Evaluating Heterogeneous Treatment Effects in Pursuit of Personalized Multiple Sclerosis Care

Multiple sclerosis (MS) is heterogeneous in its presentation, disease course, and treatment response. Recent advancements have led to more effective MS therapies and improved longitudinal outcomes, while emphasizing the considerable diversity among people with MS (pwMS) and the need for a more individualized care approach.

Treatment selection is often highly variable among MS providers¹ and is challenged further by heterogeneity in treatment response among pwMS. This has led to an increased interest in using precision medicine to individualize care for pwMS with the hope of yielding more effective disease control. Understanding and implementing heterogeneous treatment effects (HTEs) into the decision-making process is one such approach.

The Utility of HTEs for Precision Medicine

MS randomized controlled trials (RCTs) have traditionally focused on more homogenous populations, often restricted in age (18–55 years), and tend to include participants who are White, are female, and have minimal comorbidities.² Although outcome variability is reduced, the lack of representative real-world cohorts weakens external validity. This lack of diversity in MS RCTs has been addressed by well-designed observational studies which generate important insights into treatment outcomes among underrepresented pwMS.^2–4 Recently, this has prompted clinical trials to begin focusing on treatment outcomes in minority populations (eg, Prospective Study to Assess Disease Activity and Biomarkers in Minority Participants With Relapsing Multiple Sclerosis After Initiation and During Treatment With Ocrelizumab, NCT04377555), although these remain relatively small in number.⁵

RCTs provide estimates of average treatment effects (ATEs), defined as the average difference in effect size between the intervention and control cohorts generalized to the study population.^2,6 Whereas a treatment may show the average efficacy for pwMS, it may not reflect true efficacy for an individual, especially among individuals from underrepresented groups. Furthermore, ATEs do not typically account for important indicators of treatment selection and response, including medical comorbidities, social determinants of health, or genetic or immunobiologic differences.

RCTs have used subgroup analyses to attempt to address some of the issues associated with underrepresentation in study populations. However, this approach is inherently limited by strict inclusion and exclusion criteria of the parent trial, reflexively limiting heterogeneity. Therefore, relying on RCT subgroup analyses for clinical decision-making presumes that the risk-benefit tradeoffs are similar for all study participants. Furthermore, subgroup analyses are prone to smaller sample sizes with lower statistical power, and the analyses are often performed with limited forethought on relative effect modifiers, resulting in false-positive findings.^6,7 Such limitations are supported by a recent appraisal involving a systematic review of racial and ethnic characteristics of pwMS enrolled in phase 3 trials investigating disease-modifying therapies (DMTs), which found that although ~70% of included RCTs conducted subgroup analyses, they only acquired 29% power to detect an interaction effect.⁷ Thus, subgroup analyses alone may not sufficiently inform treatment effects across the heterogeneity of pwMS.

The limitations of ATEs have led to an increasing interest in the use of HTEs: the nonrandom variation in the magnitude or direction of a treatment’s effect on a specific clinical outcome based on one or more covariates defining a specific subgroup.^6,7 HTEs can maximize the potential effectiveness of DMTs at an individual level and inform the development and use of future treatments. Demographic, clinical, and radiographically meaningful covariates such as sex, race, age at MS onset, disease duration, disease course, MRI lesion burden and activity, disability level, and individual preferences are important considerations when selecting a DMT. Similarly, immunobiologic differences, genetic variations, environmental factors, health behaviors, comorbidities, and social determinants of health can affect the efficacy and tolerability of DMT, but are not used uniformly in clinical practice to guide treatment selection. The use of HTEs in MS research allows for a better understanding of how inherent baseline factors, along with clinically meaningful MS-specific covariates, are implicated in an individualized response to DMTs, and may help shape a more structured approach to DMT selection.

Methodologies for Assessing HTEs

The limitations of conventional subgroup analyses in RCTs have led to more predictive strategies for identifying HTEs. Guidelines for improved accuracy of HTE identification in RCTs have been outlined in the form of the Predictive Approaches to Treatment Effect Heterogeneity (PATH) statement,⁸ which proposes to consider multiple individual characteristics simultaneously and generate predictions that are applicable at the individual level. Two potential approaches for satisfying these criteria are risk modeling and effect modeling (Figure).⁸

Risk modeling involves generating a multivariable regression model from reference RCT data and assessing the outcome risk of interest based on individual-associated covariates. Risk modeling is generated without accounting for treatment assignment, a defining feature of this approach. Trial participants are disaggregated and stratified according to risk, and differences in treatment outcomes are assessed. The lowest-risk and highest-risk participants may be assigned to the 1st and 4th quantiles, respectively.^7–9 A study might reveal, for example, that the 4th quantile had the most significant decrease in absolute relapse risk, and the 1st quantile exhibited no change in risk when compared with the original model. A clinician could thus designate an individual to a risk quantile based on numerous characteristics simultaneously and estimate the potential treatment effect.

Conversely, effect modeling uses RCT data to design a regression model involving treatment assignment in addition to risk predictors and treatment interaction terms with treated and untreated groups considered separately.^7,8 This approach has the advantage of revealing differences in treatment effects both across risk strata and between treated and untreated participants.

Each approach has a unique set of advantages, which can inform more personalized DMT selection. Risk modeling benefits from being independent of treatment assignment, allowing for an assessment of the absolute risk difference across a cohort even when the relative treatment effect is similar.⁷ Alternatively, the use of treatment effect estimation in effect modeling can identify individuals who may benefit most from a specific treatment, which may increase its interpretability among clinicians.⁷ These methodologies are not without their own set of limitations, which are discussed in a later section.

Clinical Applications of HTEs in MS

HTE methodologies in MS may optimize DMT selection beyond ATEs alone. For example, Chalkou et al¹⁰ applied a 2-stage HTE approach (eg, risk modeling) to a network meta-analysis of 3 pivotal RCTs comparing individuals with relapsing-remitting MS taking natalizumab (Tysabri; Biogen, Cambridge, MA), glatiramer acetate, or dimethyl fumarate (Tecfidera; Biogen, Cambridge, MA).^11–13 Stage 1 involved the development of a baseline risk score model linking the outcome probability to individual characteristics. Stage 2 estimated the probability of posttreatment outcome as a function of the baseline risk score. Using individual data from 3590 participants, the authors formulated a predictive tool for estimating the 2-year risk of relapse among the 3 treatment options. Baseline risk score modified the relative and absolute treatment effects, whereas age and disability status moderated the benefit expected for each of the treatments.

Hersh et al¹⁴ assessed the potential of 2-stage HTEs to evaluate the 1-year risk of brain atrophy, measured by brain parenchymal fraction (BPF), in 870 pwMS treated with low-, moderate-, or high-efficacy DMTs. Whereas many covariates significantly predicted BPF in the mixed-effects model, baseline T2 lesion volume, sex, and broad efficacy treatment group significantly predicted the rate of BPF change over time. People with a low risk of brain atrophy had similar predicted BPF change regardless of DMT selection. However, for people with high risk of brain atrophy at baseline, high- and moderate-efficacy DMTs performed similarly, whereas a 2-fold worse BPF change was predicted for individuals taking low-efficacy DMTs. These results highlight the potential to inform a more personalized treatment approach based on the heterogeneity of baseline characteristics.¹⁴

Kalincik et al¹⁵ used a modeling approach to develop predictive algorithms of individualized treatment response to 7 DMTs across 9193 participants in a large global cohort study. They identified an increased probability of disability progression occurring with interferon-Β or glatiramer acetate treatment in individuals with higher disability at treatment initiation or with a history of severe relapses with incomplete recovery. In addition, there was a higher risk of disability progression in individuals with decreased relapse activity in the year before initiating treatment with fingolimod (Gilenya; Novartis, East Hanover, NJ), natalizumab, or mitoxantrone. As anticipated, the incidence of relapse was associated with younger age, a relapsing disease course, and pretreatment relapse activity, with the strength of these associations varying across therapies.¹⁵

A Growing Arena of Precision Medicine in MS

HTE represents just one aspect of precision medicine that is designed to tailor treatments to patients based on individual characteristics. Other facets of precision medicine, such as incorporating genetic and immunobiologic variability into the assessment of individual response predictions, are also valuable. For example, an evaluation of glatiramer acetate and interferon-Β treatment response in individuals with the HLA DRB1*1501 allele found a modest association between homozygosity and longer event-free survival while treated with glatiramer acetate compared with nonhomozygous individuals or homozygous individuals treated with interferon-Β.¹⁶ Similarly, an IRF8 polymorphism appeared to be associated with event-free survival in individuals treated with interferon-Β.¹⁶ One study evaluated peripheral blood cellular immune signatures of 309 early, treatment-naïve pwMS, and demonstrated 3 distinct MS immunologic endophenotypes, with endophenotypes 1, 2, and 3 involving alterations in CD4+ T-cell–associated neurodegeneration, natural killer cells, and CD8+ T-cell–associated inflammation, respectively. These endophenotypes may represent the heterogeneity observed across MS pathologies and were associated with differences in the efficacy of interferon-Β treatment.¹⁷

The pursuit of precision medicine for treatment of pwMS may be further supported using advanced machine learning techniques, such as deep-learning predictive enrichment strategies. This methodology was applied to data from 6 RCTs by Falet et al,¹⁸ using machine learning to predict which pwMS were most likely to respond to DMT based on clinical and imaging features. When randomizing individuals to treatment with either ocrelizumab (Ocrevus; Genentech, South San Francisco, CA) or laquinimod compared with placebo, the ATE was larger for pwMS designated as responders compared with the entire group, demonstrating that machine learning using a model for predictive enrichment can be an effective tool for evaluating HTEs.¹⁸

Barriers and Future Directions

The data-driven approach to HTEs, especially relating to effect modeling, may be prone to issues identified with the conventional one-variable-at-a-time method, such as overfitting and inadequate power. HTEs may also be prone to bias without implementing prevention strategies, such as rigorous internal validation approaches. Because previous knowledge of effect modifiers is typically limited, interpretation of statistically significant results can pose a challenge and must be approached cautiously.^8,19

These limitations highlight barriers to using HTEs in MS trials. Whereas the proposed approaches allow for consideration of multiple individual characteristics simultaneously and improve upon conventional techniques, they do not fully address that every individual with MS has a combination of features that makes them unique, and may or may not be included in the aforementioned prediction models. Personalization of medical treatment remains dependent on reference class forecasting (ie, the use of evidence from similar individuals to support clinical decision-making). Selection criteria can be established in RCTs to guarantee the existence of any number of similarities among a study population; however, each participant will differ from one another in numerous ways that may not be fully understood or captured.⁶ Furthermore, machine learning and artificial intelligence models hold promise, but they will need to be validated in multiple data sets and with diverse MS populations. The existence of large global databases of pwMS provides robust data from which predictive models can be generated and validated. This may combat some of the inherent issues of reference class forecasting, but certainly does not solve them. The Box provides an example demonstrating how physicians can implement HTEs to support a shared decision-making approach that improves patient outcomes.

Ongoing research into specific pharmacogenomic biomarkers may further enhance our understanding of HTEs in MS. Biomarkers such as variations in TPMT and CYP2C9 gene expression can inform treatment response, but examples are limited and more often related to medication safety.^7,20 Identification of biomarkers that augment treatment effect, through effect modeling or other means, could yield considerable advances in personalization of MS care.

Conclusion

Whereas the availability of numerous MS therapies allows for treatment flexibility relating to safety and efficacy, medication selection is often based on data derived from ATEs. An increased focus on HTEs and their implementation into a more personalized treatment approach has the capacity to further improve clinical outcomes. Conventional RCT subgroup analyses may have utility in generating hypotheses for potential treatment heterogeneity, but are often inherently limited. Prediction models offer an opportunity to individualize treatment selection and deserve further refinement and incorporation into real-world practice. As these models are explored further, the continued development of and nuances associated with HTEs may allow for increasingly precise and personalized care for pwMS.