Advances in Causal Generative Medical AI featuring Dr. Hongtu Zhu

Advancing Precision Medicine Through Causal Generative AI

The National Institute of Statistical Sciences (NISS) continued its AI, Statistics & Data Science in Practice webinar series with a featured presentation by Dr. Hongtu Zhu, a distinguished professor at the University of North Carolina at Chapel Hill and former Chief Scientist of Statistics at Didi Chuxing. In his talk, “Toward Causal Generative Medical AI: A Personal Perspective,” Dr. Hongtu Zhu outlined an emerging vision for transforming precision medicine through causal, data-driven artificial intelligence systems.

The webinar focused on the development of Causal Generative Medical AI (CGMAI), a framework that integrates causal reasoning, biomedical knowledge graphs, and multimodal data analytics to address limitations in current clinical prediction models. Dr. Hongtu Zhu described CGMAI as an interconnected ecosystem of AI-powered platforms capable of synthesizing diverse data streams—including electronic health records (EHRs), medical imaging, and genomic data—to generate more accurate and individualized insights for patient care.

Building the Foundations: Methods, Frameworks, and Data Integration

A central theme of the presentation was the distinction between CGMAI and conventional generative AI. Dr. Hongtu Zhu identified several defining dimensions, including advanced reasoning architectures, explicit causal decision-making processes, and scalable deployment strategies. By incorporating causal inference directly into model design, CGMAI aims to move beyond correlation-based predictions toward models that can support decision-making in real-world clinical environments.

Dr. Hongtu Zhu emphasized three foundational pillars underpinning causal decision-making systems: causal discovery, effect estimation, and policy learning. Together, these components enable researchers to uncover underlying disease mechanisms, evaluate treatment effects, and design optimized clinical strategies. Biomedical knowledge graphs play a critical role in this framework by integrating heterogeneous data sources and encoding disease-specific causal pathways.

The presentation also highlighted the rapid evolution of biomedical data analysis. Dr. Hongtu Zhu noted a shift from relatively small-scale datasets to petabyte-scale data environments, driven by large repositories such as the UK Biobank and expanding clinical data infrastructures. This transformation has necessitated new approaches to both vertical and horizontal data integration, allowing researchers to combine information across modalities and institutions.

Applications in Healthcare and Emerging AI Technologies

In discussing practical applications, Dr. Hongtu Zhu pointed to advances in EHR data analysis, where challenges such as data heterogeneity and unstructured formats remain significant barriers. Emerging solutions include time-aware multimodal fusion models and transformer-based architectures capable of extracting insights from complex clinical data. These tools, combined with AI agents and large language models, are increasingly being used to automate scientific workflows in pharmaceutical and clinical research settings.

Audience engagement during the webinar reflected strong interest in both methodological and practical considerations. Questions addressed topics such as assumptions underlying causal inference, the role of human oversight in autonomous AI systems, and strategies for validating AI-driven clinical outcomes. Dr. Hongtu Zhu underscored the importance of rigorous validation frameworks and highlighted existing tools, including causal inference software in R and Python, that support model evaluation and data fidelity.

Workforce Evolution, Ethics, and the Future of AI in Medicine

A recurring message throughout the session was the need for the statistics and data science workforce to evolve alongside advances in AI. Dr. Hongtu Zhu emphasized that statisticians must expand their expertise to include diverse data types and computational methods, while also developing the infrastructure required to support large-scale, real-world data analysis. He noted that younger researchers often bring deep familiarity with modern AI tools, suggesting that academic programs may need to adapt curricula to better prepare students for interdisciplinary work in this area.

At the same time, Dr. Hongtu Zhu acknowledged the broader implications of rapid AI adoption. While AI has the potential to significantly accelerate data analysis and scientific discovery, it also raises questions about long-term impacts on the workforce and the importance of maintaining human oversight. He called for closer collaboration between academia, healthcare systems, and industry to establish responsible AI practices, including safeguards to ensure accuracy, reliability, and ethical use in clinical settings.

The NISS webinar series continues to serve as a platform for exploring cutting-edge developments at the intersection of statistics, data science, and artificial intelligence, bringing together researchers and practitioners to address emerging challenges and opportunities in the field.

Acknowledgements

NISS extends its sincere appreciation to our featured speaker, Dr. Hongtu Zhu, for sharing his insights and advancing understanding of causal generative approaches in medical AI. We are also grateful to Dr. Hongyuan Cao for serving as the session moderator and for guiding an engaging and thoughtful discussion. Their contributions were instrumental in making this webinar a valuable experience for attendees.