Just a moment, the page is loading...
Browse ALL STUDIES
Keyword Search
View All Selected
Clear All
Login / Create Account
Login
Create Account
Home
About Us
Privacy Policy
Minimum System Requirements
How To Join
Mission
Data Sponsors
Researchers
How It Works
How to Request Data
Review of Requests
Data Sharing Agreement
Access to Data
Independent Review Panel
Metrics
FAQs
News
Help/Contact Us
Individualizing Treatment Decisions for Adult Relapsing-Remitting Multiple Sclerosis Patients Using Statistical Learning Methods
Proposal
11223
Title of Proposed Research
Individualizing Treatment Decisions for Adult Relapsing-Remitting Multiple Sclerosis Patients Using Statistical Learning Methods
Lead Researcher
Prof Dr Ulrike Held
Affiliation
Department of Biostatistics, Epidemiology, Biostatistics and Prevention Institute, University of Zurich, Switzerland
Funding Source
Potential Conflicts of Interest
Data Sharing Agreement Date
07 July 2020
Lay Summary
Multiple Sclerosis (MS) is an incurable disease of the nervous system. Many disease-modifying treatment options have been developed for MS. Individual response to these therapies varies significantly across people with MS. Specific treatments may be effective in some patients while harmful in others. Due to heterogeneous disease courses and benefit-harm balance, it is not clear which treatments provide the greatest net benefit. This question is of relevance both before starting treatment and after initiation. Also, observed disability worsening in MS usually occurs over a long period. Hence, short-term disease markers, such as lesions in brain images, are used in trials to assess a treatment's overall effect. It is not certain if these short-term markers could be reliably used in place of disability worsening in individual patients. Personalized treatment benefit/harm predcition could be easier using these short-term effect.This study will include adult multiple sclerosis patients with relapsing-remitting type. These patients will be participants in clinical trials of drugs available in the market. To answer the above questions, we are going to look at how the benefit or harm of the given treatment in clinical trials changes with different individual patients' characteristics. Many machine-learning and statistical methods will be used to detect the relationship between individual characteristics and treatment-specific response. The predictive ability of these methods for beneficial or harmful outcomes will be evaluated. These outcomes include disability worsening, acute neurological symptoms (relapses), lesions in brain images, adverse events, and infections. The relationship between short-term effects and disability worsening per patient will also be investigated.
Study Data Provided
[{ "PostingID": 4084, "Title": "NOVARTIS-CFTY720D2301", "Description": "A 24-month, Double-blind, Randomized, Multicenter, Placebo-controlled, Parallel-group Study Comparing the Efficacy and Safety of Fingolimod 1.25 mg and 0.5 mg Administered Orally Once Daily Versus Placebo in Patients With Relapsing-remitting Multiple Sclerosis" },{ "PostingID": 4085, "Title": "NOVARTIS-CFTY720D2302", "Description": "A 12-month Double-blind, Randomized, Multicenter, Active-controlled, Parallel-group Study Comparing the Efficacy and Safety of 0.5 mg and 1.25 mg Fingolimod (FTY720) Administered Orally Once Daily Versus Interferon ß-1a (Avonex) Administered im Once Weekly in Patients With Relapsing-remitting Multiple Sclerosis With Optional Extension Phase" },{ "PostingID": 4089, "Title": "NOVARTIS-CFTY720D2309", "Description": "24-month Double-blind, Randomized, Multicenter, Placebo-controlled, Parallel-group Study Comparing the Efficacy and Safety of 0.5 mg and 1.25 mg Fingolimod (FTY720) Administered Orally Once Daily Versus Placebo in Patients With Relapsing-remitting Multiple Sclerosis With Optional Extension Phase" },{ "PostingID": 4899, "Title": "NOVARTIS-CFTY720D1201", "Description": "A 6-month, Double-blind, Randomized, Placebo-controlled, Parallel-group, Multicenter Study Comparing Efficacy and Safety of FTY720 0.5 mg and 1.25 mg Administered Orally Once Daily in Patients With Relapsing Multiple Sclerosis" },{ "PostingID": 4900, "Title": "NOVARTIS-CFTY720D1201E1", "Description": "An Extension of the 6-month, Double-blind, Randomized, Placebo-controlled, Parallel-group, Multicenter Study Comparing Efficacy and Safety of FTY720 0.5 mg and 1.25 mg Administered Orally Once Daily in Patients With Relapsing Multiple Sclerosis" },{ "PostingID": 14593, "Title": "ROCHE-WA21092", "Description": "A Randomized, Double-Blind, Double-Dummy, Parallel-Group Study To Evaluate the Efficacy and Safety of Ocrelizumab in Comparison to Interferon Beta-1a (Rebif®) in Patients With Relapsing Multiple Sclerosis" },{ "PostingID": 14594, "Title": "ROCHE-WA21093", "Description": "A Randomized, Double-Blind, Double-Dummy, Parallel-Group Study To Evaluate the Efficacy and Safety of Ocrelizumab in Comparison to Interferon Beta-1a (Rebif) in Patients With Relapsing Multiple Sclerosis" },{ "PostingID": 14596, "Title": "ROCHE-WA21493", "Description": "Phase II, Multicenter, Randomized, Parallel-Group, Partially Blinded, Placebo and Avonex Controlled Dose Finding Study to Evaluate the Efficacy As Measured by Brain MRI Lesions, and Safety of 2 Dose Regimens of Ocrelizumab in Patients With RRMS" },{ "PostingID": 19692, "Title": "SANOFI-CAMMS223", "Description": "A Phase II, Randomized, Open-Label, Three-Arm Study Comparing Low- and High-Dose Alemtuzumab and High-Dose Subcutaneous Interferon Beta-1a (Rebif®) in Patients With Early, Active Relapsing-Remitting Multiple Sclerosis" },{ "PostingID": 19693, "Title": "SANOFI-CAMMS323
(CARE-MS I)", "Description": "A Phase 3 Randomized, Rater-Blinded Study Comparing Two Annual Cycles of Intravenous Alemtuzumab to Three-Times Weekly Subcutaneous Interferon Beta-1a (Rebif®) in Treatment-Naïve Patients With Relapsing-Remitting Multiple Sclerosis" },{ "PostingID": 19694, "Title": "SANOFI-CAMMS32400507
(CAMMS324/ CARE-MS II)", "Description": "A Phase 3, Randomized, Rater- and Dose-Blinded Study Comparing Two Annual Cycles of Intravenous Low- and High-Dose Alemtuzumab to Three-Times Weekly Subcutaneous Interferon Beta 1a (Rebif®) in Patients With Relapsing Remitting Multiple Sclerosis Who Have Relapsed On Therapy" }]
Statistical Analysis Plan
The study populations of interest will be the safety- as well as the ITT and PP populations depending on the objective. The main goal of the following analysis is the identification of specific predictive models. Therefore, we concentrate on splitting the data into a discovery and a validation set. We also will use simulation studies to validate our models. Since our approach does not have an inferential goal, we do not specify levels of statistical significance in advance, although this might be needed in some of the below-proposed algorithms.Data exploration: The relevant content of the data is presented in tabulated as well as graphical form. Information on baseline variables, time structure of the data, treatment effects, and structural comparison between the eleven studies of interest is given. These aspects can be specified in detail after accessing the data.The first objective (O1) focuses on different elements of treatment decision scores for people with relapsing-remitting multiple sclerosis (RRMS). It studies patient by treatment interaction (PbTI), the interaction between the characteristics of the individual patient and the treatment given, based on information collected at either baseline or shortly after treatment start. Response is regarded to mean not only treatment benefits (aspects of controlling the disease course) but also its nuisance or disadvantages (adverse effects, death) patients may suffer. The purpose is to provide scores of P(Y|X,A) with Y the corresponding binary outcome, X the baseline covariates, and A the treatment given. The PbTI is incorporated in the modeling of the relationship between (X, A) variables and Y. Lipkovich et al. (2016) provide an extensive review of data-driven yet statistically principled methods for identifying PbTI in randomized trials and estimating the outcome probabilities P(Y|X, A). These primarily search for treatment × covariate interactions, i.e., for predictive factors that modify treatment effects. We will use and, in a validation set, compare the results of three methods that already have implementations: Virtual Twins (Foster et al. 2011), Outcome Weighted Learning (Yu et al. 2015), and Model-based Recursive Partitioning alongside with its generalization to random forests (Seibold et al. 2016).The data applied for allows studying PbTI in selected treatment pairs: Fingolimod vs. Placebo, Fingolimod vs. Iβ1-a, Ocrerlizumab vs Iβ1-a, and Alemtuzumab vs. Iβ1-a. For specific pairs, there exist different dosing and follow-up schemes. These define the range of optional treatments. In the main analysis we are initially going to assume that different doses of the same active ingredient does not make a difference in treatment decision optimization but we will do a sensitivity analysis to test this assumption. These methods also explore PbTI with regard to the endpoints sketched in O1a to O1f. These endpoints specify the different outcome values. The baseline measurements in the data of the respective studies are the covariates which are potentially interacting with treatment. Given a patient with specific covariates, we can define a treatment decision for a specific class of “benefit” outcomes for which the probability of the outcome is maximized given the patient's covariates and treatment options. Similarly, given a patient with specific covariates, we can define a treatment decision for a specific class of “harm” outcomes for which the probability of the outcome is minimized given the patient's covariates and treatment options.In a second step, we are interested in models which combine benefit and harm. A cost function that incorporates individual benefit and harm outcomes will be defined to balance the individual preferences between treatment benefit and treatment harm. This approach uses the methods mentioned at the beginning of the paragraph (Lipkovich et al. (2016)).We will also try the approach by Zhang et al (Biometrics 2015) to define a population-based rule (treatment regimen). A treatment regimen formalizes subgrouping as a function from individual patient characteristics to a recommended treatment. They propose to derive a list of if-then statements (treelike structure), which are immediately interpretable and are therefore an appealing choice for broad application in practice. We plan to derive a robust estimator of the optimal regime within this class and to demonstrate its finite sample performance using simulation experiments. The treatment regimen defines the treatment regimen as a function which relates the patient characteristics with a specific treatment in a specific population. In order to estimate and validate the optimal regimen, we follow Zhang et al. The second objective (O2) is to identify factors associated with heterogeneous treatment response and to build a prediction model for non-/response using the progression outcomes described earlier, covariates drawn from the baseline data, as in Pellegrini et al. (2019), and also data collected up to 6 months after treatment initiation, as in the later work on the Rio score described previously. Models with only baseline information will address the question of important predictors of treatment response prior to initiation, while models with post-treatment information address whether short-term response to treatment adds additional information. The overall objective is to inform treatment optimization decisions. This can be considered a search for subgroups of patients that respond to therapy and subgroups that do not respond to therapy. For this purpose we turn to model-based recursive partitioning and random forests. This consists of fitting a parametric model to the data, with the outcome modeled as a function of the treatment. The algorithm detects parameter instabilities with the use of (multiplicity-adjusted) permutation tests. An optimal cutpoint for the covariate associated with the greatest instability is then estimated and the patients split into subgroups based on this decision rule. This is repeated until stopping criteria are met, resulting in a tree. The tree depicts covariates associated with differential treatment effects and their relationships to each other. They also show the optimal cut-off values for each covariate split and finally the actual treatment effect estimates in the different subgroups. To build the forest, the algorithm is repeated for many random samples of the data, creating an ensemble of trees. This provides information on the similarity between patients with respect to the treatment effect and intercept. The hypothesis that the base model fits sufficiently well can be tested to check whether the personalized models improve the base model using bootstrap methods and empirical p-values.The predictions from the resulting models will be compared to predictions from current clinical decision algorithms, such as that proposed by the AAN, and recent prediction models from the literature, such as in Pellegrini et al. (2019). In addition to the traditional statistical measures for prediction model evaluation, such as sensitivity and specificity, net benefit and more specifically decision curve analysis (Vickers et al. 2007, 2008, 2016) will be used. This measure quantifies the advantage conferred by a treatment algorithm compared to other possible treatment algorithms for a given benefit-harm threshold. The strategies of “treat none” and “treat all” are used as benchmarks. The net benefit can then be calculated for the treatment algorithm resulting from the recursive partitioning and random forest procedures, as well as other algorithms from the literature. The algorithm yielding the highest net benefit for a particular threshold value has the highest clinical value. We will also calculate concordance-statistic for benefit to measure the usefulness of our models.The definition of treatment failure may be improved by additionally considering a covariate that indicates that the patient has stopped treatment due to inefficacy. We therefore include analysis based on this modified definition (when such information is available in the dataset) as part of a sensitivity analysis. The third objective (O3) is to evaluate the surrogacy of presence of new or enlarged T2 MRI lesions for the two outcomes: annualized relapse rate and disability progression. We look at concurrent and predictive aspects. In all included studies that have data available on the surrogate, surrogacy analysis is explored by a meta-analytic approach at the individual level. We will use hierarchical models explained by Burzykowski et al. (2005) and summarized by Buyse et al. (2016). Correlation coefficients both at the individual level and across treatment arms from different studies will be reported.In concurrent surrogacy, we will analyze the relationship when the surrogate and clinical outcomes are measured at the end of study with respect to their baseline values. This concurrent surrogacy is informative for investigating homogeneity of association of surrogate outcomes with clinical outcomes.In studies lasting longer, it is of interest to assess the surrogacy of surrogate outcomes at each year to the long-term clinical outcomes. This individual-level predictive surrogacy is relevant for the traditional use of surrogate endpoint, where an early surrogate marker is indicative of a late-occurring clinical response to therapy.We also plan to develop models by meta-regression, which allow predicting the treatment effect on a clinical outcome from the treatment effect on the surrogate.General aspects:1. The Data Management will be planned and performed after having received the data. 2. As an additional step, we plan to perform network meta-analyses on the studies at hand to create treatment comparisons for each treatment pair. Bayesian approaches also allow ordering treatments with respect to their effects. It is of interest to compare this global ranking of treatment effects to the treatment decisions made for specific subgroups or for individual patients.3. Missing values in the covariates are handled by multiple imputation methods assuming missing at random (MAR). We will also perform sensitivity analyses to study missingness not at random (MNAR).
Publication Citation
Begüm Irmak Ön, Joachim Havla, Ulrich Mansmann. Multivariable prognostic prediction of efficacy and safety outcomes and response to fingolimod in people with relapsing-remitting multiple sclerosis, Multiple Sclerosis and Related Disorders, Volume 93 (2025)
DOI:
https://doi.org/10.1016/j.msard.2024.106247
eBook Prediction of prognosis and response to fingolimod in people with relapsing-remitting multiple sclerosis Beteiligte: Ön, Begüm Irmak; Mansmann, Ulrich
DOI: 10.5282/edoc.32918
© 2025 ideaPoint. All Rights Reserved.
Powered by ideaPoint.
Help
Privacy Policy
Cookie Policy
Help and Resources