What Does Big Data Mean for Wearable Sensor Systems?

Nigel H. Lovell, Guang Yang, Alexander Horsch, Paul Lukowicz, L. Murrugarra, Michael Marschollek, Stephen J. RedmondYearbook 2014 · DOI

Abstract

Wearable sensors systems have the potential to generate datasets which are currently beyond our capabilities to easily organize and interpret. In order to successfully utilize wearable sensor data to infer wellbeing, and enable proactive health management, standards and ontologies must be developed which allow for data to be shared between research groups and between commercial systems, promoting the integration of these data into health information systems. However, policy and regulation will be required to ensure that the detailed nature of wearable sensor data is not misused to invade privacies or prejudice against individuals.

Introduction

The field of wearable sensors is undergoing a radical transition from niche clinical tools—exemplified by the 1960s-era Holter monitor—to ubiquitous consumer health technology. As these devices shrink in size and explode in capability, we are entering an era where personal health devices generate "Big Data" at an unprecedented scale. However, a massive gap remains between our ability to collect this data and our ability to interpret it effectively. This position paper, synthesized by the IMIA Wearable Sensors in Healthcare Working Group, argues that while hardware is ready for the mass market, the infrastructure for data standardization, statistical inference, and privacy protection remains underdeveloped.

The "Four Vs" of Wearable Data

To understand the challenges of wearable systems, the authors apply the traditional Big Data framework to the constraints of wearable microelectromechanical systems (MEMS):

  • Volume: The scale of data is immense. A single triaxial inertial sensor (accelerometer, gyroscope, and magnetometer) sampling at 100Hz generates approximately 156 MB of raw data per day. Across a population, this creates a storage and processing burden that traditional health systems are not equipped to handle.
  • Velocity: Proactive health management requires real-time or near-real-time processing to be effective, particularly for emergency applications like fall detection.
  • Variety: Data is highly heterogeneous, originating from smartphones, smartwatches, and body-worn patches, all using disparate formats.
  • Veracity: This is perhaps the most significant hurdle. Wearable data is often "unsupervised." A loose sensor strap or normal physical activity can create noise artifacts that current algorithms may struggle to distinguish from a genuine physiological crisis, such as a heart attack or a fall.

Technical and Statistical Paradoxes

The Battery-Data Paradox

There is a fundamental trade-off between battery longevity and data utility. While pre-processing data on-device (edge computing) saves power by reducing the amount of data transmitted, it often discards raw "Big Data" that might contain yet-to-be-discovered clinical biomarkers.

The Rare Event Problem

Even with petabytes of data, researchers face a "small data" problem when training machine learning models for specific events. If a critical event, such as a fall, only happens once a year for a specific user, the training set remains sparse. This often leads to "overlearning" or overfitting on small training sets, which then fail to perform accurately in diverse, real-world environments.

The Path to Standardization and Integration

For wearable data to be clinically useful, it must be integrated into broader health information systems. The authors advocate for a "DICOM for Wearables"—referencing the universal standard that revolutionized medical imaging. They suggest BioSignalML as a potential markup language to unify disparate sensor streams.

Furthermore, the paper identifies the smartphone as the primary "hub" for this data. However, they note that the common placement of phones in pockets or bags creates significant data quality issues compared to dedicated body-worn sensors, necessitating more robust algorithms to handle varying sensor orientations.

Societal and Regulatory Considerations

The transition to ubiquitous monitoring is not merely a technical challenge but a social one. The authors note a "social friction" where fitness wearables are viewed as status symbols, while devices for chronic illness or aging are often stigmatized.

More critically, the paper warns of a "dystopian scenario" regarding privacy. Without strict government regulation, "Quantified Self" data could be misused by insurance companies to penalize individuals for lifestyle choices or by bad actors to track a user's presence at home.

Conclusion

While wearable sensors offer a path toward proactive, personalized healthcare, the current ecosystem is fragmented. Success will require the development of global standards and ontologies to allow data sharing between research and commercial sectors, alongside robust policy frameworks to protect individual privacy. Simply collecting more data is not enough; the focus must shift toward ensuring the veracity and ethical application of that data.

Summary updated

Abstract

Wearable sensors systems have the potential to generate datasets which are currently beyond our capabilities to easily organize and interpret. In order to successfully utilize wearable sensor data to infer wellbeing, and enable proactive health management, standards and ontologies must be developed which allow for data to be shared between research groups and between commercial systems, promoting the integration of these data into health information systems. However, policy and regulation will be required to ensure that the detailed nature of wearable sensor data is not misused to invade privacies or prejudice against individuals.
BibTeX
@inproceedings{lovell2014what,
  author = {Nigel H. Lovell and Guang Yang and Alexander Horsch and Paul Lukowicz and L. Murrugarra and Michael Marschollek and Stephen J. Redmond},
  title = {{What Does Big Data Mean for Wearable Sensor Systems?}},
  booktitle = {Yearbook},
  year = {2014},
  doi = {10.15265/iy-2014-0019},
}
Figure