Chair: Szu-Yu Tang, PhD (Pfizer)
Co-Chair: Shibing Deng, PhD (Pfizer)
Abstract: This session explores the transformative potential of Multistate Models (MSMs) in clinical drug development, offering a robust alternative to traditional time-to-event analyses. While standard metrics like the hazard ratio often oversimplify complex disease trajectories, MSMs provide a flexible framework for integrating longitudinal biomarkers, intermediate clinical events, and competing risks to support more robust clinical inferences.
Presentations will range from foundational applications of MSMs to cutting-edge methodological innovations, including the use of Neural Ordinary Differential Equations for modeling stochastic disease progression without restrictive parametric assumptions. We will also examine the critical role of joint modeling in linking time-dependent biomarkers to survival outcomes. Finally, the session addresses high-stakes regulatory challenges, proposing Bayesian MSM frameworks to validate surrogate endpoints and support FDA Accelerated Approvals.
Speaker: Fang-Shu Ou, PhD (Mayo Clinic)
Title: Using Multistate Models with Clinical Trial Data for a Deeper Understanding of Complex Disease Processes
Abstract: The hazard ratio is commonly used to measure treatment effects in clinical trials with time-to-event endpoints; however, it lacks a straightforward causal interpretation. In contrast, absolute risks offer a more intuitive interpretation and may be better suited for causal inference. A multistate model describes longitudinal events, enabling multiple clinical endpoints to be analyzed as outcomes while estimating covariates simultaneously. It also calculates absolute risks, such as the probability of being in a particular state at a given time, the expected number of transitions to a state, and the expected duration spent in each state. In this presentation, we will provide a brief introduction to multistate models and demonstrate their application using clinical trial data. We will showcase how multistate models can enhance our understanding of complex disease processes through the analysis of clinical trial data.
Speaker: Elsa Vazquez-Arreola, PhD (National Institute of Diabetes and Digestive and Kidney Diseases)
Title: Investigating Time-Dependent Associations of Kidney Disease Biomarkers with Mortality in People with Type 2 Diabetes Using Joint Models for Longitudinal and Multistate Processes
Abstract: High albumin to creatinine ratio (ACR) and low estimated glomerular filtration rate (eGFR) are biomarkers of kidney disease that are associated with increased risk of cardiovascular disease (CVD) and all-cause mortality in persons with type 2 diabetes (T2D). The Look AHEAD randomized clinical trial in adults with T2D measured ACR and eGFR longitudinally and followed subjects for time to fatal and non-fatal CVD events and to all-cause mortality. We use joint models for longitudinal and multistate processes to study time-dependent associations of ACR and eGFR with all-cause mortality in Look AHEAD participants. We explore the time-dependent associations of ACR and eGFR with the transitions from baseline state to first non-fatal CVD event, from baseline state to all-cause mortality and from first non-fatal CVD event to all-cause mortality. The first non-fatal CVD event during the trial is an intermediate event between the baseline state and mortality because its occurrence may affect values of ACR and/or eGFR. In the joint model, the longitudinal trajectories of ACR and eGFR are examined using a multivariate linear mixed effects model while the process of all cause mortality is described by an illness-to-death multistate model.
Speaker: Stefan Groba, PhD (Dana-Farber Cancer Institute, Harvard Medical School)
Title: Neural ODEs for Multi-State Survival Analysis
Abstract: Standard multi-state survival models often rely on restrictive parametric assumptions that fail to capture complex, idiosyncratic disease trajectories. In this talk, I present a framework that models disease progression directly as a stochastic continuous-time Markov jump process. We parameterize the infinitesimal generator matrix using Neural Ordinary Differential Equations (Neural ODEs), allowing us to estimate transition probabilities by directly solving the Kolmogorov forward equations. This approach naturally handles irregular sampling and non-linear dynamics without imposing strict proportionality. We demonstrate that this method achieves state-of-the-art performance on benchmark multi-state datasets.
Building on this core framework, I will introduce a variational latent variable formulation to quantify uncertainty in individual cause-specific hazard rates. Finally, I will present a joint modeling framework that integrates Pharmacokinetic / Pharmacodynamic (PK/PD) dynamics to capture the bidirectional dependency between drug kinetics and disease progression.
Speaker: Brian Hobbs, PhD (Telperian)
Title: Bayesian Multi-State translation Model of Response, Progression and Death Elucidates Mechanisms of Survival Benefits
Abstract: Advances in cancer biology have propelled a new generation of therapeutic innovations, including immune checkpoint inhibitors, bispecific molecules, adoptive cellular therapies, and gene-based interventions. These treatments, often administered in combination with traditional modalities such as chemotherapy or radiation, target key biological mechanisms in tumorigenesis and cancer progression. Regulatory approval of emerging cancer therapies traditionally requires evidence from large, randomized controlled trials (RCTs) that demonstrate improvement in overall survival (OS) compared to standard-of-care therapies. These confirmatory trials, while statistically robust, are time-consuming and resource-intensive, frequently spanning five to ten years due to extended enrollment and follow-up requirements. Prolonged confirmatory trials provide conclusive evidence but also deny or delay access to potentially life-saving therapies to patients. The FDA developed expedited regulatory pathways based on surrogate endpoints like objective response rate and progression-free survival. While this mechanism has enabled earlier market entry, its reliance on surrogate endpoints alone has introduced challenges in verifying clinical benefit. Notably, 23% of oncology indications granted accelerated approval (AA) between 2009 and 2022 were later withdrawn after failing confirmatory trials. In response, the FDA’s 2024 updated guidance mandates that confirmatory trials be initiated—and in some cases fully enrolled—before accelerated approval is granted. The mechanism, however, still relies on putative surrogate outcomes. This article presents a nonparametric Bayesian multi-state model for survival analysis that facilitates probabilistic inference of any collection of hypothesized mechanisms of action through intermediate outcomes such as tumor response, immunosuppression, and disease progression. Aligning with the objectives of FDA OCE’s Project Confirm, the method can be applied at interim analyses of RCTs with partially mature survival data to better synthesize the observed statistical evidence for AA using multiple outcomes. The methodology promises to improve regulatory and clinical decisions for stakeholders.
