Learning from the literature: An AI-assisted Bayesian framework for evidence-driven oncology trial design.

Abstract

e23010 Background: Designing modern oncology trials requires synthesizing prior evidence to inform hypotheses and sample size estimation. Incomplete or imprecise literature summaries can lead to mis-specified hypotheses and underpowered studies. To address this gap, we developed LEAD-ONC (Literature to Evidence for Analytics & Design in Oncology), an AI-assisted platform that extracts trial data from published studies and performs Bayesian evidence synthesis to support evidence-driven trial design. Methods: Using expert-curated publications meeting prespecified criteria, LEAD-ONC applies large language models to extract baseline characteristics from published tables and reconstruct individual-patient data (IPD) from Kaplan–Meier curves. Trials are clustered by similarity in baseline characteristics to define comparable populations, including cohorts enrolling multiple histologies (i.e., trials enrolling both squamous and non-squamous disease). Within each cluster, Bayesian hierarchical models are fitted to reconstructed IPD to generate predictive survival distributions for new trials enrolling similar populations. Results: IPD were reconstructed from five completed phase III non–small-cell lung cancer trials evaluating chemo-immunotherapy versus intensified immunotherapy strategies involving PD-1/PD-L1 inhibitors with or without CTLA-4 inhibition. LEAD-ONC identified three population groupings: non-squamous (KEYNOTE-189), squamous (KEYNOTE-407), and multi-histology trial populations (POSEIDON, CheckMate-227, CheckMate-9LA). For a hypothetical randomized trial enrolling a multi-histology population, comparing standard chemo-immunotherapy with intensified regimens incorporating dual immune checkpoint blockade (± chemotherapy), the predicted median overall survival difference was 2.8 months (95% CI, −1.8 to 7.6), with a 0.45 probability of achieving a ≥3-month OS benefit. Conclusions: Integrating AI-based data extraction with Bayesian hierarchical modeling enables scalable learning from published trials and provides an evidence-driven framework for oncology trial design. Projected overall survival by immunotherapy strategy in multi-histology NSCLC trials. Time OS (Mono-ICI) OS (Dual-ICIs) OS Differences Mean 2.5% CI 97.5% CI Mean 2.5% CI 97.5% CI Mean 2.5% CI 97.5% CI 12 0.54 0.48 0.59 0.58 0.52 0.63 0.04 -0.04 0.12 24 0.35 0.30 0.41 0.41 0.36 0.46 0.06 -0.02 0.13 36 0.24 0.20 0.29 0.30 0.25 0.36 0.06 -0.01 0.13 48 0.17 0.13 0.22 0.23 0.19 0.28 0.06 -0.01 0.12 60 0.13 0.09 0.17 0.18 0.14 0.23 0.05 -0.01 0.12 72 0.09 0.06 0.13 0.14 0.10 0.19 0.05 -0.01 0.10