SYNTHETIC PATIENT TRAJECTORIES FROM A GENERATIVE MACHINE-LEARNING MODEL: A FRAMEWORK FOR GENERATING POPULATION-ADJUSTED SURVIVAL OUTCOMES IN CHRONIC LYMPHOCYTIC LEUKAEMIA

Author(s)

Talha Munir, PhD1, Paolo Sportoletti, MD2, Leyla Mohseninejad, MSc, PhD3, Stephen Opat, MBBS4, Igor Chebuniaev, MSc5, Samuel Aballea, MSc, PhD6, Rami Zouari, MSc7, Noemi Kreif, PhD8, Mondher Toumi, MSc, PhD, MD9, Marcus Lefebure, PhD10, Rhys Williams, PhD11, Sheng Xu, PhD12, Anna Maria Frustaci, MD13, Loic Ysebaert, PhD14, Mazyar Shadman, MD15.
1Leeds Teaching Hospitals NHS Trust, Leeds, United Kingdom, 2University of Perugia and Santa Maria della Misericordia Hospital, Perugia, Italy, 3BeiGene, The Hague, Netherlands, 4Monash Health, Melbourne, Australia, 5InovIntell, Tbilisi, Georgia, 6InovIntell, Krakow, Poland, 7InvoIntell, Krakow, Poland, 8University of Washington, Seattle, WA, USA, 9Aix-Marseille University, Marseille, France, 10BeOne Medicines Ltd, London, United Kingdom, 11BeOne Medicines Ltd, San Carlos, CA, USA, 12BeOne Medicines, Shanghai, China, 13ASST Grande Ospedale Metropolitano Niguarda, Milan, Italy, 14IUCT Oncopole, Toulouse, France, 15Fred Hutchinson Cancer Center, Seattle, WA, USA.
OBJECTIVES: Synthetic data generation (SDG) is proposed to support unanchored indirect treatment comparisons (ITCs); prior work with continuous outcomes showed SDG yielding more precise treatment-effect estimates than matching-adjusted indirect comparison (MAIC). This approach was extended to time-to-event outcomes in chronic lymphocytic leukaemia, using patient-level data from SEQUOIA (NCT03336333) and ALPINE (NCT03734016) zanubrutinib trials.
METHODS: A latent ordinary differential equation (ODE) model generated full individual patient trajectories (lab measurements over time and timed clinical events: progression and death) conditional on baseline profiles, versus relying solely on observed event times. Investigator-assessed progression-free survival (PFS) was primary outcome of interest. The model was fine-tuned to reproduce PFS and overall survival (OS) in pre-specified subgroups defined by del(17p)/TP53 and IGHV mutational status (SEQUOIA) and number of prior therapies (ALPINE). To assess generalisability, we generated PFS and OS for distinct, held-out subgroups (age <65/≥65 years, β2-microglobulin normal/elevated, bulky disease) using baseline profiles created by an independent SDG method. Bootstrapping was used to quantify uncertainty around estimates based on synthetic data.
RESULTS: Across all held-out test subgroups in both trials, generated PFS and OS curves closely reproduced the original Kaplan-Meier curves with larger deviations confined to smaller subgroups and the tail of survival curves, where data are sparse (mean absolute difference in PFS and OS probabilities vs. ground truth: 0.02). On average, 90% of width of each confidence interval around synthetic-data PFS log-hazard ratios between patient subgroups fell within bounds of corresponding ground-truth confidence interval for fine-tuning subgroups, versus 86% for held-out subgroups.
CONCLUSIONS: A latent ODE model reproduced survival curves in unseen patient strata from independently generated baseline profiles, demonstrating ability to generate valid covariate-conditional time-to-event outcomes. This framework could support population-adjusted indirect treatment comparisons or external control arms and could extend to prospective outcome predictions, virtual cohort generation, treatment modelling sequences and evaluation of hypothetical clinical scenarios.

Conference/Value in Health Info

2026-11, ISPOR Europe 2026, Vienna, Austria

Value in Health, Volume 29, Issue 12S

Code

MSR91

Topic

Methodological & Statistical Research

Topic Subcategory

Artificial Intelligence, Machine Learning, Predictive Analytics

Disease

No Additional Disease & Conditions/Specialized Treatment Areas, Oncology

Your browser is out-of-date

ISPOR recommends that you update your browser for more security, speed and the best experience on ispor.org. Update my browser now

×