In recent years, the surge in wearable technology has revolutionized the way we capture and interpret human activity. Smartwatches and similar devices, embedded with accelerometers, gyroscopes, and heart rate monitors, have primarily been used within controlled laboratory environments to classify basic physical movements such as sitting, standing, or walking. However, researchers at Washington State University (WSU) have now taken a significant step forward by developing a novel algorithm capable of decoding a more comprehensive and nuanced spectrum of daily activities from smartwatch data collected “in the wild.” This innovation could usher in transformative changes in healthcare monitoring, cognitive assessment, and personalized rehabilitation.
The heart of this breakthrough lies in a sophisticated artificial intelligence system designed around a feature-augmented transformer model. Unlike traditional classification systems, this model leverages vast amounts of labeled data to detect higher-level, goal-directed behaviors such as cooking, working, socializing, or running errands by analyzing sensor signals during everyday life outside laboratory constraints. The algorithm achieves this through advanced pattern recognition that synthesizes diverse smartwatch sensor inputs over time, thereby accurately mapping complex behavioral patterns in real-world settings.
Over a span of eight years, the WSU research team gathered smartwatch data from more than 500 participants involved in various studies. Participants were prompted randomly throughout the day to self-report their ongoing activity from a predefined list comprising 12 categories, including sleeping, traveling, eating, and relaxing, among others. This unique approach yielded an unprecedented dataset exceeding 32 million labeled data points, each representing a one-minute segment of recorded activity paired with participant-reported labels. Such an expansive and richly annotated dataset provided a fertile ground for training the transformer model and testing its ability to generalize across diverse individuals.
What distinguishes this work is not only its scale but its capacity to recognize high-level functional activities essential for daily living, which are often reflective of a person’s independence and well-being. Clinical practitioners have long faced challenges in remotely assessing how individuals, especially the elderly or chronically ill, manage routine but critical activities such as handling finances, preparing meals, or performing self-care tasks. Traditional clinical visits offer limited snapshots, and prior use of wearable devices mostly focused on simple motion detection. The model developed by WSU researchers bridges this gap by enabling continuous, automated assessment of functional behaviors through passive monitoring.
The technical backbone of the system centers on cutting-edge transformer architectures, which have transformed natural language processing and signal analysis by capturing contextual relationships over time more effectively than prior recurrent or convolutional models. By augmenting features derived from multi-modal smartwatch signals, the transformer model aggregates temporal dependencies and subtle signal variations associated with different activities. This composite analysis enriches the recognizability of goal-directed behaviors despite the inherent noise and variability present in uncontrolled, “in-the-wild” data collection settings.
Achieving an overall activity recognition accuracy of nearly 78%, the system marks a notable advance in wearable sensing technology. This level of precision allows for actionable insights into patterns of behavior that correlate with cognitive and physical health metrics. For example, decreased frequency or altered sequences of specific activities can serve as early indicators of cognitive decline or diminished physical capacity. Consequently, the model offers potential pathways for proactive healthcare interventions that preserve independence and improve quality of life for at-risk populations.
Beyond clinical applications, the research lays foundational groundwork for the integration of human-centered artificial intelligence in digital health ecosystems. By capturing detailed behavioral signatures unobtrusively, such frameworks may support remote caregiving, personalized therapy regimens, and even automated clinical diagnostics once coupled with electronic health records and genetic information. Furthermore, making the anonymized dataset and modeling methods publicly available opens doors for the broader scientific community to innovate around activity recognition, behavior prediction, and human health analytics.
Lead researcher Diane Cook, a Regents Professor in WSU’s School of Electrical Engineering and Computer Science, emphasizes the societal importance of this work. She highlights how understanding an individual’s capability to perform critical activities—like bathing, feeding oneself, or managing finances—is fundamental to assessing independence. “If we can describe a person’s behavior in categories that are well recognized,” Cook notes, “we can start to talk about their behavior patterns and changes, which in turn relate to measures of cognitive health and functional independence.”
The project’s funding by the National Institutes of Health underscores the biomedical community’s recognition of wearable AI’s transformative potential. The interdisciplinary nature of the research spans computer science, behavioral health, and gerontology, reflecting a growing emphasis on data-driven precision medicine. By innovating at the intersection of robust AI modeling and real-world human activity monitoring, the WSU team exemplifies how wearable sensors can move beyond step counting to become critical tools in managing complex health trajectories.
Moving forward, researchers aim to refine the model’s capabilities, exploring automated clinical diagnosis and investigating associations between activity patterns, genetics, and environmental variables. Such studies could enrich personalized medicine by contextualizing behavioral data within an individual’s unique biological and socio-economic framework. Moreover, the deployment of this technology in everyday consumer smartwatches and healthcare settings promises to democratize access to health monitoring, empowering individuals and clinicians alike with continuous, actionable insights.
In summary, the groundbreaking work by Washington State University researchers represents a significant leap in translating raw wearable smartwatch data into meaningful, high-level assessments of human activity in naturalistic environments. Through an innovative transformer-based AI approach and an extensive, meticulously curated dataset, this system offers a promising new avenue for remote health monitoring, cognitive assessment, and functional independence evaluation. As wearable tech becomes ubiquitous, such advances pave the way for a future where continuous health insights are seamlessly integrated into daily life, potentially reducing healthcare costs and enhancing well-being worldwide.
Subject of Research: Functional activity recognition using AI from real-world smartwatch data to assess cognitive and physical health.
Article Title: Feature-Augmented Transformer Model to Recognize Functional Activities from in-the-wild Smartwatch Data
News Publication Date: 4-Jul-2025
Web References:
https://ieeexplore.ieee.org/document/11071689
http://dx.doi.org/10.1109/JBHI.2025.3586074
References: Washington State University research team; National Institutes of Health funding
Keywords: wearable technology, smartwatch data, activity recognition, transformer AI, cognitive health assessment, functional independence, digital health, behavioral monitoring, machine learning, real-world data, healthcare innovation
