Talks on Medical Data Analysis

Program

10.30–11.10 Talk by Dr. Holger Frölich: Towards Realizing an AI-Based Probabilistic Digital Twin for Personalized Disease Management and Treatment by.

11.10 – 11.20 Break

11.20 – 12.00 Talk by Dr. John Shepherd: Radiomic Phenotypes, Adiposity, and Breast Cancer Risk: Linking Imaging Biomarkers to Tissue Biology Across Populations

Towards Realizing an AI-Based Probabilistic Digital Twin for Personalized Disease Management and Treatment

Modern AI approaches in medicine are increasingly expected to go beyond prediction and enable counterfactual reasoning and individual-level effect estimation. In this talk, I present recent advances in causal and scientific machine learning that address these challenges in complex, real-world clinical settings.

The first part focuses on counterfactual outcome modeling under confounding, censoring, and competing risks. I show how Targeted Maximum Likelihood Estimation (TMLE) can be integrated with flexible neural outcome models to enable statistically efficient and robust estimation of average treatment effects from observational healthcare data. Moreover, I introduce a novel neural architecture for individual effect estimation that jointly addresses confounding and informative censoring.

The second part introduces Multi-Modal Pharmacokinetic Scientific Machine Learning (MMPKSciML), a hybrid modeling framework that combines ordinary differential equation models with neural networks and variational inference. This approach captures inter-individual variability in drug response while preserving interpretability and physical consistency, and empirically outperforms both classical population pharmacokinetic models and classical ML methods in selected settings.

Together, these contributions point toward a unifying “digital twin” paradigm that integrates causal inference, multimodal data fusion (EHRs, wearable data, genetics), and scientific machine learning to support personalized decision-making and counterfactual simulations in medicine.

Dr. Holger Frölich

Dr. Frölich is Head of Department for AI & Biomedical Data Science at Fraunhofer SCAI, Sankt Augustin, and Full professor for Biomedical Data Science at the Medical Faculty, University of Bonn / University Hospital Bonn. He is PhD from Tübingen University, 2007 and has held various positions in academia and pharmacological industry.

Statement: I am an AI-focused computer scientist with 20+ years of experience at the intersection of data science and healthcare. My personal vision is to make AI-driven medicine a reality – hence allowing to bring the right treatments to the right patients at the right time. In my research I focus on target prioritization, precision medicine, clinical trials as application domains for AI. From a methods point of view, I am interested in multimodal data fusion, (generative) modeling of patient trajectories, hybrid AI and causal ML. I am an experienced leader of international projects and teams and long-standing teacher within the University of Bonn’s Computer Science and Life Science Informatics Master programs.

Radiomic Phenotypes, Adiposity, and Breast Cancer Risk: Linking Imaging Biomarkers to Tissue Biology Across Populations

Medical imaging has long played a central role in cancer detection, yet its potential as a predictive and mechanistic data source remains underexploited. In breast cancer, imaging biomarkers such as breast density are among the strongest known risk factors, but traditional clinical risk models based on these features perform inconsistently across populations and fail to capture much of the latent information embedded in images.

This talk will describe how large-scale mammography data, combined with modern machine learning and representation learning, are reshaping our understanding of cancer risk. I will discuss emerging evidence that radiomic texture, spatial organization of tissue, and AI-derived risk representations outperform conventional clinical models, while remaining only weakly correlated with established imaging biomarkers. This surprising decoupling suggests that hybrid models, combining physics-informed imaging features with data-driven AI representations, may offer substantial gains in accuracy and generalizability.

The lecture will conclude with a set of open questions and grand challenges for the computer science community, including robustness across heterogeneous populations, interpretability of learned risk representations, and the linkage between image-derived features and underlying biology. These challenges frame medical imaging as a rich, high-dimensional learning problem where advances in computer science can directly translate into earlier detection, improved risk stratification, and more equitable cancer prevention.

Dr. John Shepherd

Dr. Shepherd is the Chief Scientific Officer of the University of Hawaii Cancer Center and the B.H. and Alice C. Beams Endowed Professor in Cancer Research at the John A Burns School of Medicine. He is also a Fulbright Scholar, a Fellow of the American Institute for Medical and Biological Engineering (AIMBE), a fellow of the American Association of Physicists in Medicine (AAPM), and a former President of the International Society for Clinical Densitometry. He received his PhD in Engineering Physics from the University of Virginia and then completed a Postdoctoral Fellowship in Biophysics at Princeton University. From there he went on to develop body composition and bone density algorithms for a major women’s health company and his patents are the basis for the accuracy of that company’s body composition algorithms. He then joined the Radiology Department of the University of California San Francisco and led his research group for 19 years studying a wide variety of imaging biomarkers for breast cancer, obesity, and osteoporosis using advanced machine-learning techniques. For the past 8 years, he has been with the University of Hawaii Cancer Center where is the Director of the Hawaii Pacific Islands Mammography Registry which is currently monitoring over 125k women for developing breast cancer risk models in this population with high disparity and underrepresentation. He also co-host the biennial International Breast Density and Risk Assessment Workshops with the next workshop to be held on the Island of Maui in June, 2027. He has been continuously funded by the NIH since 2005, led 6 R01-funded studies, and published over 400 peer-reviewed publications that have been referenced over 30,000 times. Lastly, he is an avid surfer, cyclist, and island ridge hiker!