( Editor's note: This post has been refreshed in December 2016)
Model based algorithms for Phase I dose-escalation have been in existence for nearly thirty years. Despite guarantees of increased statistical power and greater accuracy, there remains a clear preference for rule based algorithms amongst clinicians. The explanation for this is as old as the models themselves.
Bayesian models often seem, in a word, obscure. They demand complex calculations. Their results are difficult to communicate (often requiring the eerie specter of statisticians on premises). Ultimately, rule-based algorithms seem to get the job done. Why then, should anyone consider investing in model based dose-escalation techniques?
The fact is that the rules of the game have changed. The new terrain of market forces make Bayesian dose-escalation techniques nearly impossible to avoid. Here are five reasons why you should consider a model based Bayesian design for your dose-escalation study:
Thirty years ago, few IRBs and regulatory boards were familiar with the details of Bayesian models. Since then the regulatory environment has witnessed significant transformation. Innovative leaders like Novartis are currently spearheading an industry-wide shift towards the adoption of Bayesian methods, by using them for their Phase I oncology trials. Committees like the DIA Working Group on Adaptive Design provide broad education and awareness of the scope and benefits of Bayesian techniques. Even the FDA and EMA now actively encourage Bayesian model-based designs for Phase I trials in oncology. Given support from all levels of industry, we can anticipate continued efforts to establish Bayesian standards as a benchmark for regulatory committees.
(Below you can find links to the FDA's 2006 Critical Opportunities Report, and an article in the New England Journal of Medicine reporting on Novartis’ use of Bayesian methods for a recent Phase I trial of Ceritinib)
(2) Clinical Ethics:
Rule-based methods like 3+3 have the unfortunate consequence of under-dosing initial recipients of anticancer treatments and overdosing others. Under-dosing occurs because the effective use of rule-based methods requires the administration of very low doses at the outset. Tendencies for overdosing were reported by simulations conducted by James Babb, et al., and more recently by Yuan Ji and Sue-Jane Wang. Since most patients in Phase I oncology trials have advanced stage cancer, consciously under-dosing and overdosing study participants can threaten to deprive them of the best possible care.
(3) Availability of Validated Computational Packages:
Thirty years ago, there were few statistical tools to support the complex calculations required to implement model-based dose-escalation. The adoption of Bayesian methods meant long drawn-out calculations that could take weeks to perform. Now, however, software packages have begun to offer mTPI, CRM and BLRM designs (including Cytel's East 6.4 Escalate module. East 6.4 includes validated methodologies for dual agent dose escalation designs. Performing advanced calculations are now simple, swift and easily accessible to anyone in the industry.
(4) Visual Communication:
In the age of Edward Tufte, the visual display of quantitative information has enabled once arcane statistical conclusions to become communicable via thoughtfully prepared graphs, charts, figures and diagrams. Developers of computational packages have taken advantage of these advances in communication theory, to make Bayesian findings accessible to non-statisticians. For example, East 6.4 provides intuitive visuals and powerful simulations to make graphs and charts easy to interpret by non-statisticians. As a result, trial statisticians no longer need to struggle to communicate their findings to a clinical team, or to incorporate them into presentations and articles.
(5) Informed Patient Care:
Clinicians are often called upon to balance the needs of a clinical trial with responsible care for their patients. When using rule-based methods such as 3+3, critical information gained as a trial progresses cannot inform appropriate dose-recommendations for study participants. This is because rule-based models determine levels of dosages prior to the start of a study. Bayesian modeling, by contrast, has a learn-as-you-go feature. As data accumulates, clinicians are offered suggestions on how to fine-tune recommended dosages that are then administered to patients. Any clinician committed to providing the ‘best possible care’ to unhealthy patients participating in Phase I studies, has reason to transfer from rule-based to Bayesian model-based techniques.
Want to learn more about adaptive dose escalation methods? Watch our East® 6.4 Webinar: Designing Single and Dual Agent Dose Escalation Trials by clicking the link below.
Related Items of Interest
Ji, Yuan and Sue-Jane Wang. “Modiﬁed Toxicity Probability Interval Design: A Safer and More Reliable Method Than the 3+3 Design for Practical Phase I Trials.” Journal of Clinical Oncology 31.14 (2013): 1785-1791.