FormalPara Key Summary Points

Why carry out this study?

Rheumatoid arthritis has many treatment options, but there are no accurate predictors for when a patient needs to switch treatments.

We aimed to determine which clinical and demographic features collected in routine RA care were most highly associated with switching biologics at a future visit to allow for development of an interactive prediction model.

What was learned from this study?

Disease activity and its change from prior measurements, patient pain, age, and number of prior biologics were included in a model to predict switching biologic at the next visit.

This model can provide clinicians with a better idea of a patient’s trajectory and encourage these clinicians to evaluate earlier switch for patients who are likely to switch at the following visit.

Introduction

As the treatment armamentarium for rheumatoid arthritis (RA) grows, patients have more options for biologic therapy both within treatment classes and across differing mechanisms of action [1]. Unfortunately, the effectiveness of each individual biologic does not always persist on an ongoing basis. In a meta-analysis of tumor necrosis factor (TNF)-inhibitor drug survival, for example, nearly 30% of patients had discontinued TNF-inhibitors at the 1-year time point, with > 50% discontinuing at 4 years [2]. As a result this decreasing effectiveness, patients usually switch several biologics during their lifetime. When patients stop to respond to a certain biologic, pain and inflammation return and the risk of erosions and damage reemerges; steroids are frequently reintroduced or increased until a switch to a new biologic occurs. In clinical practice, this process often entails an additional visit to the clinic.

One frustrating feature of RA for patients and clinicians alike is the lack of predictability of favorable response to treatment. Failure to respond to biologics is not just discouraging, but also costly to the medical system [3]. Prior research has identified a number of factors that predict non-response to particular disease-modifying anti-rheumatic drugs (DMARDs). Laboratory-based efforts in predicting treatment failure or inadequate response have included evaluating autoantibodies and their change in titer, gene expression levels [4], blood lymphocytes [5], and drug levels, among others. A decrease in the titer of anti-citrullinated peptide antibodies after treatment has been associated with abatacept response and retention [6, 7]. Conflicting evidence exists regarding anti-adalimumab antibodies and adalimumab serum levels on determining the likelihood and long-term TNF-inhibitor response or response to a subsequent TNF-inhibitor [8, 9]. Other studies have focused on specifics of early treatment as predictors of response or failure. In a study of etanercept, combination therapy with conventional synthetic (cs)DMARDs other than methotrexate and presence of comorbidities were associated with discontinuation [10]. In another study of difficult-to-treat RA, failure to begin methotrexate within 3 months and failure to wean corticosteroids within 6 months predicted future difficult-to-treat RA [11]. Many clinical studies have focused on disease activity at different time points as a predictor of treatment failure. A study evaluating csDMARD failure in long-term RA treatment identified residual disease activity at 3 and 6 months as the primary predictor of long-term failure [12]. Similarly, non-response to certolizumab at 3 months predicted failure to achieve low disease activity at 1 year [13].

While much research has been devoted in predicting initial response, the issue of specifically predicting the need to switch biologics has been less investigated. Factors other than a patient’s underlying immunobiology and disease activity may be predictive of the need to switch biologics. For example, patient preference, demographic features, insurance status, age, and comorbidities may all affect the likelihood of switching biologics irrespective of disease activity. Rheumatologists may not accurately understand patient preferences for different types of biologic therapy, as shown in a previous qualitative study [15].

The ability to predict a biologic switch in advance may improve the quality of therapy provided to patients; for example, a patient about to need a switch to a different biologic could be seen at more frequent intervals, a switch could happen in a timely manner, flares of disease could be avoided, steroids use could be eliminated or considerably decreased. With more consistent control of RA disease activity over time, patients can also avoid erosive damage to their joints and disability.

Many rheumatologists longitudinally collect patient- and physician-derived measurements beyond disease activity as a part of routine clinical care. Such data can be used for prediction of outcomes. It is unknown, however, whether patient-reported outcomes or other clinical features at a clinic visit help predict a medication change at the next visit. In the present study, our aim was to determine which clinical and demographic features collected in routine RA care were most highly associated with switching biologics at a future visit. We also aimed to develop a model using data collected regularly at preceding visits to predict the probability of switching biologics at the time of a subsequent clinic visit.

Methods

Study Cohort and Design

Patients were drawn from the CorEvitas (formerly Corrona) registry and were adults aged > 18 years with a diagnosis of RA [16]. The registry database as of 1 January 2022 was used for analyses. As of that date, data on 57,543 patients with RA had been collected since inception, recruited from 211 private and academic practice sites across 43 states in the USA, with 911 participating rheumatologists. Data were collected from both patients and their treating rheumatologists using structured case report forms and include information on disease duration, prognosis, disease severity and activity, comorbidities, use of medications, and patient-reported outcome data.

All participating investigators were required to obtain full board approval for conducting research involving human subjects. Sponsor approval and continuing review were obtained through a central Institutional Review Board (IRB) (New England Independent Review Board [NEIRB] No. 120160610). For academic investigative sites that did not receive a waiver to use the central IRB, approval was obtained from the respective governing IRBs and documentation of approval was submitted to the Sponsor prior to initiating any study procedures. All registry subjects were required to provide written informed consent prior to participating. The study was conducted in accordance with the Helsinki Declaration of 1964 and its later amendments.

Two sets of patients were included for this analysis: switchers and controls (non-switchers)

Inclusion and Exclusion Criteria for the Group that Switched Biologics/tsDMARDs

Patients included as switchers had at least two visits prior to the switch while on drug (at least 1 visit after initiation and 1 visit prior to switching). The visit prior to the switch was between 2 and 12 months of the switch. The patient had to discontinue the original initiated drug no more than 6 months prior to the switch to a new drug. This interval of 6 months was chosen to ensure that discontinuation of the original drug was done in the context of a switch rather than an attempt to taper off medications or for some other reason associated with stopping the original drug. The visit prior to the switch must have been at least 3 months after the initiation. The patient needed to have a Clinical Disease Activity Index (CDAI) score for the two visits prior to the switch and a baseline CDAI score (CDAI score at time of initiation). In the scenario where an individual patient had ≥ 2 while in the CorEvitas registry, the first initiation/switch of a patient was used for the analysis.

Inclusion and Exclusion Criteria for Control Patient Visits (Who Did Not Switch)

Patients could be included as controls if they had at least two visits while on the initiated biologic/tsDMARD drug (one could be initiation visit; Fig. 1) and did NOT switch after those visits. We excluded patients used in the switch group from the control group. A control visit was matched to a switch visit by drug class, prior bDMARD/tsDMARD experience, and time from initiation (Fig. 1). The visit after the matched control visit could not be a discontinuation of the biologic/tsDMARD.

Fig. 1
figure 1

Timeline of visits and measurements of disease activity for switchers and controls

Statistical Analysis

The pairs of switch and control visits were randomly divided into a training dataset (60%) and a test dataset (40%). The training dataset was used to determine a best prediction model and applied to the test dataset for estimation of the probability of switching.

The set of covariates described in the baseline table were considered, including patient demographics (sex, age, race, education, work status, insurance, smoking); disease status (comorbidities, disease duration); measures of disease (CDAI, joint counts, patient and provider global assessments, patient pain, Modified Health Assessment Questionnaire [mHAQ], AM [morning] stiffness, Routine Assessment of Patient Index Data (RAPID)); and measurements of chronological time (year of initiation). For disease measures, values at the visit prior to switch, the change in values at the visit prior to switch, and the change from baseline were all considered as potential predictors. For change measures, LOWESS curves illustrated separate slopes for changes below and above zero, so linear splines with a knot at zero were used for disease change covariates.

We used data that were available for all models since there was minimal missingness for covariates examined. The covariates selected for potential use were ones that had minimal missingness in the registry (all disease activity measures combined had 0.2% missing). The final model chosen in the training dataset had an n = 3006 (vs. 3016 total — 0.3% missing); the test set had 0.2% missing. In cases where there was more than 'minimal' missing—for example, RAPID (12% missing) or disabled indicator (1.8% missing)—we examined potential prediction with its addition. The plan, if it showed improvement, would have been to examine if it was the covariate (RAPID) or the subpopulation used.

Three methods were used to examine potential models and reduce overfitting: best subset regression, lasso [17], and ElasticNet [18]. Best subset regression selects using log likelihood; lasso adds a penalty on coefficients, forcing some to zero; ElasticNet uses a combination of ridge regression and lasso. The best models from these methods were examined for prediction (in the training dataset) of estimating the probability of switching and the area under receiver operating characteristic (ROC) curve (AUC). The best models from each method were examined and further simplified if the AUC was not reduced. A final model was chosen and that model applied to the test dataset once to provide estimates of prediction of switching. Dominance statistics were used to determine ranked importance of predictors [19, 20]. All analyses were carried out using Stata 17.0 (StataCorp, LLC, College Station, TX, USA). Specific Stata commands used for model selection are listed in Electronic Supplementary Material (ESM) Appendix 1, and a summary of the dominance analysis is given in ESM Appendix 2. A probability of switching calculator was created using the coefficients from the best model selected.

For a sensitivity analysis, the model coefficients were varied to the upper and lower 95% confidence limit estimates of each coefficient, and the probability of switching and AUC were estimated to examine the stability of the model.

Results

Demographics and Clinical Characteristics

Of 4307 patients who had initiated biological (b)DMARDs or tsDMARDs and had two visits prior to switching to a different b/tsDMARD, 2564 met the other inclusion criteria of discontinuing the original drug less than 6 moths prior to switch, having a visit prior to switch > 3 months after initiation and having a CDAI available at the time points of interest (Fig. 2). Of these 2564 patients, 2525 were matched with 2525 control patients and were randomly divided into the training (60%, 3016 patients or 1508 pairs) and test set (40%, 2034 patients or 1017 pairs).

Fig. 2
figure 2

Flow chart for visit inclusion in the study. bDMARD Biologic disease-modifying anti-rheumatic drug, CDAI Clinical Disease Activity Index, IL-6 IL-6 receptor inhibitors, JAK JAK inhibitors, pts patients, RA rheumatoid arthritis, tsDMARD targeted synthetic disease-modifying anti-rheumatic drug, TNFi tumor necrosis factor-alpha inhibitors

The average age for the 5050 patients included in the analysis was 59.6 years, the majority were female (3998, 79.2%), and the average duration of RA at the time of the switch or control visit was 12.8 years (ESM Table 1A). In the training dataset, sex, race, and duration of RA did not differ between the switchers and control patients, but patients who switched were younger on average than control patients (58.3 vs. 60.2 years, respectively; Table 1). Smoking status, history of cancer, cardiovascular disease, or diabetes did not differ between the switching and control groups. There were no between-group significant differences in types of insurance or education level. Patients who switched were more likely to be disabled (p = 0.0002) but less likely to be retired (p = 0.0001). Patients who switched were less likely to be positive for rheumatoid factor (p = 0.006). The patients who switched had higher measures of disease activity, including swollen and tender joint counts, patient and physician global scores, CDAI, and RAPID score (Table 1).

Table 1 Demographic and clinical characteristics of switchers and control patients in the training dataset

As the data were randomly divided into training and test sets, there were no major differences in the distribution of clinical features in switchers and control patients in the test data (ESM Table 1B), although there were some differences in demographics between the training and test datasets. There was no statistically significant difference between the patients who were disabled in the switcher and control groups in the test set, but there were higher percentages of black patients and patients insured with Medicare in the control group in the test set (ESM Table 1B).

Prediction Model Selection, and Performance

The AUC for the model ultimately selected for its performance and simplicity was 0.690 (Fig. 3). The features included in the model, which were all significantly associated with likelihood of switching, are listed in Table 2. Age was inversely related to switching, as was the prior number of biologics. For example, those who were in the age group > 70 years had an odds ratio (OR) of switching of 0.606 compared to those who were aged < 40 years; patients with 3+ prior biologics had an OR of 0.706 of switching compared to biologic naïve patients. With each additional biologic there was a lower OR of switching; this was the same pattern with each increasing decade of age having a lower OR of switching (Table 2). Patients who had a CDAI at the visit prior to switch greater than remission had a higher OR of switching, with the patients in the high disease activity category having the highest OR of 2.964 (Table 2). Additionally, those patients with a change in CDAI > 0–that is, a worsening of disease activity–in the visit prior were more likely to switch at the next visit. Patient-rated pain on a visual analog scale of 0–100 was positively correlated with the likelihood of switching (OR 1.01).

Fig. 3
figure 3

Predicted probability of switching based on the model estimated using the training data set and then applied to the test dataset. ROC Receiver operator characteristic curve

Table 2 Prediction model in training dataset

When the model was applied to the test set, the AUC was almost identical at 0.687. The histograms show the predicted probability of switching with the actual switchers shaded in green and the control patients in the unfilled bars (Fig. 3). There is some overlap between the predicted probability of switching between actual switchers and controls, but generally patients who actually switch have higher predicted probability of switching.

Dominance analysis provided the following prediction ranking of the covariates: Prior CDAI, prior patient pain, change in CDAI from baseline, age, prior biologics/tsDMARDs used. The details of the analysis are given in ESM Appendix 2.

Varying the coefficients of the prediction model showed only a small change in estimated AUC (ESM Appendix 3), illustrating that prediction does not vary within the range of the estimated coefficients.

Probability Calculator

Based on the coefficients of the final model, an interactive probability calculator was constructed (ESM Fig. 1). Four hypothetical patient scenarios are presented in Table 3. The patients belong to different age groups, and are characterized by different changes in CDAI, number of prior biologics, and pain levels. For patient 3, although there was an increase in CDAI of 2, the lower pain score and older age group indicate a lower probability for switching biologics. Despite patient 2 having a CDAI at the visit prior to switch indicating low disease activity and not having worsened in the CDAI score (change in CDAI 0), this patient was more likely to switch than patient 3 due to younger age, higher level of pain, and being naïve to biologics.

Table 3 Probability calculations of switching biologics/targeted synthetic disease-modifying anti-rheumatic drugs at the next visit for four hypothetical patients

Discussion

In this study involving a large prospective cohort of patients with RA, we developed an interactive model to predict the likelihood of switching biologic or tsDMARD therapy at the subsequent visit. The predictors of a future switch were CDAI category and change in CDAI at the visit prior to switch, pain level at this visit, age group, and number of prior biologics. The model differentiated switchers from patients who did not switch with an AUC of about 0.69 in both the training and the test datasets. This is the first time that clinical and demographic features present prior to a switch of biologic agents have been developed and validated for clinical use.

Unlike some prior models to predict failure for particular DMARDs or biologics, the model developed in this study revealed the importance of features other than disease activity to predict therapeutic failure. This differs from a previous study in difficult-to-treat RA which showed that these patients could not be differentiated from patients with non-difficult-to-treat RA based on baseline demographic features or disease activity indices [21]. Age and number of prior biologics may affect a patient’s likelihood of switching in different ways. That patients of older age were less likely to switch could suggest that immune senescence is playing a role or that patients and physicians are more reluctant to switch older patients to a new drug possibly due to more comorbidities, or some combination of these factors. We found that higher numbers of prior biologics were associated with a lower likelihood of switching. This association possibly reflects the lack of perceived options for patients by physicians or skepticism by patients who have already experienced treatment failure with several other agents, reflected by their reluctance to embrace a new biologic which may, or may not, have more benefit than their present agent. The model performed equally well in training and test data, suggesting it was not overfit.

The development of an interactive prediction model that could be made into a mobile application or embedded in the electronic health record was a unique part of this study. There has been much interest in recent years about using mobile applications to manage rheumatic diseases such as RA. Many of the efforts have been towards developing applications for patients to improve self-management and empowerment [22]. Clinician tools, such as RheumaHelper and the DAS calculator, have been developed, but minimal research has focused on how these applications improve patient outcomes [23].

A limitation of the model developed here is the overlap in predicted probability between switchers and control patients, reflecting the AUC of 0.69. This likely reflects the variability in patient and physician preferences and practices. Decision-making may be different based on the patient population seen and the resources of the patients and the clinic to obtain biologic approval from medical insurance providers. Additionally, the registry, like all similar observational registries, lacks data on adherence to medications, on whether patients were unable to obtain prior authorization to desired medications and other limitations to medication access, and on patient preferences regarding treatment choices. All of these factors affect decisions about starting, stopping, and switching medications and could play a role in mediating who will switch bDMARDs/tsDMARDs. Finally, the registry may not be fully representative of all populations with RA, which may limit its generalizability, particularly outside of North America.

Further studies could apply this algorithm in a more limited group of patients to evaluate whether there is higher performance in a single institution where treatment algorithms are likely to be similar among physicians. This could also be examined in different healthcare systems in which commercial insurance does not play as large a role in healthcare decision-making as in the USA. As the model includes personal health information, use in mobile applications or electronic health record would need to have appropriate security measures to protect patient privacy.

Conclusions

In summary, disease activity and its change, along with pain, age, and prior biologic exposure predicted the probability of switching biologics at the subsequent visit in a prospective cohort of patients with RA from the USA. Future studies could evaluate whether incorporation of this model into an electronic database could provide a signal to inform the judgment on the likelihood of switching at a future visit and whether an earlier switch of biologic or tsDMARDs would benefit patients who are destined to change these medications.