- Top of page
- PATIENTS AND METHODS
- AUTHOR CONTRIBUTIONS
We have shown that over the course of 1 year, in a large cohort of patients with SLE, periods of PAD were more common than flare episodes, affecting nearly half of the patients. This is in keeping with the findings of Barr et al, who have also shown that chronic activity is a common disease state in SLE (15).
Flare is a term used to imply a clinically significant increase in disease activity compared with baseline. In terms of the instruments used to measure disease activity in SLE such as the SLEDAI-2K, the British Isles Lupus Assessment Group (BILAG), or the Safety of Estrogens in Lupus Erythematosus: National Assessment version of the SLEDAI (SELENA–SLEDAI), this is reflected in an increase in disease activity score (13, 16, 17). The current definitions of flare most often used are an increase in SLEDAI-2K score of ≥4 points, an increase in SELENA–SLEDAI score of ≥3 points, or 1 new category A or 2 new category B items on the BILAG score (13, 16, 17). SLEDAI-2K scores range from 0–105, with higher scores indicating more active disease. In a study of 230 patients followed in the Toronto Lupus Clinic, flare was defined by new or increased therapy for active disease, an expression of concern, or use of the term flare in the physician's notes (13). When the clinician assessed that the patient was experiencing a flare, the SLEDAI score increased by a median of 4. Based on this study, we defined flare as an increase in SLEDAI-2K score of ≥4 from the previous visit, with a minimum interval between visits of 2 months. Based on our definition of flare, we defined PAD as a SLEDAI-2K score of ≥4, excluding serology alone, on ≥2 consecutive visits with a minimum interval between visits of 2 months. In a given time interval, a patient may experience both flare episodes and periods of PAD. Because the definition of PAD entails a minimum interval of 2 months between 2 active visits, there may be a concern that a visit that is caught in the tail end of a flare episode may be incorrectly deemed PAD. However, among 176 patients with ≥2 visits designated PAD in 2004, on average, PAD episodes lasted a mean ± SD of 6.5 ± 2.6 months, which is much longer than the expected time taken for the resolution of a flare episode. An additional consideration may be that when there are several months between 2 visits designated PAD, one may in fact be dealing with a flare episode that has resolved in the interval before the beginning of a second flare episode. However, 65% of visits in 2004 were between 2 and 4 months apart, and the median time interval between visits in 2004 was 3.4 months, which is shorter than the expected time taken for one flare episode to resolve and a second to begin.
In this study we found that the annual incidence of flare and PAD were similar in 2004 and 2005. One-third of the patients experienced ≥1 flare per year, whereas half of the patients experienced ≥1 PAD episode. Nearly 60% of patients had ≥1 episode of either flare or PAD per year. Of note, ≥25% of patients had PAD without achieving the definition of flare in each year. These findings suggest that by using flare as the primary outcome variable in clinical trials, we fail to capture a clinically important outcome, namely PAD. PAD may serve as an outcome variable in SLE clinical trials in one of two ways. First, PAD may be an inclusion criterion whereby the therapeutic efficacy of a drug may be determined based on its ability to improve or resolve PAD. Alternatively, the outcome of interest may be prevention of PAD. In addition, the expected frequencies of flare and PAD per annum as observed in this study may serve a useful role in guiding the design of clinical trials, in particular the selection of the duration of treatment or followup required to ascertain the outcomes of interest.
There were no major differences in the organ systems involved in flare and PAD. The most common manifestations of flare and PAD were, in rank order, immunologic, cutaneous, renal, musculoskeletal, and neurologic, reflecting the most commonly involved organ systems in SLE (18).
We sought to determine whether in the natural history of SLE each of the outcomes of interest, namely flare or PAD, could be predicted from the previous pattern of disease in each individual. The notion of predicting the future course of disease from past events in a chronic disease as complex as SLE is intriguing and perhaps ambitious. However, if possible, this modeling of disease course has the potential to inform prognosis and guide followup and monitoring of patients. Furthermore, these models may serve a useful role in the selection of patients who are likely to flare or have PAD for inclusion in clinical trials. In the past, many therapeutic trials in SLE have been underpowered to show efficacy. Selection of the patients most likely to exhibit the outcomes of interest could potentially reduce sample size requirements.
We chose variables for inclusion in the prediction models based on existing knowledge of the determinants of course and prognosis in SLE. For example, a rise in anti-dsDNA antibody levels has been shown to herald a flare, whereas treatment with steroids and immunosuppressive agents is a surrogate for active disease (19). We have previously derived the AMS as a measure of cumulative disease activity, and shown that it is associated with the presence of damage and coronary artery disease (14, 20).
Significant predictors of PAD identified in our models were SLEDAI-2K score at the start of the outcome interval as well as cutaneous and musculoskeletal manifestations in the preceding time interval. Based on goodness-of-fit statistics, model 1 was the best prediction model for PAD (deviance 205.4, P = 0.33). In this model, SLEDAI-2K score has an OR of 1.34 (P < 0.0001), meaning that for each 1-point increase in SLEDAI-2K score at the start of the outcome interval, the odds of having PAD during the outcome interval increased by 34%. Similarly, ORs of 2.37 and 3.98 indicate that the presence of cutaneous or musculoskeletal manifestations in the preceding year increased the odds of having PAD during the outcome interval by 237% and 398%, respectively. In model 1, hematologic manifestations in the preceding year were protective, reducing the odds of having ≥1 PAD episode in the following year by 72% (OR 0.28, P = 0.035).
The association between PAD and SLEDAI-2K score at the start of the outcome interval appears intuitive, in that more disease activity at the outset forebodes yet more disease activity during followup. We postulate that the association between PAD and prior cutaneous or musculoskeletal disease activity might be due to the fact that these non–life-threatening manifestations may be less aggressively treated and therefore more likely to be persistent and in time accompanied by other features of active disease. Likewise, the protective effect of prior hematologic disease activity may be due to this manifestation being treated aggressively, thus reducing the likelihood of further disease activity during followup. However, of note, cumulative prednisone dose in the prior interval was neither implicated in nor protective of PAD in the outcome interval.
The properties of prediction models for PAD were tested using resampling and an outcome time interval that was not included in the derivation of the models. Intuitively, one might expect models that incorporate the most comprehensive past information on disease characteristics to perform best in predicting future events. However, prediction models for PAD using data from 1, 2, and 3 years prior to the outcome year (2005) all performed equally well, with areas under the ROC curves of ∼0.80. Overall, model 1 performed better than the other models by a small margin, with an overall accurate prediction of 78.5%. The sensitivity and specificity for model 1 were 70.6% and 84.3%, respectively. A PPV of 76.6% indicates that a patient with a positive prediction has a nearly 77% chance of having a period of PAD in the following year. This is substantially greater than the 50% background prevalence of PAD. A NPV of 79.7% indicates that a patient with negative prediction has a nearly 80% chance of remaining event-free. Again, this is substantially greater than the background 50% probability of remaining free of PAD. Whereas sensitivity and specificity are properties inherent to the prediction models, PPV and NPV are related not only to sensitivity and specificity, but also to prevalence. In this study, the PPVs and NPVs were based on a PAD prevalence of 46.1% and a flare prevalence of 28.0%, defined as the proportion of patients that had PAD or flare in 2005, as presented in Table 1.
In contrast to models derived for prediction of PAD, there were few significant predictors of flare and these were inconsistent across the 3 models. This is in keeping with the traditional teaching that in SLE, flares are unpredictable (18). Based on goodness-of-fit statistics and model properties, all 3 models performed poorly overall. However, although PPV was <30% in all models, NPV was >70%.
This study was strengthened by the use of a relatively large sample size for model derivation and testing. However, there were some limitations. The models derived make only short-term predictions. In addition, the definition of flare and PAD is based on SLEDAI-2K score values taken at least 2 months apart, and this must be borne in mind when applying the models in other contexts such as RCTs. Although comprehensive clinical and treatment history is needed in order to apply these models to individual patients, this type of information is often available for patients that are followed as part of a longitudinal cohort study. These are also the patients that are often considered for inclusion in RCTs. In this study we considered all cases of flare and PAD, regardless of duration or severity. An even larger sample size would allow one to determine the frequency and determinants of mild, moderate, and severe flare and PAD.
Although the time intervals used to derive prediction models were different from the time intervals used to test the predictive properties of the models, there was significant overlap in the patients used for model derivation and testing. As such, the predictive properties of the models may have been overly optimistic. It would be of interest to undertake external validation of our prediction models in other SLE cohorts and to replicate our findings using other disease activity measures, such as the BILAG or SELENA–SLEDAI, to define flare and PAD.
In conclusion, we have identified persistent activity to be a common disease state in SLE and propose that PAD should be included as an outcome variable in SLE clinical trials. Our clinical prediction models may have a role in selection of patients who are at risk for PAD and require close monitoring, and those who may be suitable candidates for inclusion in clinical trials.