AI for Medicine and Science
AI systems that accelerate biomedical discovery, disease understanding, and clinical translation.
Research
We develop scalable, multisensory, and generative AI that translates scientific insight into practical tools for medicine and health.
Research Overview
Multisensory Foundation Models
General-purpose models for wearables, sleep, language, sensors, imaging, and multimodal health data.
Personal Health Intelligence
AI that helps people understand, reason, and improve individual physiology, behavior, disease trajectories, and interventions.
AI Agents and Human-AI Ecosystem
Interactive AI systems that reason with people, tools, health data, evidence, and scientific workflows.
Generative and Physical AI for Health
Generative and physical AI models for human signals, body dynamics, interventions, and personal health digital twins.
Trustworthy and Deployable Health AI
Fair, robust, interpretable, transparent, and safe AI, especially for high-stakes medical and health settings.
Research Directions
AI for Medicine and Science
We develop AI methods that turn complex health data into clinical insight and scientific discovery. Our work studies early diagnosis, disease severity, progression, phenotypes, and treatment response across neurological disease, sleep, mental health, cardiometabolic health, and clinical imaging.
Insulin Resistance Prediction
Predicting insulin resistance from wearables and routine blood biomarkers.
Nature 2026
AI Biomarkers for Parkinson's Disease
Passive disease detection and assessment from nocturnal breathing signals.
Nature Medicine 2022
In-Home Disease Progression Tracking
Tracking gait and medication response at home using radio-wave sensing.
Science Translational Medicine 2022
Medical Vision-Language Models
Auditing expert-level foundation models for disease detection across populations.
Science Advances 2025Multisensory Foundation Models
We build general-purpose models that learn from heterogeneous multisensory signals of humans, including wearable sensors, sleep, motion, glucose, physiology, imaging, natural language, and more. These models support transfer across tasks, devices, cohorts, time, and clinical settings.
Sleep-Language Models (SleepLM)
Natural-language intelligence for human sleep from physiological time series.
ICML 2026 Spotlight
Open Sleep Foundation Models (OSF)
Open foundation models for sleep representation learning and transfer.
ICML 2026
Sensor-Language Models (SensorLM)
Learning the language of wearable sensors through large-scale sensor-text pretraining.
NeurIPS 2025
Large Sensor Models (LSM)
Scaling laws and generalization behavior for foundation models over wearable data.
ICLR 2025Personal Health Intelligence
We design AI systems that reason over individual physiology, daily context, and longitudinal data to support personalized health insight. Our goal is to model individual baselines, detect meaningful deviations, forecast risk, and support personalized monitoring, intervention, coaching, and treatments.
Everyday Passive Health Sensing
Physiological monitoring from remote visual and ambient signals.
Nature 2026
Health Reasoning over Time Series (HEARTS)
Benchmarking language-model reasoning over health time series.
ICML 2026
Personal Health Agent
Agentic systems that organize longitudinal personal health data and reasoning.
arXiv 2025AI Agents and Human-AI Ecosystem
We study interactive AI systems that can move from passive prediction to active reasoning and action: reading data, using tools, checking evidence, running analyses, and interacting with researchers, clinicians, and people.
Towards a Science of Scaling Agent Systems
When and why agent systems improve with scale, tools, and coordination.
Nature Machine Intelligence 2026
AI Co-Data-Scientist (CoDaS)
Agentic biomarker discovery from wearable sensors and health data.
arXiv 2026
Generative and Physical AI for Health
We develop generative and physical AI systems that model the structure, dynamics, and variation of human health data. Our work studies physiological rhythms, embodied signals, intervention-driven representations, and health world models to support simulation, generation, and robust health understanding.
Raptor
Scalable train-free embeddings for 3D medical volumes using pretrained 2D foundation models.
ICML 2025 Spotlight
SimPer
Self-supervised learning for periodic targets from ubiquitous physiological sensors.
ICLR 2023 OralTrustworthy and Deployable Health AI
Health AI must work across populations, hospitals, devices, and real-world data conditions. We study fairness, robustness, missingness, distribution shift, interpretability, and deployment-centered evaluation to make health AI more reliable and generalizable.
The Limits of Fair Medical Imaging AI
Understanding when clinical AI fails under real-world distribution shifts.
Nature Medicine 2024
Demographic Bias in Computational Pathology
Auditing demographic bias and misdiagnosis in pathology foundation models.
Nature Medicine 2024
Deep Imbalanced Regression
Learning continuous targets under severe label imbalance and distribution shift.
ICML 2021 Oral