Problematic Behavior

How do we use technology to power and maintain behavior change?

  • Overeating: leading to the obesity epidemic and contributing to chronic illness
  • Undernutrition: leading to malnourishment in low income countries
  • Constipation in children: reducing ER visits through behavrioal of child constipation
Our Approach

How Do We Approach Those Problems?

Approach: Combining behavioral science and computer science to build passive sensors and data analytics to help us Understand, Sense, Detect, Predict, and ultimately Prevent problematic behavior. Below we provide on-going research related to overeating.

Understand

Positive correlation between caloric intake and eating behaviors

Why: To understand what features characterize problematic behaviors and what are the antecedents to and proximal determinants of problematic behaviors (like overeating).

How: Use surveys, focus groups, interviews and human-centered design methods to help us learn how to optimize passive sensing systems.

Example: A recent study of ours showed that feeding gestures, eating duration, and number of swallows positively correlate with caloric intake. We also designed systems to detect them using wearables people will actually wear. Our survey of 93 people revealed that 88% said they would wear our wrist-worn sensor, and 62% said they would wear a personalized version of our neck-worn sensor.

Responsive image

62% of those surveyed say they would wear the device if personalized.

Project Details
  • Categories:
    Understand
  • Techniques:
    HCI, Studies/Focus Groups/Interviews, Surveys
  • Responsive image

    Calories positively correlate with feeding gestures, swallows and eating duration

Sense

Build mHealth Sensor systems to detect overeating and its contextual factors

Why: To help us detect the essential behavioral factors.

How: Build passive sensing systems comprising sensors, an information gateway (smartphone), and a backend database.

Example: Our solution involves a neck-worn sensor to detect behaviors related to swallowing and chewing; a wrist-worn sensor to detect feeding gestures, heart rate and activity; and a video camera to capture truth and a smartphone to capture everything else.

Responsive image

Non-invasive wearable sensors

Project Details
  • Categories:
    Sense
  • Techniques:
    mHealth Sensor Systems, Passive Sensing, Wearables and Embedded Systems
  • Responsive image

    Wearing multiple sensors to detect and characterize eating in the field

Detect

Applying bag-of-words to segment and detect feeding gestures

Why: To transform the sensor signals into meaningful detectable features.

How: Build human activity recognition frameworks and methods that improve our ability to detect important behavioral features.

Example: We have designed generalized and personalized machine learning models to detect low-level features like feeding gestures, swallows, and alone time.

Responsive image

Applying bag-of-words to the raw accelerometer data to create feeding gestures.

Project Details
  • Categories:
    Detect
  • Techniques:
    Segmentation, Feature Extraction, Low Level Machine Learning
  • Responsive image

    Building generalized and personalized models using classification algorithms (e.g. Random Forest).

Predict

Applying unsupervised clustering algorithms to determine behavioral phenotypes to overeating

Why: To understand which features are the most predictive of the behavioral phenotypes, and to predict the behavior in advance in order to trigger intervention.

How: Build machine learning and mixed-effects based models to discover the behavioral phenotypes and factors that are most relevant.

Example: We are applying unsupervised machine learning algorithms to understand the most predictive behavioral features and in doing so validating known clinical behavioral phenotypes and discovering new ones.

Responsive image

Predicting emotional eating

Project Details
  • Categories:
    Predict
  • Techniques:
    High Level Machine Learning, Statistical Analysis, Behavior Models
  • Responsive image

    Affect-based overeating behavioral phenotype detected

Prevent

Collaborate with domain experts to test interventions

Why: To prevent problematic behaviors.

How: Work with behaviorist, interventionist and domain experts to help prevent problematic behaviors.

Example: We are working with behaviorists to design just in time adaptive interventions (JITAIs) that help prevent problematic eating behaviors.

Responsive image

Predicting to prevent overeating while watching TV in lab and chest camera establishes ground truth in field

Project Details
  • Categories:
    Prevent
  • Techniques:
    Behaviorist, Interventionist, Medical Expert
  • Responsive image

    Just-In-Time Adaptive Intervention Model for Overeating