What is ‘Bayesian Hierarchical Modelling’ and how can it be used to evaluate oncology treatments studied in basket trials?

Should payers cover a new oncology treatment targeting specific biomarkers across multiple tumor types? One the one hand, one could require a separate trial for each tumor type. While this would be convincing evidence, it also is very expensive to conduct clinical trials for every tumor type, particularly if treatment efficacy is homogenous across tumor types with the presence of the biomarker. Conversely, simply accepting that a treatment targeting an oncology biomarker would work across all tumor types would be too far an assumption when there is heterogeneity in treatment response across tumor types.

How can one balance these twin goals of minimizing R&D cost, but maximizing certainty in treatment efficacy? One way to do this is with a basket trial. A basket trial is a type of clinical trial design that evaluates the efficacy and safety of a targeted therapy across multiple cancer types or subtypes that share a common molecular alteration or biomarker.

A paper by Sugden et al. (2024) argues that the use of Bayesian hierarchical modelling (BHM) could be useful for addressing this question. They state that:

…[BHM] is considered particularly suited to the assumption that inter-tumour site efficacy is similar within basket trials, representing a middle ground between assuming complete homogeneity (i.e. pooling all tumour sites) and complete heterogeneity (i.e. independent modelling of tumour sites).

How does BHM work conceptually and when does if fail?

Bayesian hierarchical modelling allows for the borrowing of information regarding treatment effects across histological subtypes, which is particularly useful in the context of small sample sizes in individual histological subtypes. As such, BHMs provide a foundation to allow for the treatment effect in a given histology to be informed by all histologies, increasing the utilisation of the available data. However, sufficient homogeneity is still required: “… the BHM is
advantageous only if it is considered reasonable to allow such borrowing.” [See Murphy et al. 2021]

The pros and cons of applying this approach to basket trials where sample sizes of individual trials are relatively small is described below:

Through applying BHM, tumour site-specific survival estimates are pulled towards an overall average (dependent on the permitted level of borrowing), potentially biasing survival estimates in individual tumour sites. However, in the presence of small sample sizes, complete independent modelling of individual tumour sites is likely to lead to imprecise estimates.

The authors examine a immuno-oncology HTA submission to NICE that examined treatment of an oncology biomarker across a variety of tumor types (e.g., colorectal cancer, endometrial cancer, gastric, small-intestine or biliary cancer). The BHM approach often works best when:

Homogeneity in baseline survival. It is helpful if baseline survival is relatively similar across the tumors being studied. While this isn’t completely necessary, it is more likely the treatment efficacy is more homogeneous if baseline survival is similar. With very different expected survival, it is impossible for both relative and absolute survival improvements to be similar across tumor types. Comparators. The clinical trial ideally should includes comparator arms, ideally separately by tumor type. As with any clinical trial, the comparator helps to better measure treatment efficacy. This may be problematic, however, if the population is a rare biomarker or if it is not ethical to provide patients with standard of care if no other treatments are available. RWE with biomarker data needed for single arm trials. One approach to address a lack of comparator for single arm trials is to use matching adjusted indirect comparisons (MAICs). The MAICs weight the RWE to match the characteristics of patients in the trial to serve as a counterfactual for the trial population. However, this approach only works if biomarker data is available in the real-world data. Stratified adverse event reporting. It is possible that adverse events are similar across tumor types but they could also vary by tumor types. Adverse events rates should be reported in aggregate and by tumor types.

In summary, the authors conclude that more research is needed as the NICE technology appraisal was a bit concerned that this was the first BHM application they had received (however, doesn’t someone have to be the first!). In short, the key takeaway they authors had was:

Bayesian hierarchical modelling is a useful approach in the context of histology-independent basket trials, although only under the assumption that each histological subtype can be
justifiably considered to be subgroups of an overarching population.

The article is only 3 pages and worth a read.