Tutorials

Patterns of Missingness in Wearable Data

Organizer:

Jessilyn Dunn, Duke University

Presenters: Jessilyn Dunn (Department of Biomedical Engineering, Duke University, jessilyn.dunn@duke.edu), Hayoung Jeong (Department of Biomedical Engineering, Duke University, hayoung.jeong@duke.edu), Leeor Hershkovich (Department of Biomedical Engineering, Duke University, leeor.hershkovich@duke.edu), Harrison Kane (Department of Biomedical Engineering, Duke University, harrison.kane@duke.edu), Eric Hurwitz (Department of Genetics, University of North Carolina at Chapel Hill ehurwitz@email.unc.edu

As wearable devices become increasingly popular in biomedical research and health monitoring, addressing missingness in their data streams is essential for valid and reliable analysis. Missingness due to nonwear or irregular sampling is common, yet remains under-characterized and poorly handled in most analytical workflows. This tutorial responds to a growing need in the BHI community to rigorously address missing data challenges in longitudinal observational studies. We will integrate foundational concepts, real-world case studies, and a hands-on session using the All of Us dataset. The tutorial is designed for biomedical researchers and data scientists working with wearable data, digital health, and missing data methodologies. Participants will gain practical tools and conceptual frameworks to identify, interpret, and utilize missingness patterns for more robust health research outcomes.

Interactive AI Optimization for Healthcare: A Riemannian Manifold Approach to Tailored Clinical Decision-Making

Organizer:

Farzaneh Nikbakhtsarvestani

Presenters: Farzaneh Nikbakhtsarvestani (Ontario Tech University)

The proposed topic for the tutorial introduces a novel optimization framework that leverages Riemannian manifold geometry to address the complexity and non-linearity inherent in high-dimensional biomedical data. By navigating the solution landscape through geodesic convexity, this approach enhances the ability to find global optima in constrained many-objective problems. The integration of interactive, population-based AI optimization allows practitioners to guide the search process based on clinical priorities, overcoming the limitations of one-size-fits-all pre-trained models. Additionally, real-time calibration of model performance metrics ensures alignment with decision-makers’ expectations throughout the optimization. This tutorial is ideal for researchers, clinicians, and developers interested in advanced AI methods for personalized healthcare solutions

Robust and Reproducible AI: Statistical Foundations for Machine Learning in Medical Imaging

Organizer:

Khashayar Namdar

Presenters: Ernest Namdar, (Department of Medical Imaging, University of Toronto; Department of Pharmacology, Faculty of Medicine, Dalhousie University), Pascal Tyrrell, (Department of Medical Imaging, Institute of Medical Science, Department of Statistical Sciences, University of Toronto)

Recent advances in machine learning (ML) have led to remarkable performance gains in medical imaging tasks. However, many models still suffer from poor generalizability, irreproducibility, and limited clinical trust, often due to inadequate statistical foundations. This workshop addresses the critical need to integrate robust statistical reasoning throughout the ML pipeline, from study design and data curation to model evaluation and reporting. We will focus on practical, interpretable, and reproducible workflows tailored for biomedical imaging applications. The workshop is intended for AI researchers, clinicians, biomedical engineers, and statisticians who aim to build or evaluate trustworthy AI tools for healthcare. This tutorial will be structured as a hands-on learning session with guided walkthroughs of three key ML tasks in medical imaging: classification, segmentation, and regression. Participants will work with open-source datasets and implement pipelines using state-of-the-art tools. The tutorial will begin with a 45-minute overview of statistical principles for robust ML, followed by three 30-minute lab sessions (one per task) with live coding and support. The final 15 minutes will be reserved for Q&A and discussion on reproducibility and clinical translation.
Key Concepts
  • Learning Curves and Data Sufficiency
  •  Confidence
  •  Calibration
  •  Proper Validation
  •  Uncertainty Quantification and Statistical Significance
Search this website