Introduction

Rheumatoid arthritis (RA) is a chronic autoimmune disorder that causes inflammation in the joints, leading to pain, stiffness and joint deformities. The primary objective of RA treatment is to reduce inflammation, preserve joint function and improve quality of life. If initial conventional synthetic disease-modifying antirheumatic drugs (csDMARDs), such as methotrexate, are ineffective or not tolerated by patients, biological DMARDs (bDMARDs) or targeted-synthetic DMARDs (tsDMARDs) are important treatment options. The most recent second-line treatment options for RA are the synthetic small molecules called Janus kinase inhibitors (JAKi). Currently, five JAKis are available for the treatment of RA: baricitinib, filgotinib, peficitinib (in Japan only), tofacitinib and upadacitinib.1 JAKi have demonstrated superior efficacy over placebo or csDMARDs, and some have even demonstrated superiority over adalimumab or abatacept in specific outcomes.2 3 However, an open-label post-marketing safety randomised controlled trial, the ORAL-surveillance study,4 has raised concerns of an increased risk of malignancies and major adverse cardiovascular events, with tofacitinib compared with tumour necrosis factor inhibitors (TNFi), in patients with RA with particular risk factors. Considering these results, the use of JAKi has been restricted by health authorities worldwide.

The objective of this study was to assess and compare the overall tolerability and safety of JAKi, TNFi and bDMARDs with other modes of action (OMA), by evaluating treatment discontinuations for adverse events (AEs) in a real-world population. In light of the ORAL-surveillance findings, a secondary objective was to explore whether the observed effects are attributable to the entire JAKi class or are specific to tofacitinib.

MethodsPatient and public involvement

Patients and/or the public were not involved in the design, conduct, reporting or dissemination plans of this research.

Patient sample

The JAK-pot register collaboration conducted an investigator-initiated observational study aiming at evaluating clinical aspects of JAKi and bDMARDs in RA.5 6 Patients with a clinical diagnosis of RA and starting treatment with a JAKi, a TNFi or a bDMARD-OMA were included. Treatment courses of interest initiated before the first JAKi was commercially available in each participating country were excluded from the analysis to avoid confounding by time-trends. Rituximab was not included as the time to discontinuation is difficult to assess. Patients could contribute to multiple treatment courses. This study included patients from the following 17 registers, which all provided individual treatment course-level data: ATTRA from the Czech Republic, ARBITER from Russia, BIOBADASER from Spain, BIOREG from Austria, BioRx.si from Slovenia, BSRBR-RA from the UK, GISEA from Italy, I-RECORD from Israel, METEOR from the Netherlands, NOR-DMARD from Norway, RABBIT from Germany, REUMA.PT from Portugal, RHUMADATA from Canada, ROB-FIN from Finland, RRBR from Romania, SCQM from Switzerland, TURKBIO from Turkey.

Exposure of interest

The exposure of interest was the type of treatment course: JAKi (baricitinib, filgotinib, tofacitinib or upadacitinib), TNFi (adalimumab, certolizumab, etanercept, golimumab or infliximab) or bDMARD-OMA (abatacept, sarilumab or tocilizumab).

Time point definitions

Each treatment course was defined as the time between treatment initiation (baseline), and treatment discontinuation, start of a new treatment, end of participation in the register or the end of the study period (November 2022), whichever came first. The end of the study period was set at an earlier date for some registers who could not provide their most recent data (see online supplemental material).

Study outcomes

The primary outcome was the rate of treatment discontinuation due to AE, as decided by treating clinician, compared across treatment groups. Each treatment course was linked to a single reason for discontinuation. When several reasons were provided the order of precedence was (1) AEs, (2) others and (3) ineffectiveness.

Covariates of interest

Baseline covariates were chosen a priori based on clinical knowledge and literature. They included patient, disease and treatment characteristics, namely gender, age, body mass index, tobacco smoking (ever/never), disease activity (Clinical Disease Activity Index (CDAI) or Disease Activity Score 28 if CDAI was unavailable), disease duration, seropositivity (rheumatoid factors and/or anti-citrullinated protein antibody), number of previously used bDMARDs/tsDMARDs (0, 1, 2, ≥3), concomitant csDMARDs treatment (none; methotrexate; other csDMARDs without methotrexate; methotrexate and at least one other csDMARDs), concomitant glucocorticoids (presence/absence), functional status (Health Assessment Questionnaire Disability Index) and C reactive protein. Because AEs are significantly influenced by comorbidities, we further adjusted for pre-existing interstitial lung disease, cardiovascular disease, infections, malignancies, diabetes, depression, hypertension and hyperlipidaemia. The comorbidity variables were dichotomised to represent the presence or absence of a specific condition at treatment baseline. This approach was taken to adapt to data granularity and harmonise data from different registers, making our pooled analysis feasible.

Statistical methods

The results are reported according to the Strengthening the Reporting of Observational Studies in Epidemiology recommendations.7 Baseline disease and patient characteristics by treatment groups are summarised using counts and percentages for categorical variables, mean and SD for continuous variables. The number of treatment courses with valid values (non-missing) are provided. Raw incidence rates were plotted using the crude non-parametric cumulative incidence functions of treatment cessation for AE in JAKi, TNFi and bDMARD-OMA patients. We used unadjusted and adjusted (for baseline confounders, as described above) cause-specific Cox proportional hazard models,8 considering stopping for ineffectiveness or other reasons as competing events, to compare treatment discontinuations for AE between treatment groups. Treatment discontinuation associated with each treatment group was compared using HR estimates, with JAKi as the reference level. Multiple treatment courses from the same patient were included if the individual had received more than one second-line treatment during the study period. To account for this within-patient correlation, we incorporated a cluster term for the patient identity in the cause-specific Cox model, in order to obtain robust SEs. We also used the Fine-Gray subdistribution hazard model,9 to account for competing risks (stopping for ineffectiveness, pregnancy, remission or other reasons) as a sensitivity analysis. The cause-specific hazard approach focuses on the hazard function for each specific cause separately and treats competing events as censoring events. The cause-specific hazards are more relevant when exploring aetiological inquiries as it quantifies the event rate among individuals at risk of encountering the event of interest (treatment cessation for AE, in this scenario). In contrast, the Fine-Gray method takes competing risks into account when estimating the cumulative incidence function, modelling the subdistribution hazard without treating competing events as censoring events, becoming particularly useful when the objective is predicting an individual’s risk. Missing covariates were imputed using multiple imputations with chained equations (50 samples and 25 iterations, predictive mean matching algorithm for continuous variables and logistic and polytomous regression for categorical variables). All baseline covariates were used in the imputation. The outcome variable (reason for discontinuation), an indicator variable (the event of discontinuation) and the untransformed time-to-event were used as predictors in the multiple imputation process but were not imputed.

Subgroup analysesRandomised controlled trial-cohorts

We considered two subgroup analyses of patients at higher risk of AEs:

The ‘randomised controlled trial (RCT) - duplicate cohort’, mimicking the main inclusion criteria of the ORAL surveillance trial,4 namely patients 50 years or older and with at least one cardiovascular risk factor (hypertension, hyperlipidaemia, diabetes, smoking, prior history of strokes or myocardial infarctions).

The ‘RCT – high cardiovascular disease (CVD) risk cohort’, a subset of patients aged ≥65 years with at least one cardiovascular risk factor.

Separate comparison of tofacitinib, baricitinib and other JAKis

To explore signals of AEs reported specifically with tofacitinib, baricitinib and other JAKis, we also analysed them separately, to explore effect modification by the type of JAKi, both in the general cohort and in the higher risk subgroups.

Results

A total of 46 913 treatment courses, started by 33 511 patients, were included: 12 523 on JAKi, 23 391 on TNFi and 10 999 on OMA (table 1). Patients were predominantly women (79%) and seropositive (77%), 57.3 years old on average, with a mean disease duration of 12.7 years. Most of the patients on JAKi were treated with baricitinib (43%), followed by tofacitinib (40%), upadacitinib (15%) and filgotinib (2%). According to routine clinical practice in the participating countries, patients starting a TNFi had shorter disease durations and were more often on their first second-line therapy and less often on monotherapy compared with non-TNFi treatments. Non-TNFi treatments were more often on third-line (or more) therapy, and presented higher counts for some comorbidities, indicating a more refractory disease.

Table 1

Baseline characteristics by treatment group

Treatment discontinuation due to AE

During a mean treatment follow-up of 1.9 (SD: 1.6) years, and a mean patient follow-up of 2.9 (SD: 1.9) years, 899 JAKi treatments, 1387 TNFi treatments and 870 OMA treatments were discontinued for an AE (table 2). The crude incidence rate of treatment discontinuation for AEs was 4.3 with JAKi, 3.3 with TNFi and 4.4 with OMA per 100 patient-years of treatment. Crude incidence rates and cause-specific HRs for competing events can be found in the online supplemental table S8. Among patients who stopped their treatment, 47% did so for ineffectiveness or other reasons.

Table 2

Discontinuation for adverse events among JAKi, TNFi and OMA patients

The unadjusted cumulative incidence function (figure 1) showed higher treatment discontinuations for AEs with JAKi compared with TNFi, but lower compared with bDMARDs-OMA. The adjusted cause-specific HR (csHR) of treatment discontinuation for AEs was higher for OMA versus JAKi (csHR 1.11, 95% CI 1.01 to 1.23) and not statistically different for TNFi versus JAKi (csHR 1.00, 95% CI 0.92 to 1.10) (table 2). The significantly higher rate of treatment discontinuations for OMA compared with JAKi is mainly driven by anti-interleukin (IL)-6 medications, whereas abatacept patients have lower treatment discontinuation rates (online supplemental table S7). CsHR for the main analysis and subgroup analysis are shown in figure 2.

Figure 1

Crude cumulative incidence of stopping for adverse events. JAKi, Janus kinase inhibitors; OMA, biological disease-modifying rheumatic drugs with other modes of action; TNFi, tumour necrosis factor inhibitors

Figure 2

Adjusted cause-specific HRs (csHR) of stopping for AEs for TNFi and OMA against reference JAKi, in the overall cohort and in the different subgroups. AE, adverse event; csHR, cause-specific HR; JAKi, Janus kinase inhibitors; OMA, biological disease-modifying rheumatic drugs with other modes of action; RCT, randomised controlled trial; TNFi, tumour necrosis factor inhibitors.

The sensitivity analysis using Fine-Gray approach instead of cause-specific Cox demonstrated a similar trend for OMA versus JAKi, although the risk of discontinuation for AEs did not reach statistical significance (subdistribution HR (sHR) 1.08, 95% CI 0.98 to 1.19). No significant difference was found for TNFi versus JAKi (sHR 0.98, 95% CI 0.87 to 1.09) in the discontinuation for AEs. Adjusting for countries did not change the csHRs.

Subgroup analysesRCT-cohorts

The ‘RCT-duplicate cohort’ represented 36% of all treatment courses (n=17 146), and crude incidence rates of treatment discontinuations for AEs were comparable to the overall cohort (4.2, 3.2 and 4.5 per 100 patient-years for JAKi, TNFi and OMA, respectively). The results were almost similar to the full cohort (table 3), with a higher csHR for OMA versus JAKi (csHR 1.19, 95% CI 1.03 to 1.39) and no significant differences for TNFi versus JAKi (csHR 0.99, 95% CI 0.86 to 1.14).

Table 3

Discontinuation for adverse events among JAKi, TNFi and OMA patients, in the RCT-duplicate cohort and in the RCT—high CVD risk cohort (aged ≥65 years with ≥1 cardiovascular risk factor)

In the ‘RCT-high CVD risk cohort’, the incidence rate of treatment discontinuations for AEs was higher (5.3, 3.4 and 4.9 per 100 patient-years for JAKi, TNFi and OMA, respectively). We no longer found a significant difference in the cause-specific hazard rate of discontinuation for AE between OMA versus JAKi (csHR 1.01, 95% CI 0.82 to 1.25), but the hazard was lower for TNFi versus JAKi (csHR 0.79, 95% CI 0.65 to 0.97).

Separate comparison of tofacitinib, baricitinib and other JAKis

Baseline characteristics of JAKi treatments indicated a more refractory disease for patients on baricitinib and other types of JAKi than patients on tofacitinib (online supplemental table S1).

When exploring effect modification in the overall cohort by the type of JAKi, specifically for tofacitinib and baricitinib (table 4), we found some suggestions for an effect modification by type of JAKi. The hazard rate of discontinuation for AE for TNFi was significantly lower compared with tofacitinib (csHR 0.81, 95% CI 0.71 to 0.90), a significant difference which was inversed when comparing TNFi to baricitinib (csHR 1.15, 95% CI 1.01 to 1.30) and other types of JAKi (csHR 1.26, 95% CI 1.02 to 1.55) (online supplemental table S2). Ninety per cent of patients on baricitinib received the 4 mg dose. Similar trends were found in the ‘RCT-duplicate cohort’ and the ‘RCT-high CVD risk cohort’ (online supplemental tables S3 and S4). Results remained similar when looking at the time period before ORAL-surveillance results were published (online supplemental table S5).

Table 4

Discontinuation for adverse events for tofacitinib and baricitinib, against TNFi and OMA patients

Discussion

In this large international collaborative study, we did not find an increased incidence of treatment discontinuation due to AEs with JAKi compared with other biological DMARDs in the overall cohort. After adjusting for potential confounders, the incidence and rate of treatment discontinuation for AEs were similar for TNFi versus JAKi, and the range of the CI supports the idea of very similar rates of stopping for AE. The rates of treatment discontinuation for AE tended to be higher with OMA versus JAKi, a difference mainly driven by anti-IL-6 medications. When selecting patients at increased risk for AEs, similar to the inclusion criteria of the ORAL-surveillance trial,4 the incidence of treatment discontinuation for AEs was higher, but the relative effect sizes between the different treatment groups were similar to the overall group. However, the incidence and the rate of discontinuation for AEs of patients on JAKi was increased compared with TNFi in the ‘RCT – high CVD risk cohort’ of patients aged 65 years and older. When exploring the possibility of a differential effect between the different types of JAKi, we found a higher rate of discontinuation for AEs with tofacitinib compared with TNFi in the overall cohort and in the ‘RCT-high CVD risk cohort’, a difference that we did not see with the other JAKi. The apparent signal of an elevated rate of treatment discontinuations for AEs with tofacitinib was also increased in the ‘RCT-high CVD risk cohort’.

There are several possible explanations for the discrepancy between available observational studies and the ORAL-surveillance trial. First, it is paramount to highlight that our study aims to address a distinct research question. While ORAL-surveillance specifically targets certain AEs, our study is aimed at assessing treatment tolerability. This broader scope encompasses a range of patient experiences, from nuanced feelings of ‘not feeling very well’ to severe AEs. Consequently, our primary outcome, discontinuation due to AEs, captures a more subjective spectrum of patient reactions than ORAL-surveillance. Second, although we used discontinuation for AEs as our primary outcome, there is always the possibility that the decisions to discontinue therapy is also influenced by other reasons, such as insufficient effectiveness, even though the recorded reason for discontinuation was an AE. Finally, it is possible that the incidence of AEs is only increased in older patients with specific risk factors and not in the overall population. This last explanation is supported by the result of the sensitivity analysis including only patients >65 years and with at least one cardiovascular risk factor. Subgroup analyses of the ORAL surveillance study also suggested similar differential safety results in high-risk groups, such as a history of arteriosclerosis,10 or a history of tobacco smoking.11

It is currently unknown whether the safety signals observed in the ORAL-surveillance study with tofacitinib represent a class effect of the JAKi group or a drug-specific effect of this particular drug. We explored a potential effect modification by type of JAKi and examined tofacitinib, baricitinib and other JAKi separately. Our findings indicate that patients taking tofacitinib were more likely to discontinue treatment due to AEs compared with those taking TNFi. This was not observed with other JAKi, although caution is needed due to limited data on filgotinib and upadacitinib. These results may hint at differences in safety profiles between tofacitinib and other JAKi but should not be considered definitive evidence against a JAKi class-effect. Recent translational research suggested that individual JAKi agents differ in their biological effects on specific immune cells, such as the natural killer cell activation.12 A recent study from the Swedish ARTIS register13 did not find any overall differences in discontinuation for AEs between the JAKis tofacitinib and baricitinib and other bDMARDs. However, the study’s sample size was limited, with only 170 cases of discontinuation resulting from AEs (114 with baricitinib and 56 with tofacitinib), making it challenging to detect differences if they exist. Subjective factors such as patients’ tolerability for AEs can play a significant role in determining treatment discontinuation rates for a particular medication. AEs can vary in severity, and some patients and physicians may be more willing to tolerate mild or moderate AEs if they perceive that the benefits of the medication outweigh the risks, which may have been less the case for tofacitinib than the other JAKi recently, although we should point out that the data collection occurred mostly before the publication of the ORAL-surveillance trial, and results from the analysis did not change when considering data before ORAL-surveillance. The availability of alternative treatments and their accessibility may also vary from country-to-country affecting the decision to change or not the ongoing treatment.

The HR using the Fine-Gray model used for a sensitivity analysis was overall not different from the cause specific Cox hazard model used in the main analysis. In the Fine-Gray model, the cumulative incidence of treatment discontinuation for AEs depends on the hazards of the competing causes, such as ineffectiveness. As there is an increased cause-specific hazard rate for ineffectiveness for TNFi compared with JAKi, the HR for stop for AEs with the Fine-Gray model is less than we would expect if the exposure was not associated with ineffectiveness, and so is smaller than with the cause-specific Cox model.

One of the strengths of this study is the use of several registers from different countries with a large sample size and the possibility to adjust for many covariates, including disease activity and disease and treatment characteristics. In contrast with claims database, disease and treatment registers are also able to incorporate clinical information such as disease activity, function and smoking status. We also used discontinuation for AEs, which is an outcome that is consistently reported among registers and includes not only safety, but also the patients’ perceived tolerability. Ineffectiveness being the most common reason for drug discontinuation,5 accounting for it through a competing risk approach reduces the risk of bias.

The main limitation is the lack of details on the type and seriousness of AEs leading to discontinuation, as this information cannot always be simply linked to the discontinuation of the treatment and is not as consistently reported among the registers. While our results provide an overview of tolerability across treatment groups, the exact nature and severity of AEs leading to discontinuation remains an area that requires further investigation. This absence of details means clinicians must exercise caution when interpreting these findings in terms of specific safety concerns, such as cardiovascular events or serious infections. Furthermore, drug discontinuation is inherently complex and blends clinical, patient-specific and sometimes socioeconomic variables. Our research highlights some elements, but a comprehensive exploration remains essential for future studies. A second limitation is the absence of randomisation between treatment groups, although we tried to emulate as much as possible a trial, by adjusting for relevant confounding factors, including a wide range of comorbidities known to play a role in the occurrence of AEs, including patients from the start of their treatment, selecting a comparable population, and having a well-defined intervention, follow-up period and outcome. We did not prespecify any direct comparison between OMA and TNFi, so it is not possible to determine if, in addition to having comparable or potentially inferior outcomes to JAKi, OMA could lead to higher discontinuation rates due to AEs when compared with TNFis. Missing data can also present a significant concern, but it was addressed with multiple imputation. The follow-up period is relatively short and may limit our ability to detect long-term AEs or late-onset discontinuations. Finally, misclassification of the cause for drug discontinuation can be a concern. In interpreting our findings, it is essential to acknowledge the potential for physician bias in attributing the reason for drug discontinuation. Physicians might be influenced by pre-existing clinical experiences and knowledge, possibly leading to differential reporting across treatments.

In conclusion, in this international cohort of patients with RA, patients on JAKi did not demonstrate higher rates of discontinuation for AEs compared with patients on TNFi or bDMARD-OMA in the overall population. However, differences were observed in specific subgroup of patients with older age and cardiovascular risk factors or with different JAKi molecules, with a higher risk of discontinuation of AEs among those with cardiovascular risk factors and older age and with tofacitinib compared with TNFi but not the other JAKi. Treatment discontinuation for AEs comprises a wide range of AEs and includes both the patient perceived intolerability and the physician described safety issues. Our findings suggest that tolerability might differ depending on the subpopulation and possibly the type of JAKi. Further studies are warranted to assess which antirheumatic treatment fits best for which subgroups of patients and if the differences in tolerability by type of JAKi suggested by this study are associated with true differences in safety profile when looking at specific AEs. The conclusions drawn are specific to RA and merit further exploration within the context of other rheumatic and musculoskeletal diseases.