII analyses are shown in Table 1 . The NHS cohorts are both > 95% of European ancestry and those selected from the biobank for this nested study were > 98% European ancestry. Average age at SLE onset among cases was approximately 50 years and approximately 52 years at the matched index date among the controls. Approximately one-quarter of the SLE cases were current smokers, and 19% reported a positive family history of SLE or RA. Their initial SLE clinical manifestations are also summarized in Table 1 . The four sequential models we developed to predict SLE risk among women yielded the following AUCs: 1) wGRS 0.63 (95%CI 0.58-0.68), 2) SLE/RA family history alone 0.64 (95%CI 0.59-0.68), 3) selected lifestyle, reproductive and environmental factors 0.71 (95%CI 0.66-0.75), 4) wGRS, family history and all lifestyle/environmental risk factors with p< 0.10 in model 3, 0.76 (95%CI 0.72-0.81) ( Figure 1 ). Our final model thus included: SLE wGRS, family history, current cigarette smoking, low alcohol consumption, obesity, depression, early age at menarche, and oral contraceptive use as defined above, as well and NHS/NHSII cohort. The beta estimates, standard errors, and odds ratios for these eight SLE-related risk factors in the most complete model, model 4, are shown in Table 2 . Age of diagnose, race/ethnicity, postmenopausal status and hormone use, sleep duration, residential income, U.S. geographic residential region, and wGRS × smoking interaction were not included as they were not strong predictors in this case-control dataset and did not improve prediction accuracy. The estimated optimism of the AUC for model 4 was 0.016, providing an optimism-corrected AUC of 0.75. In our similar sequential models to predict anti-dsDNA+ SLE, AUCs were 0.60 (95% CI 0.51-0.69), 0.63(95% CI 0.56-0.70), 0.81(95% CI 0.75-0.87) and 0.82 (95% CI 0.76-0.88) respectively, with optimism of AUC estimated at 0.028, providing optimism-corrected AUC of 0.79 for the most complete model predicting anti-dsDNA + SLE with all factors considered ( Figure 2 ).

We utilized genetic and self-reported data from both the NHS (started in 1976) and NHS II (started in 1989; both collectively referred to here as the NHS). The NHS cohorts are longitudinal female cohort studies in which data on lifestyle, behavioral factors and disease outcomes have been collected every two years by questionnaire 39 . The NHS began in 1976 and enrolled 121,700 registered female nurses ages 30-55 from 11 U.S. states; the NHSII began in 1989 and enrolled 116,430 registered female nurses ages 25-42 from 14 U.S. states. The preponderance of the NHS cohorts’ participants is of European ancestry, given the demographics of the nursing profession during the years of enrollment. Participants were asked to report new physician diagnoses of SLE on each biennial questionnaire. Those indicating a new diagnosis were asked to complete the Connective Tissue Disease Screening Questionnaire and to consent to the release of their medical records 53 . Medical records of all nurses who indicated SLE symptoms on this questionnaire and agreed to release them to the NHS were independently reviewed by two board-certified rheumatologists. Incident SLE cases were identified based on four or more of the 1997 Updated American College of Rheumatology (ACR) criteria for the classification of SLE and reviewers’ consensus 54 - 56 . (Cases were collected and validated 1978-2017, thus prior to the publication of the 2019 European League against Rheumatism/ACR criteria for SLE classification 57 , and the performance of the three currently used sets of criteria (1997, 2012 SLICC and 2019 EULAR/ACR) are highly similar with essentially identical sensitivity and specificity for SLE 57 .) For each case in the NHS, ten controls without any history of reporting a connective tissue disease and with available GWAS data, were selected, matched on age at index date of SLE diagnosis (within four years), self-reported race, and genotyping platform. In the NHSII, four controls were similarly matched to each SLE case (as fewer women had genotyping results available in NHSII). We considered factors that have been associated with SLE in our past analyses as potential predictors, in addition to race/ethnicity, income level, and geographic region of residence 37 - 47 . Lifestyle and environmental exposure variables were assessed on NHS questionnaires at least one cycle prior to the SLE index date (matched date in controls), corresponding to exposures prior to disease onset for the cases or matched date for controls. We examined the following updated health factors, assessed via biennial questionnaire: age, race (non-European vs. European ancestry), updated body mass index (obese > 35 kg/m 2 vs. non-obese ≤ 35 kg/m 2 ), cigarette smoking (current/recently quit within 4 years vs. past/never), alcohol consumption (cumulative average of daily intake in grams, ≤ vs. >5 grams/day), cumulative average sleep duration (hours per day as a continuous variable), history of major depression (antidepressant use, a physician’s diagnosis of depression, or a Mental Health Index-5 [MHI-5 score <60 indicating probable depression vs. no depression) 58 , updated household income (low vs. high, dichotomized by zip-code level median household income), U.S .geographic region of residence (Mid-Atlantic, Midwest, New England, Southeast, West), age at menarche (< vs. ≥ 10 years old), oral contraceptive use (current or past use vs. never), and post-menopausal status and hormone use (pre-menopausal, post-menopausal-never, postmenopausal-current, postmenopausal-past). Family history of RA or SLE was self-reported in the NHS in 2008 and in the NHSII in both 2013 and 2017. Any RA or SLE family history was defined as self-reported RA/SLE cases in family members (the NHS specified first-degree family members, and the NHSII did not specify). Missing values were kept in a missing group for categorical variables, and the sample mean was used for continuous variables. Except for family history, all time-varying cumulative updated information through the questionnaire of one cycle (two years) prior to SLE onset for cases, and the same questionnaire cycle for their matched controls, was employed for our prediction modeling. Our continuous weighted Genetic Risk Score (wGRS) for SLE was recently derived and included 86 SNPs previously reported associated with SLE risk with genome-wide significance (p ≤ 5 × 10 −8 ) and ten classical HLA alleles were included in the wGRS using SLE association results from Langefeld et al 17 , 59 . We employed it here as a continuous variable of increasing SLE risk. Demographic characteristics of cases and controls were described using means and standard deviations for continuous variables and frequency and proportions for categorical variables. The odds ratios for SLE were estimated using logistic regression for all factors including lifestyle variables, family history of SLE or RA, and SLE wGRS as a continuous variable. Given our past work on gene-smoking interactions in SLE risk using a larger dataset 33 , we also tested wGRS × smoking interactions for SLE prediction in our model. Multivariable logistic regression models were built to predict SLE risk. We started with the complete covariate list above, including previously reported SLE risk factors, traditional confounders and matching factors (age, race). Four sequential models were generated: 1) SLE wGRS, 2) family history of SLE or RA, 3) all lifestyle and environmental factors as described in methods, and 4) combining the risk factors from models 1-3 and ultimately including only the most significant and strongest model 3 risk factors given our current and past findings 37 - 40 , 42 - 44 , 48 and to avoid overfitting given relatively few numbers of incident cases and many potential predictors that may be correlated. Similar models were also developed to predict anti-double stranded DNA (dsDNA)-positive SLE. The discriminatory abilities of all models to define case group vs. control group at different combinations of sensitivity and specificity were assessed using a Receiver Operating Characteristic (ROC) curve and computing the Area Under Curve (AUC). Since we did not have access to an external dataset with all the variables that we considered prior to the onset of SLE with non-SLE controls, we carried out an internal validation procedure using correction for optimism. The bootstrap-based optimism correction procedure was used to limit the possibility of overfitting in the prediction models to obtain an optimism-corrected ROC AUC 60 . Bootstrapping was done with 500 replications to obtain a stable estimate of optimism. The dataset was repeatedly resampled with replacement to 500 datasets. For each random sampled dataset, the prediction model was fitted. Each fitted model was then applied both to the resampled data and the original data. The C statistic was calculated for both, and the optimism estimate was the average difference in these AUCs. Optimism was subtracted from the initial AUC estimate from the full model to calculate optimism-corrected AUC. All analyses were performed on SAS Version 9.4 (SAS Institute, Cary, NC).

In this study nested within the large longitudinal female NHS cohorts, we have used a combination of lifestyle, reproductive risk factors, family history, and wGRS to develop quite accurate models for SLE risk prediction. We found that a combination of readily assessed variables and a SLE wGRS led to 76% (75% after optimism correction) accurate prediction of SLE risk in women. In our best-performing model combining all of these, with optimism-corrected AUC of 0.75, some modifiable SLE risk factors, i.e., smoking and obesity, were no longer strongly predictive, possibly due to collinearity with other variables such as depression. These results suggest that our increasing knowledge about risk factors for SLE could lead to the identification of those at highest risk, and potentially then to early interventions prior to the onset of symptoms , to intercept and prevent this often-devastating disease. Genetic predisposition and exposure to environmental stimuli are thought to interact in SLE pathogenesis. A complex and enigmatic disease, SLE is characterized by autoantibody production, complement activation and immune complex deposition, producing inflammation and damage in tissues. SLE develops over a period of years; autoantibodies (including antinuclear antibodies, ANAs, and more specific SLE-related autoantibodies) are present years prior to SLE symptoms and identify SLE subtypes 61 , 62 . Type I interferon (IFN), as well as tumor necrosis factor (TNF) α, interleukin (IL)4, 5, 6, and IFN-γ-induced protein (IP)10, are involved in SLE pathogenesis and elevations are detectable years before clinical SLE 50 , 51 . In a study of banked Department of Defense blood samples, Th1, Th2 and Th17-cytokines were found to be dysregulated in 84 subjects who later developed SLE compared to matched controls 50 . In that study, ANA and anti-Ro/SSA antibody positivity, with high levels of IL-5, IL-6, and the monokine induced by IFN-γ (MIG), distinguished future SLE patients with 92% accuracy. It is not known whether this SLE-related autoimmunity can be halted or reversed once detected in high-risk individuals. The main limitation of this work is that the NHS cohorts are almost entirely of European ancestry. It will be important to test and validate these and other SLE risk prediction models in younger and more diverse populations, including males and younger individuals as well. We were also unable to include other known and suspected risk factors, such as PTSD, child abuse and endometriosis, which have only been assessed in the NHSII, but not the NHS cohort, or have not been assessed in these cohorts, and pesticides, mercury and silica exposure 41 , 45 , 47 , 63 , 64 . Additionally, family history of SLE or RA was assessed as a composite question in both cohorts and in different years and in slightly different manners resulting in a high proportion of participants (16%) missing these data. Our study suggests that increasing knowledge about risk factors for SLE may allow the identification of those at highest risk, or at least quantification of their risk, and potentially then to early interventions prior to the onset of symptoms . Accurate risk prediction models could enable earlier and more accurate identification of at-risk patients in the general population for disease interception and design of prevention trials. The results of SLE risk prediction models based on early signs and symptoms, autoantibodies and cytokine/chemokine signatures may be another option, but require more intensive screening and laboratories than those we have developed and might be used as a second step for predicting more imminent SLE 50 , 51 . Here we show that using available genetic and population risk factors, SLE risk can be quantified fairly accurately. Moreover, many of these risk factors- e.g. smoking, obesity, and oral contraceptive use-- are modifiable and might then lead to important conversations about potential disease prevention through environmental or lifestyle change strategies, or early and safe therapeutic interventions such as vitamin D or omega-3 fatty acids, which have been shown to reduce the risk of autoimmune disease overall 65 .

Systemic lupus erythematosus (SLE) is a chronic autoimmune disease characterized by widespread immune dysregulation, systemic inflammation and, in many cases, progressive and irreversible organ damage 1 , 2 . Although SLE is relatively uncommon, it disproportionately affects females, who represent approximately 90% of cases, during their reproductive years and remains a leading cause of mortality in young women, underscoring its impact as an important public health issue 3 , 4 . SLE can affect multiple organ systems, including the joints, kidneys, central nervous system, skin, heart, and lungs 5 . Delays in diagnosis and treatment are associated with higher disease activity, more organ damage, healthcare utilization and mortality 6 - 10 . Currently, therapies are still inadequate for many patients, and morbidity and mortality remain unacceptably high 3 , 4 . Both genetic and environmental factors influence SLE susceptibility. SLE genome-wide association studies (GWAS) have identified > 100 risk loci 11 - 23 . Most GWAS have included those of European descent and most individual variants have small effect sizes, with odds ratios < 1.5, each explaining a small fraction of genetic risk of SLE 24 - 28 . HLA genes, Class I, II and II, including classical loci (HLA-DRB1*01:02, *03:01, *08:01 and HLA-DQA1 *01:02), as well as many single nucleotide polymorphisms (SNPs) are associated with SLE risk 29 - 31 . For genes including IRF5, STAT4, TNFAIP3 and PTPN22 , our understanding of the effects of polymorphisms on immune function is advanced 32 . Far less is known about many SNPs in non-protein coding regions that may influence gene expression 24 . Weighted genetic risk scores (GRS) have been developed by our group and others to estimate an individual’s cumulative genetic susceptibility to SLE risk 33 - 35 . Having a high SLE GRS has been associated with earlier onset SLE and more severe disease phenotypes 34 . Monozygotic twin studies estimate SLE heritability ~24%, leaving much to be explained by environmental exposures 24 , 36 . As SLE is multifactorial and multigenic, an individual’s risk for SLE development cannot be well estimated using only known genetic risk factors. In the Nurses’ Health Study (NHS) cohorts and the Black Women’s Health Study (BWHS), we have identified a range of factors related to increased SLE risk among women: airborne exposures (e.g., current or recently ceased cigarette smoking); psychosocial factors (e.g., major depression, sleep deprivation, child abuse and post-traumatic stress disorder [PTSD]); reproductive and hormonal risk factors among women (e.g., early menarche, oral contraceptives, and postmenopausal hormones); and others potentially related to SLE through their effects in systemic inflammation (e.g., obesity and low alcohol intake -the only exposure associated with decreased risk) 37 - 49 . To date, there have been efforts to develop models to predict SLE among patients presenting with potential early SLE symptoms, and among those who have a family history of the disease, based on autoantibodies and cytokines 50 - 52 . However, to our knowledge no models have been developed to predict SLE based on a wide range of previously identified risk factors. Thus, we undertook to use existing genetic and collected lifestyle, environmental, reproductive, exposure, and family history data from the large NHS cohorts to develop prediction models of SLE risk and assess their performance.