Prediction, Inference and Overparameterization in Machine Learning: Guiding Intuition for Clinical AI
Jeffrey D. Blume, PhD
November 12, 2025
Machine learning models are the engines of AI and are increasingly applied in clinical practice, yet their goals and behavior often differ from traditional statistical models. This talk will develop intuition for how prediction, inference, and overparameterization interact; using simple examples to reveal when complex models help, and when they can mislead. We’ll discuss practical ways to evaluate and trust model outputs in clinical settings where interpretability and generalization matter most.
Learning objectives:
- Differentiate prediction from inference
- Recognize the implications of overparameterization in modern machine learning
- Integrate statistical reasoning into clinical evaluation of AI
To view previous Grand Rounds lectures, click here.
About the Lecturer — Jeffrey D. Blume, PhD
Dr. Jeffrey D. Blume is Professor of Data Science and Quantitative Foundation Associate Dean for Academics and Faculty Affairs at the University of Virginia School of Data Science. A biostatistician by training, he is a leading expert in likelihood‐based methods for measuring statistical evidence and the inventor of the second-generation p-value, a framework for improving statistical inference and reproducibility in data-rich settings. His research spans prediction modeling – including methods for handling missing data, high-dimensional model selection and post-selection inference – mediation modeling, ROC analysis, and clinical trial design, particularly in diagnostic and imaging studies. Dr. Blume has served as lead statistician on numerous large-scale multisite trials and currently co-leads an NCI-funded project examining disparities in lung-cancer screening, for which he and his collaborators received Vanderbilt’s Chancellor’s Award for Research. He also served for more than a decade on the faculty of the Radiological Society of North America’s (RSNA’s) Clinical Trials Methodology Workshop (CTMW), including five years as its co-leader. His collaborative experience spans radiology, nephrology, women’s health, and neuroscience, including service on multiple data safety monitoring boards. As Director of the PRISM Lab (“Prediction, Inference, and Scale in Data Science Models”), Dr. Blume mentors PhD students developing novel tools to bridge methodological innovation with applied biomedical research. A Fellow of the American Statistical Association and the American Association for the Advancement of Science, he was also awarded the Spinoza Chair in Medicine by the University of Amsterdam in 2019.
Comments