Invited Session 6: From Data to Evidence - Statistical Learning and AI Methods for Oncology Trials

10:10-11:30 AM, November 23, 2026; Room 7C-D

Moderator
Speakers

GRACIAS: Gradient-Augmented Calibration Incorporating AI for Covariate Adjustment in Randomized Studies

Speaker: Ruitao Lin, PhD, Associate Professor, The University of Texas MD Anderson Cancer Center
Abstract

Pretrained AI models can encode rich prognostic information, but predictive accuracy alone does not guarantee improved efficiency for treatment-effect estimation in randomized trials. We develop a general calibration framework for augmented inverse-probability weighted (AIPW) estimation in multi-arm trials and characterize conditions for simultaneous variance optimality across treatment contrasts. Under stated regularity and arm-specificity assumptions, requiring this optimality uniformly over admissible data-generating distributions implies that calibration is affine in its parameters, up to reparameterization. Motivated by this characterization, we propose GRACIAS, which combines AI-generated outcome predictions under all treatment arms with low-rank features obtained from model-parameter gradients. Separate Oaxaca–Blinder least-squares regressions within each treatment arm calibrate these features on a common basis, and the resulting outcome models enter the AIPW estimator. Under simple randomization, this construction simultaneously minimizes asymptotic variance within the chosen linear calibration space for every zero-sum treatment contrast. We further establish a local approximation result: under smoothness conditions, when the true outcome regression is represented by nearby model parameters and the relevant gradient directions are retained, the augmented calibration space contains an approximation that removes the leading first-order error. We extend the calibration framework and its optimality conditions to covariate-adaptive randomization. Simulations and analyses of 21 randomized trials demonstrate precision gains from calibration relative to directly using AI predictions, with additional improvements from gradient- and representation-augmented implementations. These results connect flexible AI prediction with classical regression adjustment and provide a principled framework for using AI-derived information to improve treatment-effect estimation.


Evidence Synthesis of Drug Safety from Oncology Trials

Speaker: Shouhao Zhou, PhD, Associate Professor of Biostatistics, College of Medicine, Pennsylvania State University
Biography

Shouhao Zhou is an Associate Professor of Biostatistics at Penn State College of Medicine, where he also directs the Center for Methodology Excellence and Innovation and the Biostatistics PhD Program. He received his PhD in Statistics from Columbia University and previously served on the faculty at The University of Texas MD Anderson Cancer Center. His research focuses on Bayesian statistical methods, evidence synthesis, pharmacovigilance, and innovative cancer trial designs.

Abstract

Meta-analysis is essential for synthesizing clinical trial evidence, but safety data pose challenges that differ fundamentally from efficacy outcomes. Adverse events (AEs) are multidimensional, graded by ordered severity levels, often rare, and frequently incompletely reported, making conventional meta-analysis methods inadequate for comparative safety evaluation. In this talk, we introduce Meta-analysis of Adverse druG Effects with Censoring (MAGEC), a Bayesian learning framework for integrating complex AE data from oncology trials. MAGEC jointly models multiple AEs and ordinal severity grades while accounting for left-censoring due to incomplete reporting, within-study correlations across AE categories, and sparsity-inducing methods to identify AE-specific treatment differences. An application to targeted cancer therapy illustrates its practical utility for comparative oncology drug safety evaluation.


AI Augmented Clinical Trials

Speaker: Arsela Prelaj, ESMO / IRCCS INT Milan
Abstract

Oncology clinical trials often face significant challenges, including slow patient recruitment, high failure rates, and limited applicability of results to broader patient populations. Artificial intelligence (AI) offers promising opportunities to address these issues by improving trial design, accelerating patient recruitment, optimizing eligibility screening, and enhancing data collection and patient monitoring. By analyzing large amounts of clinical data, AI can help identify suitable patients, improve trial efficiency, and support more informed decision-making throughout the clinical trial process. While some AI applications are already being implemented in cancer centers, more advanced approaches, such as digital twins and synthetic control groups, still require further research and validation. However, challenges related to data quality, patient privacy, fairness, and regulatory oversight must also be considered. Overall, AI has the potential to make oncology clinical trials faster, more efficient, and more inclusive, while maintaining the importance of clinical validation, human oversight, and patient safety.


Learning from Literature: Integrating Large Language Models and Bayesian Evidence Synthesis for Oncology Trial Design

Speaker: Wei Wei, PhD, Associate Professor, Division of Computational Oncology and Biostatistics, Yale School of Medicine
Abstract

Published clinical trials provide valuable information for designing future oncology studies, but extracting information from unstructured multimodal metadata and synthesizing evidence across heterogeneous trial populations remain major challenges. In this talk, we introduce an AI-assisted evidence synthesis framework that integrates large language models (LLMs) with Bayesian hierarchical modeling to generate target-trial-aligned historical benchmarks. Building on LEAD-ONC (Literature to Evidence for Analytics & Design in Oncology), the proposed framework uses LLMs to extract trial characteristics and baseline covariate summaries from published reports, together with reconstructed individual patient survival data from Kaplan–Meier curves. Expert-curated eligibility criteria and a Bayesian compatibility model are used to identify historical trials most relevant to a specified target population. Reconstructed individual survival data from compatible trials is then synthesized using a Bayesian hierarchical model to generate posterior predictive survival distributions that account for residual between-trial heterogeneity. We evaluate the framework using eleven randomized phase III trials in advanced gastrointestinal cancer, demonstrating its ability to generate target-relevant survival benchmarks with improved predictive coverage compared with conventional pooling. This approach bridges automated literature extraction and Bayesian evidence synthesis to support more informed oncology trial design.


Additional speaker biographies, presentation titles, and abstracts will be posted when received.