Presentation (CTI)

OBJECTIVES: Clinical trials often report results using Kaplan-Meier (KM) survival curves measuring progression-free survival (PFS) and overall survival (OS). For settings with multiple health states, such as cancer, survival curve data typically includes events which happen in subsequent health states. However, since trials do not typically report whether deaths occur pre- or post-progression, available methods are inadequate to estimate state transition probabilities and hazard rates using only OS and PFS KM survival curves. Our goal is to introduce a publicly accessible method which fills this gap.
METHODS: We analyze the structure of KM survival curve data in the context of multi-state settings to probe why traditional methods are lacking. We then develop a general methodological approach to jointly estimate multi-state transition probabilities using KM survival curves. Using this general approach, we develop a flexible and simple hands-on form that can be used by the public in an open-source statistical software program.
RESULTS: We find that underlying issue is that the structure of typically available KM survival curve data, even with infinitely large samples, is consistent with a range of parameter combinations. We describe this so-called partial identification problem. We then develop and describe a general method to solve it to jointly estimate state transition probabilities using PFS and OS Kaplan-Meier survival curves. We also develop and describe a specific and flexible ready-to-use functional form, implementing it in an R package which allows users to estimate multi-state transition probabilities from KM survival curves and visualize results.
CONCLUSIONS: Estimating state-specific parameters is necessary settings such as modelling sequential treatment paths, but current methods are often restrictive or otherwise lacking. We describe the underlying problem and using this knowledge, we develop a methodological approach and software implementation allowing researchers to jointly estimate multi-state transition probabilities from Kaplan-Meier survival curve data, enabling more flexible and accurate multi-state modelling.