A multistate model for early decision-making in oncology.

Area of Science:

• Clinical Pharmacology
• Biostatistics
• Oncology Drug Development

Background:

• Oncology drug development relies on phase-based decisions, often using Response Evaluation Criteria in Solid Tumors (RECIST) from single-arm trials.
• RECIST implicitly assumes tumor response surrogacy for long-term outcomes like progression-free survival (PFS) and overall survival (OS).
• New therapies may have weaker or inaccessible links between RECIST response and long-term endpoints, challenging traditional decision-making.

Purpose of the Study:

• To explore a multistate model for enhanced decision-making in early-phase single-arm oncology trials.
• To utilize richer patient data beyond RECIST status, including time to response, duration of response, PFS, and time to death.
• To predict overall survival (OS) hazard ratios (HR) for experimental treatments using early phase data with limited survival follow-up.

Main Methods:

• Development and application of a multistate model to analyze early phase oncology trial data.
• Prediction of OS HR by comparing historical controls to experimental treatment data derived from the multistate model.
• Validation using two case studies to demonstrate the feasibility of OS HR estimation.

Main Results:

• The multistate model successfully incorporates detailed temporal patient data, offering more information than standard RECIST criteria.
• Feasibility of estimating OS HR was demonstrated through case studies, showing the model's practical application.
• The model provides acceptable OS predictions even with limited follow-up and small sample sizes.

Conclusions:

• The proposed multistate model offers a more robust approach for early efficacy decisions in oncology drug development.
• This method enhances decision-making by providing reliable OS predictions, particularly for novel therapies with uncertain RECIST-outcome links.
• The approach supports better-informed decisions early in the drug development pipeline, even with limited patient data.