Session 2: Clinical Development

AI Agents in Clinical Development: Integrating Clinical Insights with Statistical Innovations

Will Ma, Founder & CEO at HopeAI

Will Ma is the founder and CEO of HopeAI, a Mayo Clinic Platform_Accelerate company on a mission to bring hope to patients through AI-accelerated clinical development. HopeAI has developed AI assistants that combine clinical insights with statistical innovations, enabling faster and more precise clinical trials. Prior to founding HopeAI, Will had over 10 years of experience in clinical development, including working as a statistician at Sanofi and BMS, and serving as a faculty member at Moffitt Cancer Center.

Abstract: While Large Language Models excel at general knowledge, clinical development demands a more nuanced approach that combines domain expertise with statistical rigor. We present a novel dual-agent system consisting of an AI Clinical Scientist that synthesizes complex clinical evidence, working in tandem with an AI Statistician that implements real-time statistical methods. This AI-augmented approach shows promise for accelerating and strengthening clinical trial design, providing clinical development teams with powerful new tools for evidence-based decision-making.

Evaluating Predictive Performance of Statistical and Machine Learning Models for Survival Outcomes with Post-Baseline Time-Dependent Covariates

Zhaohua Lu

Zhaohua Lu is an Associate Director in Biostatistics at Daiichi Sankyo and a Ph.D.-trained statistician with over ten years of experience in statistical modeling, data science, and clinical trial design and implementation. Prior to this role, he served as a Faculty Member in the Biostatistics Department at St. Jude Children’s Research Hospital for six years. His expertise spans the analysis of large neuroimaging, genomic, and natural language datasets, with a strong focus on predictive modeling, machine learning, Bayesian methods, and statistical computations and simulations. He has authored approximately 40 peer-reviewed clinical papers and 30 methodological publications, contributing significantly to both applied and methodological advancements in biostatistics.

Abstract: Time-to-event (TTE) endpoints are fundamental in drug development and biomedical research. While traditional statistical approaches, such as the Cox proportional hazards model, have long been utilized for predicting TTE outcomes, recent research has highlighted the potential of flexible machine learning (ML) techniques, including tree-based models, to achieve improved predictive performance. Furthermore, incorporating post-baseline time-varying predictors has been shown to enhance prediction accuracy, particularly in ML models. In this study, we explored the predictive performance of both the Cox model and various ML methods using a combination of baseline and post-baseline predictors. Model performance was assessed through multiple evaluation metrics, including time-dependent area under the curve (AUC), concordance index (C-index), and Brier scores. These metrics also served as criteria for selecting relevant predictors to optimize model performance. Our findings suggest that the Cox model can deliver comparable performance to ML methods in practical scenarios, particularly when the sample size is moderate and key model assumptions—such as proportional hazards and linear effects of predictors on the log hazard ratio—are reasonably satisfied. These results highlight the critical role of thoughtful model selection and comprehensive comparison in TTE outcome prediction.