Comparative effectiveness of dimethyl fumarate in multiple sclerosis

To assess the effectiveness of dimethyl fumarate (DMF) on annual rate of relapse subject to treatment (ARRt) and disability progression in multiple sclerosis (MS) compared to injectable immunomodulators (IMM), teriflunomide (TERI) and fingolimob (FTY), in real‐life setting.


| INTRODUCTION
To date, disease-modifying therapies (DMTs) represent the main therapeutic strategy in relapsing-remitting multiple sclerosis (RRMS), to reduce the risk of relapse and delay disability progression. The first generation of medications approved were the injectable immunomodulators (IMM) such as interferon beta-1a and 1b (IFN) and glatiramer acetate (GA). Since then, treatment options have broadened to include the orally administered DMTs fingolimod (FTY), which is predominantly indicated as second-line therapy in Europe, and more recently, teriflunomide (TERI) and dimethyl fumarate (DMF). All these drugs showed a significant treatment effect compared with placebo on the occurrence of relapses, disease activity and disability progression. 1 Although head-to-head randomized clinical trials have been performed for IMM, FTY and TERI, 1 well-designed head-to-head trials are lacking for DMF. Its clinical efficacy cannot thus be directly compared with other oral DMTs. A systematic review and meta-analysis of randomized clinical trials found that DMF significantly reduced the occurrence of relapse compared to IFNs, GA and TERI. 2 However indirect comparison studies are not sufficient to conclude the superiority of DMF due to the variability of study population and endpoint definitions.
Many observational studies have been used to assess the realworld comparative effectiveness of DMF and a number of alternative treatments, using claims-based analyses or registries with mixed results. The variability of these results may be explained by the various durations of follow-up limited to 1 year in some studies, [3][4][5][6] or by the use of simple propensity score (PS)-based methods in other studies, 3,5,[7][8][9][10][11] which, although they can balance observed baseline covariates between groups, do nothing to balance unmeasured characteristics and confounders. In this context, comparative studies from real-world practice in large population-based healthcare databases using robust statistical methods to handle confounders are needed to provide valuable evidence that will help clinicians in the choice of treatment.
The aim of this study is to assess the effectiveness of DMF in comparison with the injectable or oral DMTs used in RRMS, in terms of frequency of treated relapses and disability progression using the French nationwide claims and hospital database, SNDS (Système National des Données de Santé).

| Design and population
We conducted a cohort study within the general scheme of the SNDS claims database. It included all naïve patients initiating an IMM (IFN or GA) or an oral DMT (FTY, TERI or DMF) between 1 July 2015 and 31 December 2017, with a 4.5-year history of data. The index date was the earliest date of dispensing of IMM or oral DMT or of hospitalization for administration of drugs that are used under temporary use authorization (i.e., French procedure allowing the use of a drug before its market authorization) or reimbursed in addition to the procedurebased hospital payment system. Initiation was defined as having no dispensing or hospitalization for one of these drugs or any other drugs for MS (i.e., natalizumab, methotrexate, cyclophosphamide, mycophenolate, azathioprine, rituximab or tacrolimus) during the 4.5-year history period. MS patients were followed from the index date until index treatment switch or discontinuation (i.e., no dispensing of the index drug during 60 days after the end of the last dispensing), death or 31 December 2018, whichever occurred first.

| Data source
The SNDS has been described in detail elsewhere. 12 showed a significant effect compared with placebo in clinical trials on the occurrence of relapses, disease activity and disability progression.
• Results of indirect comparisons suggested that DMF significantly reduces the occurrence of relapse compared to other RRMS treatments, but well-designed head-to-head trials are lacking to draw definitive conclusions.
• Current findings on DMF effectiveness from observational studies are mixed and additional data are needed to define the place of this drug within the current treatment strategy.

What this study adds
• This study highlighted that DMF was associated with better results than teriflunomide (TERI) and injectable immunomodulators (IMM) regarding relapse activity requiring steroid treatment.
• However, no significant difference was found on surrogate measures of disability.
• These data will be useful to feed into physician choices of patient treatment. BOSCO-L

| Outcomes
The primary outcome was the annual rate of relapse subject to treatment (ARRt) during the index treatment period. Treated relapses were identified through an algorithm, developed based on national guidelines 15 and clinicians' medical expertise, that included dispensing of a high dose of corticosteroids for outpatients and hospitalizations with MS relapse diagnosis potentially followed or preceded by dispensing of a high dose of corticosteroids (Table S1). To be considered as independent events, treated relapses must be separated by at least 31 days. The diagnostic performance of this algorithm was assessed in a validation study with 95.0% positive predictive value. The secondary outcomes were MS disability progression during the index treatment period defined according to new reimbursements related to equipment for motor or sphincter disability or the implementation of a neuromodulation device for the treatment of chronic pain identified in the index period compared with the pre-index treatment period.

| Statistical analysis
Baseline characteristics included treated relapses, MS-related hospitalizations, steroid use, medical visits to the neurologist, lab test and encephalic or spinal cord magnetic resonance imaging in the 2 years prior to the index date. Data related to patient comorbidities and disability were collected during the 4.5 years prior to index date in order to ensure the completeness of the information. Chronic disease burden was measured using a version of the Charlson Comorbidity Index (CCI) score adapted for the SNDS database. 16 The probability of discontinuation or switch of the index treatment as well as the probability of the first treated relapse occurrence during the index treatment period were described for each treatment group using Kaplan-Meier survival analysis. Head-to-head comparisons were performed in three separate matched analyses of DMF vs IMMs, DMF vs TERI or DMF vs FTY based on an "as treated" analysis. For each head-to-head comparison, we computed a high dimensional Propensity Score (hdPS), which reflects the probability to be treated with DMF vs one of the other treatments, given the comparison of interest, 17,18 adjusting for hidden or unknown confounders. 19 The first step of the development of the hdPS algorithm is the creation of data dimensions within the healthcare database to distinguish various subsets of information, with codes mostly related to inpatient diagnoses, outpatient procedures and outpatient pharmacy drug dispensing. The hdPS algorithm considers distinct codes in each dimension without needing to understand their medical meaning and creates binary variables indicating the presence of each factor during a defined pre-exposure covariate assessment period. The hdPS considered the most prevalent codes in each data dimension and for each code created three binary variables, indicating at least one occurrence of the code, sporadic occurrences, and many occurrences during the covariate assessment period. These variables are then selected according to their greatest potential to adjust for confounding using a formula by Bross that depends on the covariate-outcome and covariate-exposure associations. The top-ranked variables are selected, and then entered into a logistic regression propensity score model. 20 Some of the covariates selected by the hdPS algorithm may not be direct confounders but may actually be proxies of unmeasured or unknown confounders. Adjusting for a perfect proxy of an unmeasured confounder is equivalent to directly adjusting for this confounder. In that way, the hdPS is able to adequately adjust for confounders that are hidden to the algorithm or unknown by the investigator. 19 It was demonstrated that this approach achieved more plausible effect estimates than conventional PS modelling based on clinically selected variables or simple multivariable modelling. In the present study, the hdPS was estimated using a multivariable logistic regression model with treatment group as dependent variable and a large data set of independent variables collected in the pre-index period provided by six data dimensions (i.e., outpatient drug dispensing, diagnoses related to hospitalizations and LTD registrations, outpatient and inpatient medical and paramedical visits, lab tests, medical procedures and medical devices) and fixed baseline characteristics (i.e., age, sex, number of treated relapses, MS medical device and outpatient and hospital costs identified in the preindex period). The variables were selected on a bias-based approach.
The number and the type of variables selected in the hdPS are detailed in the Table S2 in the Supporting Information. We excluded the subjects in both treatment groups who were at or below the 2.5th percentile of the hdPS in the group of subjects who received the treatment predicted by the hdPS, as well as those at or above the 97.5th percentile in the alternative treatment group. This hdPS trimming ensures that each subject has a reasonable probability of receiving either compared treatment, given relevant confounding variables. 21 The remaining patients were 1:1 matched for the comparison. The analysis of an hdPS matched sample can mimic that of an RCT: one can directly compare outcomes between treated and untreated subjects within the hdPS matched sample. 22 We assessed the predictive performance of the hdPS using the c-statistic calculated after matching and the balancing effect of the hdPS matching with standardized differences in baseline variables before and after matching, knowing that an absolute standardized difference of 10% or less indicates a negligible difference between groups. 22 For treatment effects, we estimated for the ARRt, the rate ratio (RR) and its corresponding confidence interval (95% CI) by negative binomial regression and for the disability progression, the odds ratio (OR) and its 95% CI by logistic regression. For each outcome, estimates were calculated crude and after hdPS matching. Because some individuals end up not matched and hence excluded after hdPS matching, and therefore resulting in a loss of both precision and generalizability, we performed sensitivity analyses with two other hdPS-based methods including all patients discarded by the previous approach: an adjusted analysis on hdPS deciles and an analysis modelled by the inverse probability of treatment weights (IPTW), which is an hdPS score-based weight used to control for confounding by indication. 23 We performed statistical analysis using SAS software (version 9.4; SAS Institute, Cary, NC) and hdPS using the routines from Harvard Medical School. 24 3 | RESULTS were female with an age of 39.9 (standard deviation, SD: 12.1) years on average, ranging from 37.5 (11.9) years for IMM to 43.1 (11.7) years for TERI (Table S1). The mean CCI score was 0.

| DMF versus TERI
Of the 2697 DMF patients and the 3089 TERI patients, 571 (21.2%) and 402 (15.0%), respectively, were excluded by trimming, and 1679 patients were then matched in each group with a satisfying hdPS overlapping ( Figure S1 in the Supporting Information) and a c-statistic at 0.56, as well as good balance on all covariates (  Figure 4).

| DISCUSSION
With the increasing number of DMTs developed in these last 10 years, and the lack of head-to-head randomized controlled trials to assess their comparative efficacy, robust observational data studies are needed to support decision making by stakeholders and to assist clinicians in choosing the most favourable treatment option for their patients. This nationwide population-based observational study conducted to assess the effectiveness of DMTs in a population of DMT initiators on MS activity using robust hdPS-based methods found that the DMF treatment proved to be associated with better results than TERI and IMM regarding relapse activity, without significant difference on disability progression. F I G U R E 2 Probability of first relapse occurrence during the index treatment exposure period (Kaplan-Meier curve), according to treatment groups: dimethyl fumarate (DMF), teriflunomide (TERI), fingolimob (FTY), injectable immunomodulatory drugs (IMM) The 9304 individuals newly treated with a DMT identified in the SNDS are broadly consistent with the national epidemiological estimates that evaluates at around 110 000 individuals the prevalence of MS in France, and from 4000 to 6000 its annual incidence. 25 The number of patients included are a little lower than expected probably because patients initiating an MS medication other than the specific ones (i.e., natalizumab, methotrexate, cyclophosphamide, mycophenolate, azathioprine, rituximab or tacrolimus) were not included in this cohort. At treatment initiation, DMF patient characteristics were very similar to those of TERI and IMM patients: their age was 40 years on average, they were mostly female and had an annual rate of treated relapses of 0.13 during the pre-index period.
These findings differ slightly from those of randomized clinical trials (i.e., age, duration of drug exposure period), 26     Although it underestimates the real number of relapses, our algorithm identifies treated relapses with a high level of accuracy (PPV = 95%) and has allowed strong mitigation of any potential misclassification bias. To strengthen the validity of this algorithm, we also conducted an additional analysis by extending the period between two occurrences of treated relapses from 31 to 60 days, which did not affect fundamentally the performance of the algorithm. Some treated relapses may nevertheless not have been captured, and frequency of relapses could be underestimated, but this should affect all treatment groups similarly and thus not bias the results.
Like any claims-based real-world studies, this study presents an inherent risk of unmeasured confounding. To address this limitation, we applied hdPS-based methods in the analysis of the outcome. The hdPS is a well-known statistical technique that attempts to estimate the effect of a treatment, policy or other intervention by accounting for the measured and unmeasured covariates that predict receiving the treatment. It summarizes a large set of variables that characterize each subject for status and unmeasured confounders not recorded in a database (i.e., drugs, medical status, hospitalization, other comorbidities directly, or indirectly linked with unmeasured confounders). 17,18 Matching, adjusting or weighting on large numbers of covariates ascertained from subject healthcare claims data may improve control of confounding, as these variables may collectively be proxies for unobserved factors. HdPS matching appears to be a reliable method, in that it provides excellent covariate balance in most circumstances. It has the advantage of being simple to analyse, present and interpret. Matching can eliminate a greater portion of bias when estimating the more precise treatment effect as compared to other approaches. It creates a balanced dataset, allowing a simple and direct comparison of baseline covariates between treated and untreated patients. The analysis of an hdPS matched sample can mimic that of an RCT: one can directly compare outcomes between treated and untreated subjects within the matched sample. 32 Unlike covariate adjustment using the propensity score, propensity score matching does not require that an outcomes model be specified. In contrast, covariate adjustment using the propensity score requires that the outcomes model be correctly specified but searchers using covariate adjustment with the hdPS rarely examine the fit of their model compared to more complex models. 33 A further technical detail is that hdPS matching (such as IPTW method) estimates a marginal treatment effect (i.e., average effect of treatment on the population), whereas multivariate regression estimates a conditional treatment effect (i.e., average effect of treatment on the individual). Furthermore, using hdPS adjusting or weighting methods in addition to matching allowed us to include all patients meeting the eligibility criteria of the study, whereas matching method excludes patients not finding a match.
Using hdPS weighting or adjustment ensures transparency of included patients and validity of results in the predefined study population.
In conclusion, this study provides further insight into the therapeutic benefit of DMF in the real-life setting compared to other commonly used agents for RRMS, including IMM and another oral drug, TERI. The ARRt was significantly lower in patients treated with DMF vs IMM and TERI using robust hdPS-based methods. These data will be useful to feed into physician choices of patients' treatment.