LB-004 - COUPLING BIOMARKER KNOWLEDGE WITH THE STRENGTHS OF PHARMACOMETRICS AND MACHINE LEARNING FOR INDIVIDUAL PROGNOSIS IN ONCOLOGY: A PRECISION MEDICINE FRAMEWORK FOR HEALTH CARE PROVIDERS

C. JAMOIS¹, A. FOCHESATO², F. MERCIER³, K. DIAZ-ORTAZ⁴; ¹Pharma Research and Early Development, Roche Innovation Center Basel, PHARMA RESEARCH AND EARLY DEVELOPMENT, ROCHE INNOVATION CENTER BASEL, SWITZERLAND, ²Pharma Research and Early Development, Roche Innovation Center Basel, Basel, Switzerland, ³Genentech Research and Early Development, San Francisco, Genentech Pharma Research and Early Development, San Francisco, United States of America, ⁴University College London, United Kingdom, University College London, United Kingdom.

Background: Although immune-based combination trials show promise in oncology, improving patient benefit-risk remains crucial for healthcare providers (HCPs). Anticipating primary endpoints from early pharmacodynamic (PD) biomarker readouts can inform patient management and reduce burden. Here, we couple pharmacometrics (PMx) and machine learning (ML) to develop a survival decision-making algorithm for patients with Non-Small Cell Lung Cancer (NSCLC) treated with Atezolizumab (ATZ), offering clinical teams a digital tool for therapy matching and patient stratification.

Methods: Clinical data from BIRCH (n = 595), FIR (n = 133), and POPLAR (n = 134) ATZ Phase 2 arms were used for development; OAK (n = 553) Phase 3 data for validation; ATZ + Carboplatin + (nab-) Paclitaxel (n = 377/319 + (325)) and ATZ + Carboplatin + Paclitaxel + Bevacizumab (n = 368) combination arms for applications.
6-/12-/24-week PD biomarker kinetics (tumor size, albumin, lactate dehydrogenase, and neutrophils) were modeled using non-linear mixed effects and pooled to pre-treatment information for building landmark-based survival random forests. As new patients are recruited, the model evaluates whether individual data are informative for survival outcome prediction, or if more on-treatment data are needed to meet the desired confidence threshold.

Results: Developed on monotherapy data, our tool showed good generalizability to combination trials, providing most patients with a trial continuation/discontinuation suggestion already at 6 weeks with enough confidence. The tool requires less than 10 clinical information to provide 1-year predictions, with accuracy levels diagnosed via both calibration and discriminative metrics. Patient-centric risk factors for survival outcome were derived to ease clinical engagement on patient stratification.

Conclusion: Shaped as an online patient evaluation, monitoring, and re-evaluation loop, our tool can digitally assist oncologists in proposing informed next therapeutic steps.