While there are many good treatment options for RA, the multitude of treatments has led to greater confusion. Which treatments are best for which patients in which order? These questions are complicated to approach but must begin with an understanding of how different treatments are sequenced in RA. We used a large real-world data set from the Corrona RA registry to examine methods for describing the path to a given treatment. We embarked on this sequence analysis focusing on TCZ, and specifically TCZm, as a “proof of concept” study. We found several dominant sequences of treatments that led to TCZm and some patient characteristics associated with TCZm use over time, including prior TCZc use, older age, longer disease duration, seronegative status, higher disease activity and no immediate past prior use of a TNF inhibitor. Further work will analyze sequences of competing treatments. For a brief comparison, see Supplemental Figure 1 which illustrates sequence of drugs toward TNF inhibitor monotherapy.
The goal of these analyses was not to derive a better treatment sequence (we did not focus on clinical outcomes), but rather to develop ways of assessing sequences of DMARDs. The typical method for examining comparative effectiveness in RA has focused on comparing one drug with another, sometimes in randomized controlled trials but often using observational epidemiologic methods(19, 20). While comparing active drugs with one another (versus placebo) is critical, the therapeutic armamentarium for RA has at least five dominant mechanisms each with several agents. As the current analyses demonstrate (Figure 1 and Table 2), patients switch drugs commonly. The switching to an option like TCZm has strong correlates, including the use of specific prior DMARDs. Thus, to evaluate current practice, sequential trials must be considered. For example, early RA is typically responsive to multiple different therapies, but the effectiveness of treatments after initial therapy varies. Identifying patient characteristics (e.g., serologic status, gender, disease duration, HLA status) that might identify more or less successful treatment sequences would assist clinicians and patients determine the next treatment of choice. However, structuring these decision nodes in a trial format requires an adaptive trial design.
Other implications follow from these findings. First, the differences in predictors between baseline and follow-up were subtle and somewhat to be expected. The variables that became important during follow-up were prior treatments and disease activity, emphasizing the importance of DMARD sequence and DMARD response. Second, the variables associated with TCZ and TCZm use were not substantially different. This likely reflects the fact that almost one-third of patients who try TCZm have used TCZc, possibly because TCZm is similarly effective to TCZc(12-15). Third, the path to TCZm typically takes 25.6 months and a median of 1 other bDMARD and up to 5 others. This illustrates the tremendous amount of trial and error that is typical of the treatment course for RA. Several factors may contribute to the relative infrequent use of TCZm as first line bDMARD. First, In the US, drug insurance often dictates which drugs are used first-line as a bDMARD. The structure and rules of patients’ drug insurance is not a variable contained in the Corrona database. Second, it may be that rheumatologists are less comfortable using TCZm since the drugs are newer to the market than TNFi’s or abatacept. Finally, it is also possible that patients are reluctant to use a drug that may worsen their lipid profile.
While this study did not aim to define which treatments are best for a given set of patients, we focused on describing the complex sequence of RA treatments and how patients transition between treatments toward TCZm. TCZm is just one example of a bDMARD that helps illustrate the haphazard treatment sequence that patients with RA experience. We believe that there needs to be a better framework for explicitly testing treatment sequences in RA. The concept of the dynamic treatment regime describes treatment decisions based on patient states that are recognized to evolve over time (21). For example, early RA is typically responsive to multiple different therapies, but the effectiveness of treatments after initial therapy varies. Identifying patient characteristics (e.g., serologic status, gender, disease duration, HLA status) that might identify more or less successful treatment sequences would assist clinicians and patients determine the next treatment of choice. However, structuring these decision nodes in a trial format requires an adaptive trial design. An example of a relatively simple dynamic treatment regime is the dosing of warfarin over time in a patient with changing clotting times and changing clinical characteristics. A clinician considers the current state and prior states to determine the optimal treatment at a given point in time. The same set of issues, but slightly more complicated, can be considered for RA treatment: what are the patient’s current characteristics (e.g., disease activity), what are the prior characteristics (e.g., what treatments have been tried), and what are the features of the patient’s disease (e.g., serologic status, erosion status, disease duration).
The dynamic treatment regimens can be tested in the setting of sequential multiple assignment randomized (SMART) trials(22). As the name implies, this type of a randomized trial allows for sequences of treatments that may be different across patient subgroups and are determined based on treatment response(23). SMART trials have grown in popularity but we are not aware of any such trials occurring in RA.
The current set of analyses have important limitations. Similar to most registries, Corrona has longitudinal data but it is “left-censored”, people often enter as prevalent cases of RA and thus we may not have full information about their treatment history. Corrona is a very large registry but represents patients with RA in rheumatology practices. Thus, patients whose RA is cared for by primary care clinicians are not well represented. As noted, we did not focus on clinical outcomes, making it impossible for us to comment on which treatment is better or worse for a given patient. Like all good science, this type of analysis should be repeated in other datasets.
Strengths of the analysis include the large size of Corrona and the fact that it reflects real world evidence. The dataset contains many longitudinal variables, allowing us to consider sequential predictors of DMARD treatment. Including variables that change over time, such as prior treatments and disease activity, had a major impact on the regression coefficients. This suggests that responses to treatment are critical for determining sequences of treatment.
In conclusion, we examined sequences of RA treatments in a large US-based real-world dataset, focusing on TCZm. We characterized time until TCZm, dominant paths to TCZm, and longitudinal predictors of TCZm. This work highlights the variable treatment sequences experienced by patients with RA starting a bDMARD. This variability is likely because of an evidence deficit regarding the comparative effectiveness of different treatment sequences in RA. While observational datasets may provide useful information regarding dynamic treatment regimes, it is more likely that SMART trials could substantially impact future care in RA.