Novel Methods for Modeling Treatment Sequencing
Martin Vu, MPH, BSc and Koen Degeling, PhD, MSc, BSc, Cancer Health Services Research, The University of Melbourne, Melbourne, Australia
The Bigger Picture: Modeling Sequential Management Decisions
Novel therapeutics are introduced at rapid speed, resulting in an ever-increasing number of potential management strategies for many diseases. This challenges the traditional approaches towards health economic evaluations, as they struggle to appropriately reflect the complexities in today’s clinical pathways. In this third and final spotlight session the panelists (Figure 1) discussed how individual-patient modelling techniques and causal inference methods using real-world data can be used to overcome these challenges.
Figure 1. (clockwise) Maiwenn J Al, Mark Sculpher, James Robins, Huajie Jin
Challenging the Status Quo
The growing number of available interventions emerging into clinical practice introduces substantial challenges from both a clinical and health economic perspective. Changes in up-front treatment will have downstream consequences, which may not only impact the effectiveness of subsequent treatment but also economic outcomes. Furthermore, the total number of possible treatment sequences may be large, especially considering that some drugs may be used throughout multiple lines of treatment by “rechallenging” the disease.
Maiwenn J Al, PhD, Institute of Health Policy & Management, Erasmus University Rotterdam, Rotterdam, The Netherlands, discussed that the 8 available treatments for metastatic renal cell cancer (mRCC) result in 40,320 possible sequences. Even though data from a registry in The Netherlands suggests the number of treatment sequences observed in the real world reduces to 336 given that very few patients will receive more than three lines of treatment, this still is a substantial number of possible management strategies.
The complexity is further increased when considering multiple outcomes, as was highlighted by Huajie Jin, MBBS, MSc, King’s Health Economics, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, London, United Kingdom. Clinical decision making and health economic analyses are based on multiple outcomes, such as overall survival, toxicities, adverse events and quality-of-life. Considering these multiple outcomes for events tens of sequences becomes unmanageably complex.
"Clinical decision making and health economic analyses are based on multiple outcomes, such as overall survival, toxicities, adverse events, and quality-of-life."—Huajie Jin, MBBS, MSc
Shifting the Perspective From the Population to the Individual
The majority of health economic evaluations has been model-based. When considering treatment sequencing, it is very unlikely that single randomized clinical trials will provide all evidence needed for a specific disease. Even if that would be the case, results from such a trial would soon be outdated when new therapeutics are introduced. The use of modelling methods therefore is evident.
When the number of sequences that need to be modeled increases, traditional techniques, such as decision tree analysis and Markov modelling, are unlikely able to represent the complex pathways transparently according to Jin. Individual-level modeling techniques, such as discrete event simulation, allow for such complex pathways to be represented more naturally as they provide more control to the modeler.
Given that clinical care is getting increasingly personalized, it does seem logical modelling methods do so as well. But that these more advanced techniques simulate individuals and provide more flexibility, does not mean they require patient-level data. Jin pointed out that discrete event simulation models, for example, can be populated based on aggregated data from published literature.
More Data, Additional Assumptions or Smart Alternatives?
To account for all downstream consequences in a piecewise comparison, or when modelling sequences, substantial evidence is required. Preferably, this evidence will come from carefully collected data, but that may not always be possible. Al highlighted several assumptions that can be made when modelling sequences of treatments: 1) treatment effects are stable over lines, 2) treatment effects may depend on the line of treatment, but not on the type(s) of previous treatment, and, 3) treatment effects decrease over treatment lines. Although such assumptions evidently need to be made in some cases, it is even more evident that they will not be representative of the true underlying effects.
But are there alternatives? James Robins, MD, Harvard T.H. Chan School of Public Health, Boston, MA, USA, shared a study in which the “no direct effect” assumption was used to optimize the cost-effectiveness of CD4 testing for HIV treatment in African countries. This approach is based on the assumption that a test has no effect on outcomes other than through the choice of treatment based on the test result. Professor Robins applied time-specific inverse probability weights to estimate the health and economic outcomes for different test intervals based on a randomized trial of a 12-week testing interval. By taking this novel approach, the need for trials for each test interval was circumvented and an interval of 48 weeks was found to be optimal.
Observational Data to the Rescue?
If there is no data from randomized studies available, observational data can inform health economic analyses. Al highlighted several considerations regarding the use and availability of such real-world datasets. There are now well-established methods for causal inference based on observational data, such as propensity-scores based methods. More specifically regarding the use of real-world data for treatment sequencing, g-methods have recently gained more attention, which are better suited to account for the time-varying aspect of sequential modelling.
Regardless of the specific methods used, data of sufficient quantity and quality is required. To illustrate that this may be challenging, Al presented information on the Castration-Resistant Prostate Cancer Registry (CAPRI). Even though 2000 patients started first-line treatment, sample size decreased substantially in subsequent lines, with only 500 patients receiving third-line treatment. Furthermore, although clinical registries often include fields on important prognostic factors, missing data of up to 50% is very common, which further reduces the amount of information available.
Ready, Set, Go!
It is evident that the concept of treatment sequencing challenges our traditional approaches towards health economic analyses. However, with an increasing amount of high-quality observational data becoming available, better understanding of causal inference based on such data, and rapidly building experience in using individual-level modelling methods, all pieces of the puzzle are coming together to appropriately represent the complex dynamics of today’s clinical pathways.
