Workshop Schedule

Date: Saturday, July 11, 2026

Location: HALL D2, Coex Convention & Exhibition Center, Seoul, South Korea

Time	Duration	Event
8:30 – 8:40 AM	10 mins	Opening Remarks
8:40 – 9:15 AM	35 mins	David Q. Sun - Building Personalized Wellness Optimization through Consumer Health Hardware Consumer sleep hardware generates millions of nightly physiological recordings, but turning passive sensing into actionable intelligence demands new approaches. This talk presents three case studies. First, Autopilot Pro: an agentic reinforcement-learning system personalizing bed temperature across sleep phases, trained on over 10 million user-nights. Second, BCG-FM: a foundation model pretrained on 2.75M hours of ambient cardiac signals, achieving 3.26-year biological age estimation and clinically relevant disease discrimination — suggesting sensor and health foundation models need not require a wearable or a clinic. Third, a hybrid architecture partitioning deterministic and neural computation for structured health text generation. Together, these illustrate how every night becomes both a check-up and an intervention.
9:15 – 9:50 AM	35 mins	James Zou - Learning the language of sleep. I will present our work on building a sleep foundation model from millions of hours of multi-modal sleep recordings. Our model can predict risks for diverse diseases as well as individual's subjective experience of sleep.
9:50 – 10:50 AM	60 mins	Poster Session 1 in Hall A 807-817, 900-913, inclusive
10:50 – 11:25 AM	35 mins	Cyrus Tanade - Crafting Wearable Foundation Models Leveraging Physiological Priors Wearable health poses unique challenges for foundation models due to the complexity of physiological signals and constraints of resource-limited devices. This talk presents two complementary approaches for building wearable foundation models through efficient learning and domain-aware model design. The first approach, HiMAE, leverages hierarchical convolutional masked autoencoders to exploit the inductive biases of physiological signals. By learning multi-resolution representations, HiMAE enables systematic analysis of which temporal scales are most informative for different downstream tasks while remaining compact enough for real-time on-watch deployment. The second approach, xMAE, incorporates physiology-aware cross-modal pretraining by reconstructing ECG from PPG using directional cross-attention that captures meaningful cardiovascular timing relationships. Despite being trained on the MIMIC fingertip recordings, the learned representations transfer effectively to wrist-based wearable data. Together, these approaches illustrate how domain structure, physiological priors, and efficient model design can enable deployable foundation models for wearable health.
11:25 – 12:00 PM	35 mins	Shenda Hong - From Physiological Signal Foundation Models to AI Family Doctor This talk presents a systematic vision toward an "AI Family Doctor," bridging physiological signal foundation models with real-world health services. We address three core challenges: acquiring daily health monitoring data, enabling large models to understand physiological time-series signals, and delivering intuitive health indicators to general users. Key topics include: the construction of HEEDB, the world's largest ECG dataset, and MEETI, the first multimodal ECG dataset; the training of ECGFounder, a foundation model built on the in-house Net1D architecture, together with multimodal large models such as GEM and ECG-R1 that achieve alignment, understanding, and generation between ECG signals and large language models; the introduction of "AI Digital Biomarkers" as a new perspective, spanning cardiac age, vascular age, continuous sleep depth index, and other multidimensional health indicators; and the deployment of the "Wen-Xin-Wu-Yang" AI ECG Monitor (medical device certifications, 300,000+ users) and the ZhunXin Agent in clinical decision support, chest pain triage, and personalized home health management.
12:00 – 12:35 PM	35 mins	Su-in Lee - Auditing AI in Healthcare: Toward Transparency through Explainable AI
12:35 – 2:05 PM	90 mins	Lunch Break
2:05 – 3:05 PM	60 mins	Poster Session 2 in Hall A 914-917, 1000-1015, 1100-1103, inclusive
3:05 – 3:40 PM	35 mins	Suchi Saria
3:40 – 4:15 PM	35 mins	Cecilia Mascolo - Listening to Health: From Digital Biomarkers to Foundation Models Recent advances in AI have accelerated the development of foundation models for health, yet much of this progress has focused on structured clinical records and physiological signals. In this talk, I will argue that everyday sounds represent a rich but largely untapped source of health information and show how AI can transform unstructured audio into structured health representations for clinical and public health applications. Drawing on research spanning respiratory disease screening, passive mobile sensing, population-scale health monitoring, and foundation models for health audio, I will illustrate how machine learning has progressively enabled the extraction of increasingly rich health information from everyday acoustic signals. These examples demonstrate the evolution from task-specific prediction to scalable digital biomarkers and transferable representations, highlighting how unstructured audio can complement structured health data for individual health assessment, population health surveillance, and multimodal AI systems for healthcare.
4:15 – 4:30 PM	15 mins	Coffee Break
4:30 – 5:00 PM	30 mins	Closing Remarks & Award Presentations

Poster Sessions

Sponsors