See the Campus Event Calendar for details about upcoming seminars
Day and Time: Fridays from noon to 1 pm central time unless otherwise noted.
Location: Morgridge Hall Seminar Room – 7560 unless otherwise noted.
Zoom: https://uwmadison.zoom.us/j/96633729112
Passcode: 621125
Further details should be available about a week before the seminar.
Upcoming Seminars:
October 23
Speaker: Zhenhua Lin, National University of Singapore:
LOCATION CHANGE: 1524 Morgridge Hall
Poster: 1023-Zhenhua-Lin-poster
Title: Regression with Complex Data Objects: Geometry and Biomedical Applications.
Abstract: Modern biomedical studies increasingly generate complex data objects such as covariance and connectivity matrices, probability distributions, and other observations that naturally reside in nonlinear metric spaces, where standard regression methods based on vector operations are no longer directly applicable. In this talk, I will consider regression from two complementary perspectives. First, when the response is a metric-space-valued object indexed by a scalar predictor such as time, I will introduce total-variation-regularized Fréchet regression, which provides an intrinsic way to recover piecewise-constant structure while respecting the geometry of the response space; an application to task-related fMRI data illustrates its use for detecting changes in dynamic functional connectivity. I will then consider the reverse setting, where the response is binary but the covariate is a metric-space-valued object. Motivated by classical logistic regression, we construct a geometrically defined binary regression model using distances and angles between geodesics, and apply it to functional-connectivity matrices from the Human Connectome Project to distinguish motor and language-processing tasks. Together, these results illustrate how familiar regression ideas can be extended beyond Euclidean spaces while retaining interpretability and relevance to complex biomedical data.
October 30: Guanhua Chen, Biostatistics & Medical Informatics, UW-Madison
November 6: Zhengzheng Tang, Biostatistics & Medical Informatics, UW-Madison
DeMets Lectures
Poster: DeMets_LecturePoster2026
November 12
DeMets Lecture 1
Alternate Location and Time: 3:45-4:45 PM. Room 1306 Health Sciences Learning Center. Reception to follow.
Speaker: Thomas Louis, Johns Hopkins Bloomberg School of Public Health
Title: With Reference to What?
Abstract: Goals and criteria, either explicit or implicit, guide much of statistical practice. In this context, I discuss statistical considerations associated with inference to a reference population, and those guided by a goal or a criterion. Reference population examples come from clinical trials, epidemiology and surveys, with “representation” an overarching issue. Goal/criterion guided examples are drawn from summarizing laboratory assays and from hospital profiling. I close with comments on a related administrative and political issue: the what and the how.
November 13
DeMets Lecture 2
Morgridge Hall Seminar Room 7560
Speaker: Thomas Louis, Johns Hopkins Bloomberg School of Public Health
Title: Improved Small Domain Estimation via Compromise Regression Weights
Abstract: Shrinkage estimation of small domain parameters typically combines a noisy direct estimate with a more stable regression estimate. However, if the regression model is misspecified, the “optimal” (MLE) estimate of regression parameters can degrade performance for the relatively unstable domains due to substantial shrinkage toward the misspecified regression surface. Jiang and co-authors addressed this issue via the Observed Best Predictor (OBP) that, in estimating the regression, over-weights the small domains, thereby giving more weight to those that, “care about” the regression. Going a step further, we introduced a class of empirically-informed regression weights that are a convex combination of the MLE and OBP values–a Compromise Best Predictor (CBP). The mixing parameter is found by minimizing an unbiased estimate of the mean-squared prediction error, a SURE estimate. This approach enables the CBP to preserve the robustness of the OBP while retaining the main advantages of the MLE when the regression model is well-specified. We compute performance of the MLE, OBP and CBP approaches via simulation, and compare data analysis results in estimating gait speed in older adults.
November 27: NO SEMINAR
December 4: Student Rotation Presentations
Past Seminars:
September 11
Speaker: Andreas Velten, UW-Madison Dept of Biostatistics & Medical Informatics
CHANGE OF LOCATION: Morgridge Hall Room 1524
Title: Of Photons and Pixels: How the transition to single photon pixels changes imaging
Poster: 0911-Andreas-Velten-poster
Abstract: Cameras and microscopes make measurements of photons to form images or extract other information from a scene. We now have access to various types of single photon or quanta cameras that are able to directly convert individual photons to digital information. This results in fundamental changes in what cameras can do, and how they should be designed.
This presentation shows examples from our recent research to illustrate how the difference between classical and quantum light substantially changes how we might approach designing a camera and what cameras can do. Our examples include low-light and high dynamic range imaging as well as imaging of scintillator fluorescence for radiation detection.
September 18
Speaker: Daifeng Wang, Biostatistics & Medical Informatics, UW-Madison
Title: AI-Driven Single-Cell Analysis: From Disease Mechanisms to Therapeutic Discovery
Poster: 0918-Daifeng-Wang-poster
Abstract:
Emerging single-cell genomics provides an unprecedented view of cellular and molecular heterogeneity across diseases and clinical phenotypes, yet translating these complex data into mechanistic and therapeutic insight remains challenging. This talk presents an AI-driven computational framework for single-cell analysis that integrates multimodal, graph-based, and physics-informed learning to connect genes, cell functions and dynamics, disease phenotypes, and repurposed drugs. First, cooperative multi-view integration and interpretable model explanation prioritize molecular features and higher-order interactions across single-cell and spatial modalities for Alzheimer’s disease (AD). Next, graph learning approaches refine and associate cell subtypes with AD phenotypes, including cognitive resilience, neuropsychiatric symptoms, and pathological progression, reveal personalized molecular diversity, and identify network-based repurposed drugs for neuropsychiatric disorders. Finally, physics-informed AI models infer continuous dynamics of gene expression, cell populations, and perturbation responses. Multi-agent reinforcement learning further reveals virtual cell environments that represent cellular dynamics and interactions. Together, these methods advance interpretable, mechanistic single-cell analyses for disease modeling, patient stratification, and drug discovery.
September 25
Speaker: Evgeny Shlevkov, Lieber Institute for Brain Development
Title: From Genetics To Therapeutics: Mapping Druggable Mechanisms in Brain Disorders
Poster: 0925-Evgeny-Shlevkov-poster
Abstract: In my talk, I will illustrate how integrating pharmacology with advanced cellular technologies—including high-content imaging, iPSC modeling, and machine learning—complements human genetics in the quest to map druggable cellular signaling nodes in neurodegenerative and neurodevelopmental disorders. I will present how the combination of genetics and cell biology guides the nomination and prioritization of drug targets, how high-content imaging can be used to identify druggable mechanisms driving disease-associated cellular phenotypes, and how machine learning can help identify convergent phenotypes in iPSC models derived from patients. These methods and concepts can be applied across indications, offering a powerful platform for the translational investigation of brain disorders.
October 2
Speaker: David Brundage, School of Veterinary Medicine, UW-Madison
Title: Veterinary Informatics: Out of Distribution by Default
Poster: 1002-David-Brundage-poster
Abstract:
In veterinary medicine, almost every dataset is out of distribution relative to every other one. Species, clinics, laboratories, scanners, stains and report language all change from one dataset to the next, so distribution shift is the normal condition. Veterinary medicine lacks regulation for AI in the clinic, and often builds on existing human clinical AI. The Brundage Lab explores how models transfer across domains and the impact to the clinician. I will describe three lines of work. across 1,171 animal whole-slide images from 14 clades, pathology foundation models trained mostly on human tissue identify a slide’s source dataset almost perfectly within a species, and species recall falls from 0.79 to 0.20 once whole sources are held out. open-weight language models running inside UW turn archived free-text pathology reports into labeled slide cohorts. Anatomic specimen availability constrains how we can provide educational access to students. We explore how Gaussian-splat capture can be leveraged for educating the future veterinarian.
October 9
Speaker: Yin Li, Biostatistics & Medical Informatics, UW-Madison
Title: From Medical Video Analysis to Multimodal Intelligence
Poster: 1009-Yin-Li-poster
Abstract: Video is becoming an increasingly important form of data in healthcare. Cameras are now deployed in operating rooms, intensive care units, exam rooms, and other clinical settings, generating a growing volume of medical video that captures rich information about disease signals, care delivery, and patients’ responses. Yet much of this video is collected primarily for documentation or quality assurance, and its potential to support clinical care and medical research remains largely unexplored. At the same time, recent advances in multimodal large language models (MLLMs) offer new opportunities for medical video analytics. By jointly processing video, audio, text, and other modalities, these models provide a flexible interface for understanding complex clinical events and supporting a broad range of analysis tasks.
In this talk, I will first introduce our prior work on extracting clinically relevant signals from high-dimensional video data. We have developed specialized deep learning models to detect medical tools, recognize care activities, characterize patients’ responses, and quantify diagnostic signals. Building on these efforts, I will then present our current and ongoing work toward generalist MLLMs for streaming video understanding, with the goal of supporting diverse medical video analysis tasks within a unified framework. I will highlight several key directions, including integrating multimodal signals beyond video and text, developing long-term memory for continuous video streams, and enabling richer interaction between AI models and human users. Collectively, these efforts aim to move medical video analysis beyond narrowly tailored perception models toward multimodal AI systems that can continuously observe, reason about, and interact within complex clinical environments.
October 16
Speaker: Muxuan Liang, MD Anderson Cancer Center
Title: Gradient-Augmented AI Calibration for Covariate Adjustment in Randomized Trials.
Poster: 1016-Muxuan-Liang-poster
Abstract: Pretrained AI models learn from vast clinical and medical knowledge and can encode rich prognostic information into predictions for patient outcomes, which can be used to improve the efficiency of treatment-effect estimation in randomized trials. However, predictions from AI models can be random and noisy, and data adaptation is often needed for better predictive accuracy. In this work, we develop a general calibration framework based on augmented inverse-probability weighted (AIPW) estimation in multi-arm trials combining in-context learning and data adaptation. Specifically, we propose GRACIAS, which combines repeated AI-generated outcome predictions based on randomly sampled examples/demonstrations with low-rank features obtained from model gradients. Separate Oaxaca–Blinder least-squares regressions within each treatment arm further 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 extend the calibration framework and its optimality 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-augmented calibrations. 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.
To subscribe to the BMI Seminar mailing list email join-biostat-seminar@lists.wisc.edu.