Abstract
Annals of Behavioral Medicine (SBM)
Annals of Behavioral Medicine (SBM)
Boyang Wei, Glenn Fernandes, Farzad Shahabi, Christopher Romano, Annie W. Lin, Nabil Alshurafa
Link: TBD
Year: 2025 （accepted）
Background: Accurate monitoring of eating behaviors is crucial for managing conditions like obesity, diabetes, and eating disorders. Wearable devices like smartwatches offer a promising solution for continuous, objective monitoring, which increase user adherence compared to self-report. However, implementing real-time eating detection on these devices is challenging due to their limited processing power, memory, and battery life. To address these issues, we've developed a new approach for efficient, accurate real-time eating detection on wearable devices. Our solution uses a dataset of typical eating cycle patterns collected on wrist-worn devices to create models that overcome device limitations while providing reliable, continuous monitoring.
Method: We proposed an advanced machine learning approach called Neural Architecture Search (NAS) to find the best model design for detecting eating behaviors. This process explored various deep learning architectures that recognize eating gestures in inertial measurement unit (IMU) data from the wrist. The search process was guided by a strategy that balanced the tradeoff between accuracy and the device's limitations, such as memory size and processing speed. The data we used came from motion sensors (accelerometer and gyroscope) on the wrist collected from 12 participants for 246 minutes. We analyzed this data in 5-second intervals, with each interval overlapping the previous one by half.
Result: Our approach identifies the best model architecture (2-D convolutional neural network (CNN) with long short-term memory (LSTM) layers), achieving high accuracy in eating detection (84.81% F1-score) while meeting strict on-device constraints. The best model requires less than 225.38 KB of memory, performs real-time inference with latency under 36.74 ms on typical wrist-worn hardware, and maintains battery life around 25.7 hours under continuous monitoring.
Conclusion: Clinically, our proposed method allows healthcare providers to detect patients' eating patterns using wrist-worn devices, enabling accurate assessment of dietary adherence and timely treatment adjustments. Continuous detection allows identification of eating triggers over time, enabling personalized behavior modification strategies and just-in-time Interventions. For conditions like diabetes or obesity, this facilitates precise insulin dosing or immediate feedback on caloric intake, potentially reducing complications and improving health outcomes.
Nabil Alshurafa, Boyang Wei, Farzad Shahabi, Christopher Romano, Tammy Stump, Annie W. Lin, Angela F. Pfammatter
Link: https://doi.org/10.1093/abm/kaae014 (page 248)
Year: 2024
Background: Dietary intake often extends beyond calorie and nutrient intake, which are not often captured by current dietary self-reported assessments. Advances in mobile computing can improve the validity of diet assessments using objective measurement and enable real-time capture of factors that surround one’s eating behavior. Using wearable sensing and event-based ecological momentary assessment (EMA), we build machine-learned (ML) models using behavioral, social, and psychological features to assess which are most predictive of overeating in people with obesity. Our results aim to empower our ability to design timely interventions to detect, predict, and preempt overeating.
Method: 60 participants with obesity (BMI ≧ 30 kg/m2) completed eventbased EMAs and wore a smartwatch, necklace sensor, and activity-oriented camera for 14 waking days. The EMAs captured social and psychological features, and foods consumed during each eating episode. To verify food records, participants completed daily 24-hour diet recalls. The sensor devices continuously captured video and physiological metrics. Overeating was defined as any meal exceeding 1 standard deviation of a participant's mean per-meal calorie intake. We build ML models using XGBoost (a highly efficient, flexible, and portable algorithm) based on features captured from sensors [sensor only], features estimated from EMA [EMA only], and features captured from both [sensor+EMA]) to classify meals as overeating or non-overeating.
Results: The EMA only dataset (n=48, 2,302 meals, 368 overeating) achieved an AUC score of 0.814, the 5 most predictive features being food type, food source, night eating, eating in absence of hunger (EAH), and pre-meal biological hunger. The sensor only dataset (n=41,700 meals, 109 overeating) resulted in an AUC score of 0.732, and the 5 most predictive features were meal duration, # of bites, inter-chew interval, and eating speed. The combined sensor+EMA dataset resulted in an AUC score of 0.826, and the 5 most predictive features were # of bites, pre-meal biological hunger, food source, post-meal loss of control (LOC), and pre-meal loneliness.
Conclusions: We report on a novel SenseWhy system that predicts overeating from EMA, passive sensor, and combined sensor+EMA datasets. Features identified across all datasets may guide future studies of the contexts and behavioral aspects of eating and overeating episodes. Results from our analysis of sensor-only data suggests the possibility of wearable systems capable of predicting and intervening on overeating episodes in real-time without any active input from the user.
Farzad Shahabi, Christopher Romano, Boyang Wei, Mahdi Pedram, Angela F. Pfammatter, Annie W. Lin, Tammy Stump, Nabil Alshurafa
Link: https://doi.org/10.1093/abm/kaae014 (page 248)
Year: 2024
*Citation and Meritorious Award Winner
Despite the known complexity of overeating, prior research largely examines single causal mechanism of overeating (e.g., stress, loss of control). With recent advancements in mobile and wearable computing, we can examine—in natural settings—multiple behavioral, environmental, and psychological factors and how they co-occur with each other. This allows us to identify behavioral phenotypes warranting different treatments. In our SenseWhy study, we use wearable sensing and EMA to record rich, multimodal data from thousands of eating episodes and identify phenotype clusters associated with overeating behaviors.
During waking hours, 48 participants wore a wearable camera, 2 passive sensing devices, and a food-tracking app for 2 weeks. We define overeating as any meal exceeding the 1 standard deviation of a participant’s mean calorie intake per meal. We employed a semi-supervised learning clustering pipeline by leveraging a deep neural net and UMAP nonlinear manifold learner to eliminate irrelevant features and improve visual interpretability. We then utilized K-means clustering followed by Z-score analysis on the separable clusters to pinpoint leading factors driving problematic eating behaviors. We applied these ML approaches separately to EMA-only data, passive-sensing only data, and the combined EMA and passive-sensing data.
Our EMA-only, passive-sensing only, and combined approach yielded six, three, and three distinct clusters within overeating meals. Within our most comprehensive analysis (i.e., EMA-only), out of 2246 recorded meals (369 (16.4%) overeating meals), we identified six distinct overeating phenotypes, each reported here with a descriptive name, cluster purity (%), count of overeating meals (n), and concise interpretation:
“Stress-driven Night Nibbling”: (48%, 27), eating while stressed or lonely at night
"Night Craving": (86%, 113), heightened biological hunger at night and opting for self-prepared meals
"Uncontrolled Pleasure Eating": (78%, 39), hedonic and loss-of-control eating during task-oriented distractions (e.g., working/studying)
"Nighttime Restaurant Reveling": (95%, 61), pleasurable social dining during nights
"Take-out Feasting": (100%, 71), indulgent take-out meals in social settings
"Unstressed Home Dining": (50%, 4), low-stress meals during task-oriented distractions (e.g., watching TV)
The identified phenotypes represent a unique approach to understanding problematic eating behavior in naturalistic settings, identifying specific combinations of multi-level factors that are likely to co-occur within overeating episodes. This methodology has the potential to support populations struggling with obesity by fostering a more comprehensive understanding of their eating patterns, paving the way for personalized treatment strategies and empowering them to better regulate their eating behaviors.
Boyang Wei, Renee C. Edwards, Yudong Zhang, Stephanie Krislov, Aditi Rangarajan, Peter Cummings, Mahdi Pedram, Darius Tandon, Lauren S. Wakschlag, Nabil Alshurafa
Link: https://doi.org/10.1093/abm/kaad011 (page 338)
Year: 2023
Background: Prenatal stress can contribute to adverse health and developmental outcomes for both mother and infant. Mothers and Babies (MB) is an evidence-based intervention based on cognitive behavioral therapy (CBT) and attachment theory that aims to reduce maternal stress during pregnancy. Traditional CBT has organized intervention delivery schedules. However, rigidity in the number and content of intervention sessions may not be maximally benef icial for all participants. Predicting later stress in the early intervention stages may help clinicians adjust the scope and intensity of the intervention to help clients manage mood and stress more effectively. In this study, we explored data from the early stages of the MB intervention period and used machine learning models to predict perceived stress during the post-intervention assessment.
Method: Pregnant women (N=99) were randomized into a control (N=51) or MB intervention group (N=48) at study enrollment, with average gestational age around 11.6 weeks.. Women in the MB group were offered 12 weekly one-on-one sessions focused on behavioral activation, social support, cognitive restructuring, and individualized “just-in-time” stress reduction content. In addition, ecological momentary assessments (EMAs) of perceived stress (PSS-4) were sent 4 times daily via text to participants in both groups. Participants completed the Perceived Stress Scale (PSS-10) at enrollment and post-intervention. Participant baseline demographics, the number of intervention sessions received, and PSS-4 scores up to weeks 2, 4, 6, and 8 were applied to five commonly used machine learning models for detecting post-intervention stress at each of the four timepoints. Results are averaged from 5-fold cross-validation for the control and MB intervention group.
Results: Random forest overall performed the best for predicting postintervention stress in both groups from weeks 2 through 8. The best performance was achieved at week 4 for the MB intervention group (F1: 0.85, Precision: 0.83, Recall: 0.87) and the control group (F1: 0.81, Precision: 0.81, Recall: 0.80). From our exploration of different combinations of predictors, we found that weekly PSS-4 score at week 2, 3 and 4, marital status and income-to-need ratio were the most predictive features.
Conclusions: Our findings show that late pregnancy stress can be predicted early and robustly in pregnancy by EMAs and demographic features. Future studies can further strengthen the findings by using large sample sizes and participants from more diverse backgrounds. Early prediction can help identify mothers at risk of high stress and enable clinicians to adjust the intensity and content of interventions before the emergence of high stress. This study also supports the utility of applying machine learning models to promote the perinatal well-being of pregnant women.