Identifying Predictors of First-Line Subcutaneous TNF-Inhibitor Persistence in Patients with Inflammatory Arthritis: A Decision Tree Analysis by Indication

This registry linkage study was conducted in Sweden. The Swedish health care system provides automatic and universal coverage to all legal residents, and co-payment is limited [23]. Therefore, associations between exposures and outcomes are practically unaffected by affordability issues. Furthermore, in Sweden, it is possible to deterministically link administrative registers using personal identifiers unique to each Swedish resident [24].

The national registers included in this study are held by The National Board of Health and Welfare (NBHW). Data from the different registries were linked and anonymized by the NBHW prior to data extraction. Ethics approval for the study was granted by the Regional Ethical Review Board in Stockholm, Sweden (approval number 2019-02774). The reporting of this paper follows the STROBE statement [25].

The data for this study were obtained from three nationwide Swedish administrative registers with practically complete coverage, frequently used for medical research [26‐28]. The Prescribed Drug Register (PDR) contains data on all prescriptions dispensed in pharmacies, including anatomical therapeutic chemical (ATC) codes, date of dispensation and the specialty of the prescriber. The quality of the PDR is high, with the loss of patient information estimated to be < 0.3% [27]. The National Patient Register (NPR) is a population-based register on all inpatient and specialized outpatient care in Sweden. All healthcare providers are mandated by law to report in- and outpatient episodes to the NPR (including dates and diagnoses). It has been estimated that > 99% of all somatic and psychiatric discharges in Sweden are registered in the NPR [26]. The Causes of Death Register contains information on all deaths in Sweden and their underlying causes. Less than two percent of deaths have a missing cause of death [28].

After the application of a wash-out period (no prescriptions filled for biologics between July 1, 2005, and May 5, 2010), we identified adult patients (≥ 18 years) with RA, PSA or AS/uSpA who filled their first prescription for a SC-TNFi from May 6, 2010, to December 31, 2017. The PDR does not contain information on the indication for a prescription. Therefore, indication for a prescription was determined using an algorithm based on previously described ICD-10 diagnoses [29‐31]. After identification, patients were excluded if they met one of five different criteria: (1) if they were < 18 years old at treatment initiation; (2) if they had records of a TNFi administered in an in- or outpatient hospital setting before the date of treatment initiation; (3) if they had < 12 months of both baseline and follow-up periods due to data availability, emigration or death; (4) if they had a filled prescription of SC-TNFi from any department besides rheumatology, orthopedics or rehabilitation departments; (5) if they had no diagnosis record for IA as defined in this study (i.e., AS/uSpA [ICD10: M08.1, M45, M46.1, M46.8 and M46.9], PsA [ICD-10: L40.5, M07.0, M07.1, M07.2 and M07.3] or RA [ICD10: M05.8, M05.9, M06.0 M06.9 and M12.3]). If several IA diagnoses were present, the diagnosis recorded closest in time to the first SC-TNFi prescription was used for patient classification. The rationale for the second exclusion criterion was twofold. First, it was implemented to minimize the risk of deriving incorrect estimates of treatment duration because of ambiguous gaps in the PDR data. Second, this criterion ensured that included patients were biologics-naïve given the exclusion of patients who received biologics via intravenous in a hospital setting.

The outcome of the study was treatment persistence. Consistent with methodologic guidance [32] and previous studies [29‐31], we defined treatment persistence as the duration of time from initiation to discontinuation of treatment. The index date was set at treatment initiation and occurred when patients filled their first prescription of adalimumab, certolizumab pegol, etanercept or golimumab. Discontinuation was defined to occur at the first of three events: (1) when the last prescription was depleted with no further prescription filled, (2) when the patient had a treatment gap exceeding 60 days between filled prescriptions and (3) when the patient filled a prescription of another biologic or targeted synthetic immunomodulatory treatment (a second SC-TNFi, anakinra, ustekinumab, secukinumab, tocilizumab, rituximab, abatacept, baricitinib or tofacitinib). The duration of individual prescriptions was derived based on a standard defined daily dose and pack size. Accumulated medication from overlapping prescriptions was used to offset any prospective gaps between filled prescriptions to account for the Swedish Pharmacy system, which allows for prescription refills when two-thirds of the previous prescription has been consumed.

The exposures of interest in this study were variables potentially associated with treatment persistence [21, 22] and comprised: specific SC-TNFis, age, sex, number of health care contacts, comorbidities, use of steroids, non-steroidal anti-inflammatory drugs (NSAIDs) and conventional systemic DMARDs (csDMARDs) as well as the highest level of education. (Table 1). Health care contacts were further broken down by setting (inpatient vs. specialized outpatient care) and association to IA based on ICD-10 code. Comorbidities were operationalized as the age-adjusted Charlson Comorbidity Index [CCI] excluding the RA component [33, 34]. All analyses were performed separately for RA, PSA and AS/uSpA.

Descriptive statistics of baseline characteristics were presented using mean and standard deviation or median and interquartile range for continuous variables depending on the approximate normality of the underlying distributions. For the CCI and inpatient care resource use, both mean and standard deviation and median and interquartile range were presented reflecting the mass of observations at zero. Categorical variables were presented using frequencies and percentages.

We used recursive partitioning in a conditional inference framework [35] to identify patient groups with distinct treatment persistence by IA diagnosis. Recursive partitioning is a statistical learning technique that classifies subjects according to their observed outcomes. It is arguably more intuitive and provides more clinically interpretable results than traditional regression analysis [36]. The reason is that the output of the analysis comprises patient groups defined by specific characteristics instead of a coefficient describing the change in the outcome for a one-unit change in a predictor controlled for covariates. Other benefits include that recursive partitioning allows for inclusion of an unlimited number of predictors, is adept at finding non-linear relationships and does not assume that ordinal variables are interval or ratio scales in nature [35].

The recursive partitioning algorithm identifies subgroups with differences in outcomes using sequential splits that each maximize the difference in outcomes between groups. The partitioning (“splitting”) process in this study used a conditional inference framework and comprised two steps: The first step consisted of testing the global null-hypothesis of independence between all included predictors and the outcomes. If the null hypothesis of independence was rejected, a second step was initiated where the algorithm identified the predictor with the strongest association to the outcome, and the data were split on this variable. Splits were always in two, and for continuous and categorical variables the algorithm identified optimal cut-offs (i.e., the cut-off that maximized the difference in outcomes between the two groups). The process is recursive in the sense that the splitting process is repeated until the null hypothesis of independence in the first step cannot be rejected, and then the algorithm stops. An α of 0.05 was used for this analysis, and the p values were adjusted for multiple testing using the Bonferroni correction [37, 38]. Finally, we compared the treatment persistence in the identified groups and combined groups for which treatment persistence was not statistically different from each other (p > 0.05).

The analysis was performed in R version 3.54.2. Recursive partitioning was implemented using ctree from the partykit package [35]. Differences in drug survival between different identified groups were tested using log-rank tests starting with the two groups with the lowest treatment persistence.

The overall study population included 13,913 patients (Fig. 1). Approximately 50% of the patients had RA and a quarter of patients had PsA and AS/uSpA, respectively (Table 2). Patients with RA were predominately women (76%) whereas the sex distribution in AS/uSpA (47 women) and PsA (54% women) was more even. In the year prior to index date, most patients had taken NSAIDs whereas the use of csDMARDs and systemic steroids varied more across indications.

The recursive partitioning identified two groups with distinct treatment persistence in RA, four in PsA and four in AS/uSpA (Fig. 2, Table 3). In RA, the two groups were combinations of three and four separate subgroups with similar treatment persistence, respectively (Fig. 2, Table 3).

The factors most strongly associated with SC-TNFi persistence differed between RA and spondyloarthropathies as indicated by the first split in the trees (Fig. 2). In RA, the number of outpatient visits the year prior to treatment start was most influential: Patients with few outpatient visits (≤ 3 visits) had better treatment persistence than patients with many outpatient visits (> 3 visits). In contrast, in both AS/uSpA and PsA, sex was the most influential variable. Men had better treatment persistence than women across both indications.

In RA, there was substantial variation after the initial split, and prior csDMARD treatment appeared to be the most influential variable overall (Fig. 2A, D, Table 3): All patients in the group with high SC-TNFi survival were treated with csDMARDs prior to treatment start, whereas two of the four subgroups in the group with low treatment persistence were not. The subgroups with low treatment persistence without prior csDMARDs were characterized by a (1) low age (≤ 39 years), female sex and a high number of outpatient visits (n > 3) and (2) high age (> 39 years) and a very high number of outpatient visits (> 12 visits).

There was also variation in the treatment persistence of SC-TNFi in PsA and AS/uSpA. However, the initial split on sex appeared to be the most important overall with all groups comprising men having higher treatment persistence than all groups comprising women. In PsA (Fig. 2B, E, Table 3), steroids and specific SC-TNFi agent were also associated with treatment persistence in men: The highest treatment persistence was observed in patients without prior steroids treated with golimumab. Among women, no further subgroups were identified after the initial split, and the group with the lowest treatment persistence comprised women. In AS/uSpA (Fig. 2C, F, Table 3), men treated with golimumab had the highest treatment persistence whereas women < 45 years had the lowest treatment persistence.

Time on treatment in the group with the lowest treatment persistence was similar across all three indications: The median time on treatment in these groups ranged from 9.5 [95% CI 8.8–10.7] months in PsA to 11.3 [9.7–14.0] months in AS/uSpA (Table 3). In contrast, there was more variation in time on treatment across the groups with the highest treatment persistence: The median time on treatment in these groups ranged from 18.4 [17.0–19.6] months in RA to 48.9 [34.7–69.7] months in PsA (Table 3).

The Cox-PH analyses, presented in Table S1 in the electronic supplementary material, identified more predictors than the recursive partitioning analysis. However, in AS/uSpA, age was not identified as a significant predictor in either the bivariable or multivariable models Cox PH models.