Do Treatment Sequencing / Pathways Models Have a Place in Health Technology Assessment?

That is the title of an ISPOR Europe panel I saw with Hugo Pedder, Jeroen Jansen, Dawn Lee and Mark Harries. The goal of sequential pathways model is to be able to answers questions like (i) how does a new treatment impact the cost and effectiveness for patients within a sequence of treatment, or (ii) which sequence provides the best value in terms of health benefits and treatment cost.

Dr. Jansen talked about two open-source sequential models in rheumatoid arthritis and EGFR+ non-small-cell lung cancer that were produced for the Innovation and Value Initiative (IVI), now known as Center for Innovation & Value Research.  I actually participated in the create of both the IVI-RA and IVI-NSCLC models. Most models mean look at parameter uncertainty through deterministic sensitivity analysis (DSA) or probabilistic sensitivity analysis (PSA); however, one key learning from the IVI models was that model structure for sequential models lead to vastly different estimates of treatment value.  A second learning was that the model structure should mirror the evidence synthesis structure. Third, if difference in treatment history are prognostic factors or effect modifiers, the analysis will be biased. In other words, if a treatment’s effectiveness during later lines of therapy depends the treatments received earlier in care, then it is important to model those impacts explicitly in a treatment sequence model or the results will be biased.

Dr. Lee and Dr. Marries examined renal cell carcinoma (RCC) as part of a NICE pathways pilot. One key learning is that these models are complex: researchers examined 4 lines of therapy for 3 different risk populations; 744 treatment sequences and 90 scenario analyses were conducted.  Second, sequential models take much more time than standard models and typically require more computing power. Not only is Excel not feasible to use, but the authors relied on a university supercomuter to run some sensitivity analyses.  Third, earlier treatments impact treatment options in later lines; however, one key simplifying assumption is independence of effect (i.e., treatment efficacy is independent previous treatment).  Fourth, additionally, sequencing models require much more data as there are many more parameters to estimate.  Access to data is key: both real-world data as well as more detailed results from clinical trials. Fifth, one must clearly specify the decision question.  As described above, one decision question would answer whether adding a treatment to current standard of care treatment sequence is cost effective; another is identifying which treatment sequence is most cost effective.  Sixth, the model results in very different results depending on whether a state transition or partitioned survival model was used, mainly due to differences in the estimated health gains.  Unsurprisingly, the presenters noted that there is less variability in estimated value when the treatment pathways are widely known as compared to when they are highly uncertain.  The full pathways pilots sequential model was published by Lee et al. (2024).

Both presentations noted that treatment sequence models are largely not feasible in Excel and are require to do in R. R is a much more powerful tool than Excel, but stakeholders were a bit more skeptical of the results compared to Excel-based models. Dr. Jansen noted that the hesim package in R can help develop these sequential models.