TL;DR
This scoping review synthesizes the current landscape of wearable sensors for automatically detecting eating behavior in real-world settings. It highlights the growing potential of these technologies to overcome the limitations of traditional self-report methods in public health research, while also identifying key challenges in sensor development, data collection, and evaluation.
What Problem Does This Paper Address?
Accurately measuring dietary intake and eating behaviors is crucial for understanding and preventing chronic diseases like obesity, type 2 diabetes, and heart disease. However, traditional methods like 24-hour recalls and food diaries are prone to participant burden and recall bias, leading to inaccurate data. While objective measurement tools, particularly wearable sensors, offer a promising solution by passively collecting data in naturalistic settings, there's a lack of clarity on their current capabilities, the variety of approaches used, and how their performance is evaluated. This paper addresses this gap by providing a comprehensive overview of the existing research on wearable-based in-field eating detection for public health.
Methods
This paper presents a scoping review, a systematic approach to map the existing literature on a topic. The authors searched five major databases (PubMed, SCOPUS, IEEE Xplore, ACM Digital Library, and Google Scholar) for studies published up to December 2019. They included studies that used wearable devices/sensors for automatic eating detection in "in-field" (naturalistic) settings, reported at least one evaluation metric, and were peer-reviewed. Data extraction focused on study characteristics such as sample size, sensor types, ground-truth methods, eating outcomes, and evaluation metrics. The review then synthesized these findings to identify trends, challenges, and future directions.
Datasets
This paper is a review of existing literature and does not introduce or use new datasets for training or evaluation. The "datasets" in this context are the individual studies that were included in the review, each with its own participant data collected in real-world settings.
Key Insights and Ideas
- The Rise of Smartwatches: Accelerometers and gyroscopes, commonly found in smartwatches, are the dominant sensor types for eating detection, reflecting the trend towards user-friendly and accessible wearable technology.
- Self-Report Still Reigns for Ground Truth: Despite the push for objective measurement, self-report methods (like eating logs) remain the most common way to validate sensor data, highlighting a persistent reliance on subjective measures.
- In-Field is Key, But Challenging: The paper strongly emphasizes the importance of "in-field" or "free-living" studies, as behavior naturally differs from lab settings. However, deploying and evaluating sensors in these environments presents significant technical and logistical hurdles.
- Lack of Standardization is a Major Hurdle: The wide variation in how eating outcomes are defined and how sensor performance is evaluated (e.g., different metrics like accuracy, F1-score, precision) makes direct comparison between studies difficult and hinders progress.
- Context is the Next Frontier: While current research focuses on detecting eating events, there's a growing recognition that incorporating contextual information (social setting, location, activity) is crucial for a deeper understanding of eating behavior.
Key Results
- 40 Studies Included: The review analyzed 40 studies that met the eligibility criteria, published between 2011 and 2019.
- Multi-Sensor Systems Common: The majority of studies (65%) used multi-sensor systems (more than one sensor) to improve detection accuracy.
- Accelerometer Dominance: Accelerometers were the most frequently used sensor (62.5%), followed by gyroscopes (37.5%).
- "Eating Occasions" Most Studied Outcome: The most common eating outcome measured was "inferred eating occasions" (e.g., eating events, episodes), reported in 37 out of 40 studies.
- Accuracy and F1-Score Most Reported Metrics: Accuracy was the most frequently reported evaluation metric (12 studies), followed closely by F1-score (10 studies).
- Varied Sample Sizes and Durations: Sample sizes ranged from 1 to 104 participants, with a mean of 10.83. Data collection durations varied widely, from 2 hours to 1 month.
Limitations and Future Work
Limitations:
- Strict Inclusion Criteria: The review excluded studies that used semi-controlled settings, potentially missing valuable insights.
- Lack of Standardized Outcomes/Metrics: The heterogeneity in defining eating outcomes and reporting evaluation metrics made direct comparison and synthesis challenging.
- Potential Bias in Reporting Metrics: The authors acknowledge that reporting the "best" evaluation metric from each study might introduce bias.
- Omitted Data: Some evaluation metric results were only available in figures and not textually or numerically, leading to their omission from the review's data extraction.
Future Work:
- Standardization is Crucial: The authors strongly advocate for the development of standardized terminology for eating outcomes and standardized evaluation metrics to enable better comparability across studies.
- Multidisciplinary Collaboration: Input from computer scientists, health behaviorists, and nutritionists is needed to identify meaningful eating outcomes and develop robust detection algorithms.
- Integrating Context: Future research should focus on integrating contextual information (social, physical, psychological) with eating behavior data to develop more comprehensive models.
- Exploring Novel Ground-Truth Methods: While wearable video cameras are promising, challenges with privacy and participant comfort remain. Exploring methods like mobile ecological momentary assessment (mEMA) for ground-truth validation is recommended.
- Focus on Real-World Robustness: For successful long-term deployment, the focus needs to shift from novelty to accuracy, robustness, and reliability of sensors and algorithms in unconstrained, real-world settings.
Summary updated