X-PAI: Real-Time AI for Sleep Diagnosis

Wednesday 19 Nov 2025, 10:30 → 17:30 US/Pacific

403 (UW ECE)

Shih-Chieh Hsu (University of Washington Seattle (US))

The Accelerated AI Algorithms for Data-Driven Discovery (A3D3) Institute at the University of Washington is proud to announce its partnership with the International Institute for Integrative Sleep Medicine (IIIS) at the University of Tsukuba through the Cross-Pacific AI Initiative (X-PAI) program, a global initiative dedicated to advancing trustworthy and efficient artificial intelligence in health and longevity research. This collaboration marks the launch of a multifaceted program, including not only cutting-edge research but also a structured exchange program that fosters cross-institutional scientific exchange, joint workshops, and joint data analysis for rapid innovation and training of future leaders in AI for sleep medicine.

The project, sponsored by Amazon, titled "Real-Time Adaptive AI for Sleep Diagnosis, Treatment and Therapy," brings together an interdisciplinary team led by Principal Investigator Shun Nakajima (IIIS, University of Tsukuba) and Co-PIs Shih-Chieh Hsu (A3D3, University of Washington). Key team members include Amy Orsborn, Eli Shlizerman, and Scott Hauck from A3D3 (University of Washington); Kei Muroi, Shio Maeda, Hiroku Noma, and Hiroyuki Kitagawa from IIIS (University of Tsukuba); and Haruhito Tanaka, M.D., and Hiroaki Yamamoto from Gifu Mates Sleep Clinic. Together, their expertise spans neuroscience, sleep medicine, real-time AI, brain-machine interfaces, and trustworthy machine learning.

This seminar is the inaugural public event hosted at the University of Washington for this transformative international partnership. Attendees will learn about the project’s aim to make sleep diagnosis more accurate, accessible, and ethically grounded, as well as the innovative training and exchange opportunities supported by X-PAI.

Zoom (required registration):
https://washington.zoom.us/meeting/register/511WLU4fSWC9LgBpy-I35g

IMG_6403.HEIC

IMG_6403.jpg

IMG_6404.HEIC

IMG_6404.jpg

IMG_6410.HEIC

IMG_6410.jpg

IMG_6417.HEIC

IMG_6417.jpg

IMG_6418.HEIC

IMG_6424.HEIC

IMG_6424.jpg

IMG_6426.HEIC

IMG_6426.jpg

IMG_6632.jpg

10:30 → 11:00 Closed-loop BCI for the treatment of brain disorders

Closed-loop brain-computer interface (BCI) for controlling brain states is an emerging technology for treating brain disorders. Current time-invariant controllers have limited ability to address the nonlinearity, time-variation, and disturbances in brain networks, which can cause performance degrading or even instability of control. We design, analyze, and validate a robust adaptive BCI control algorithm that can track nonlinearity and time-variation and is robust to noise and disturbances. Our algorithm achieves accurate, stable, and robust control and significantly outperforms state-of-the-art controllers.Our algorithm can help future designs of precise and safe brain stimulation systems to treat brain disorders and facilitate brain functions.

Speaker: Hao Fang (University of Washington)

11.19 X-PAI Seminar.pdf

11.19 X-PAI Seminar.pptx

YouTube

15:30 → 16:00 Application of AI to Psychotherapy and Assessment for Sleep Disorder

We are building an AI therapist for sleep problems based on rich real-world data from insomnia and sleep apnea patients. Using remote CBT-I trials and large multimodal registries (sleep studies, questionnaires, speech, video, and transcripts), we study how therapists and patients interact and which communication patterns relate to better outcomes. These findings are then translated into concrete “skills” for the AI ​​therapist (for example, how to listen, reflect, and respond safely). Finally, we combine clinical results and public attitudes to evaluate when, for whom, and under what conditions AI-supported psychotherapy can be used safely instead of—or alongside—human therapists.

Speakers: Hiroku Noma (University of Tsukuba), Shio Maeda (University of Tsukuba)

YouTube

16:10 → 16:40 Learning Representations from Neural Population Activity: Addressing Neural Variability Across Scales

Interactions between individual neurons, each characterized by distinct intrinsic physiological properties, collectively give rise to the population responses underlying complex animal behaviors. These responses exhibit highly variable dynamics across trials, recording sessions, and behavioral contexts, hampering effective extraction of useful information from neural activity. Consequently, modeling and decoding from population activity necessitate methods capable of learning stable representations that capture the underlying structure of neuronal activity, despite the noise and partial observability inherent in population recordings. In this talk, I will present my previous studies aimed at learning robust representations from population activity in the presence of neural variability, with applications targeting few-shot motor behavior decoding and silent speech decoding. Together, these contributions advance computational methods for analyzing neural dynamics across diverse experimental conditions, paving a way towards high performing and robust brain-computer interfaces.

Speaker: Trung Le (University of Washington)

YouTube

16:50 → 17:20 Toward Multimodal Spatial Intelligence: Representing and Reasoning About Spatial Environments Through Vision and Audio

Reliable spatial intelligence is essential for AI systems that aim to represent and reason about complex real-world environments for applications in immersive experiences, autonomous monitoring, and scientific sensing. Many spatial properties are difficult to infer from vision alone, while acoustics provides complementary information about geometry, materials, and occlusions. Integrating these modalities enables richer and more stable spatial representations that support both reconstruction and higher-level reasoning. In this talk, we will present a sequence of multimodal systems that progressively advance joint visual and acoustic representations of real environments. INRAS, BEE, AVCloud, and SoundVista show how integrated sight and sound can reconstruct audio-visual spaces with increasing efficiency, fidelity, and robustness, moving from static scenes toward dynamic environments. Building on this foundation, Savvy introduces spatial reasoning and contextual environmental understanding through unified multimodal cues, along with a benchmark formulation that offers a closer evaluation of this emerging problem. These works outline a coherent pathway toward AI systems that represent and reason about spatial environments through multimodal perception.

Speaker: Mingfei Chen (University of Washington)

YouTube