MyDJ: Sensing Food Intakes with an Attachable on Your Eyeglass Frame

Title of the Paper

MyDJ: Sensing Food Intakes with an Attachable on Your Eyeglass Frame

Paper Information

• Research Area: Wearable devices, automated dietary monitoring, health technology
• Keywords: wearable device, automated dietary monitoring, multimodal sensing, energy efficiency, social acceptability, eyeglass frame

Research Background and Issues

• Identified Problems or Challenges

  1. Traditional manual dietary recording methods are cumbersome and often forgotten.
  2. Existing automated dietary monitoring systems have limited accuracy outside laboratory settings or are overly intrusive (e.g., head-mounted devices).
  3. Previous studies using eyeglasses for dietary monitoring often require custom frames, consume significant power, and are unsuitable for widespread use.

• Significance Dietary monitoring is crucial for maintaining healthy eating habits, managing weight, and controlling chronic diseases. Developing accurate, efficient, and socially acceptable monitoring devices can increase adoption rates and provide users with automated dietary intervention tools.

• Research Motivation and Related Work

  • Automated dietary monitoring has immense potential for societal impact, but current methods are limited by social acceptability and long-term usability.
  • Eyeglasses, as a widely accepted daily wearable form, are structurally advantageous for detecting chewing signals due to their proximity to the mouth.
  • This study aims to design a device that can be easily attached to any eyeglass frame while addressing issues of power consumption and sensor reliability.

Solution

• Proposed Method or Solution

  1. Develop the MyDJ system, a dietary monitoring device attachable to any eyeglass frame.
  2. Utilize piezoelectric sensors and accelerometers to capture distinct yet complementary chewing signals through multimodal sensing.
  3. Design low-power hardware combined with feature extraction and deep neural network (DNN) models for real-time data processing and classification.

• Innovations

  1. No need for custom eyeglass frames, offering broad compatibility.
  2. Multimodal fusion of low-power sensors significantly improves detection accuracy and environmental robustness.
  3. Enhanced data processing algorithms and lightweight model design achieve more efficient power management.

• Implementation Steps and Key Technologies

  1. Hardware Architecture Design:
     • Piezoelectric sensors detect mechanical vibrations of the temporalis muscle.
     • Accelerometers capture mechanical wave propagation caused by chewing.
     • Custom PCB with microcontroller unit (MCU), Bluetooth communication, and small battery integration.
  2. Data Preprocessing and Feature Extraction:
     • Generate signal windows using Short-Time Fourier Transform (STFT) and Mel-Frequency Cepstral Coefficients (MFCC).
     • Feature selection algorithm (JMIM) identifies the optimal feature set, balancing power consumption and accuracy.
  3. Deep Neural Network Classification:
     • Use a DNN model to classify window-level data into "eating/non-eating."
     • Detect complete eating events based on a 15-second time frame.

Research Results

• Specific Outcomes

  1. In experiments, MyDJ achieved an F1 score of 0.919 and an accuracy of 0.984 for dietary monitoring.
  2. Battery life significantly extended: 66 hours of operation with a 220mAh battery, 4.03 times longer than existing solutions.
  3. During a one-week long-term study, MyDJ detected 93% of eating events.

• Advantages

  1. High social acceptability: Users reported a comfort level of 94.95% when wearing MyDJ-equipped glasses, comparable to regular glasses.
  2. Compared to existing dietary monitoring methods, MyDJ significantly reduces false alarms and power consumption.
  3. Strong adaptability: MyDJ is compatible with various eyeglass frames without customization.

• Experimental or Evaluation Results

  1. Data Collection and Evaluation:
     • 237 hours of data (24 participants, daytime study).
     • 477 hours of data (6 participants, week-long study).
  2. In the week-long experiment, MyDJ's average F1 score of 0.777 was further optimized to 0.849 through personalized models.
  3. The combined mode of piezoelectric sensors and accelerometers performed better than single-sensor setups.

• Limitations and Future Directions

  1. Limitations:
     • Not suitable for individuals who do not need or prefer not to wear glasses.
     • Generalizability of training data across different eyeglass frames requires further validation.
  2. Future Directions:
     • Enhance compatibility testing and dataset expansion for various eyeglass frames.
     • Develop lighter, more compact, and better-packaged device versions.
     • Explore user-friendly annotation tools to reduce the time cost of personalized model tuning.

Remarks

This study comprehensively explores wearable device design, data analysis, and user acceptability, representing a significant step toward the practical and societal application of dietary monitoring systems.