Embedded Files
Research
What we work on
Graduate and undergraduate students in the Wearable AI Lab study human health data and build AI models for detection, characterization and just-in-time intervention.
Current projects at SJSU
Fall 2023 – present · Graduate & undergraduate students
Collaborative Wearable Sensor Data
Focus
Collecting multimodal wearable data (accelerometer, gyroscope, PPG) and building preprocessing pipelines for missing or noisy signals.
Impact
Shared data infrastructure for every other project in the lab.
Spring 2024 – present · Graduate students
Complex and Rare Physical Activity
Focus
Detecting low-frequency events such as falls and transitional movements in long-duration wearable recordings.
Impact
Activity recognition that holds up beyond common daily activities.
Fall 2023 – present · Graduate & undergraduate students
Robot–Human Collaboration
Focus
Wearable AI for human motion detection to improve communication between people and robots in manufacturing.
Impact
Moves manufacturing toward Industry 5.0 readiness.
Spring 2024 – present · Graduate & undergraduate students
Ethics and Bias in AI
Focus
Evaluating fairness, equity and gender bias in wearable AI models.
Impact
Groundwork for health-equity-focused AI.
Spring 2025 – present · Graduate & undergraduate students
Humanities and AI Collaboration
Focus
Where wearable AI meets ethics and human values.
Impact
Brings wearable AI research to new disciplines and communities.
Foundations
Doctoral research at MD2K, University of Memphis
MD2K (Mobile Data to Knowledge) is one of 11 NSF Big Data Centers of Excellence. Advisor: Dr. Santosh Kumar.
2017 – 2020
ROBAS: Remote Oral Behaviors Assessment System
Field study with synchronized video capture; first labeled dataset of oral behaviors in natural environments, released to the community.
3,440 h sensor data · 412 annotated videos · 30 participants
2017 – 2019
mORAL: oral-hygiene behavior detection
Micro-event hand-to-mouth detection that cut computation by 90%. Published at UbiComp 2019; Distinguished Paper nomination.
3,886 h · 1B+ data points · 95% F1 for brushing/flossing
2019 – 2021
mTeeth: which tooth surfaces get brushed
Lightweight stroke detection and a Bayesian ensemble for surface-level brushing duration. Published at UbiComp 2021.
10,230 brushing instances · <4.2% stroke error · <5% median duration error
2022 – 2023
Detecting "about-to-brush" moments
Predictive models for pre-brushing moments, to time personalized interventions.
2016 – 2017
Physical activity recognition from wrist sensors
Robust activity models extended to heart-failure readmission data and a 300-participant study to find smoking hotspots (SmokingOpp, UbiComp 2020).
4,200 person-days · +10–15% productivity inference accuracy
Report abuse