PURPOSE: It is common for the long term overall survival (OS) benefit to represent a major source of uncertainty due to the immaturity of data from clinical trials. The emergence of immuno-oncology (IO) agents and their new mechanism of action add further layers of complexity in OS extrapolation. For typical OS extrapolation relying on fitting parametric curves under a set of standard distribution assumptions, the main challenges for IO agents are the large variation in the extrapolations, and how to decide the best clinically plausible curve. In recent reimbursement submissions, rich patient-level biomedical data collected in IO trials were not typically used to help with reducing the uncertainty of OS extrapolation, presumably due to their complex nature. The objective of this workshop is to discuss the scientific rationale for looking beyond Kaplan-Meier curves in OS extrapolation and harnessing patient-level biomedical data, such as repeated longitudinal tumor measurement data, to help inform the best clinically plausible scenario and thus reduce uncertainty in OS extrapolation. The workshop will build upon previous workshops held at the 2016 and 2017 ISPOR meetings and provide practical examples of how patient-level data could be used to improve the precision and validity in OS extrapolation. DESCRIPTION: The workshop will begin with a description of biomedical measures in oncology and their potential in informing OS extrapolation. The high degree of uncertainty introduced by standard parametric curve fitting will be reiterated with an overview of recent reimbursement submissions. Several emerging approaches for OS extrapolation employing patient-level biomedical information will be presented, including examples from IO trials. The strengths and weaknesses of using biomedical measures to support OS projections will also be highlighted by the panelists, together with suggestions for future research. Attendees will be invited to contribute their own experiences of using patient-level biomedical data for supporting OS assumptions.