Integrated image and sensor-based food intake detection in free-living

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Dietary Monitoring

Background:

• Accurate dietary monitoring relies on detecting eating episodes.
• Wearable sensors like the Automatic Ingestion Monitor v2 (AIM-2) can detect eating episodes but may produce false positives.
• Existing methods using image or sensor data alone have limitations in precision.

Purpose of the Study:

• To reduce false-positive detections of eating episodes using the AIM-2 wearable sensor.
• To enhance the accuracy of dietary monitoring systems.
• To evaluate a combined image and sensor-based approach for eating episode detection.

Main Methods:

• Utilized the AIM-2 wearable sensor with thirty participants in pseudo-free-living and free-living conditions.
• Employed three detection methods: image recognition of food/beverages, accelerometer recognition of chewing, and a hierarchical classification combining both.
• Collected sensor and image data over two days per participant.

Main Results:

• The integrated image and sensor-based method achieved 94.59% sensitivity, 70.47% precision, and 80.77% F1-score in a free-living environment.
• This combined approach demonstrated an 8% higher sensitivity compared to individual methods.
• The hierarchical classification significantly improved the accuracy of eating episode detection.

Conclusions:

• Combining image and accelerometer data from the AIM-2 sensor effectively reduces false positives in eating episode detection.
• The proposed hierarchical classification method offers a more robust solution for dietary monitoring.
• This advancement contributes to more accurate and reliable automated dietary assessment tools.