**Background:** Over 42.4% of US adults have obesity, and frequent overeating is a key risk factor. Prior research often focuses on single determinants (e.g., stress, emotion) and relies on subjective or non-personalized measures. Wearable sensing advances now allow detection of complex feature patterns characterizing excess calorie intake. The SenseWhy project aims to identify automatically detectable features that predict overeating and to cluster eating episodes into meaningful phenotypes (e.g., emotional eating, mindless eating, night eating) using a multi-sensor passive sensing system combined with ecological momentary assessments (EMAs) and 24-hour dietary recalls.
**Methods:** Up to 60 adults with obesity (BMI ≥ 30 kg/m², ages 18–65) in the Chicagoland area will be recruited for a 14-day free-living observational study. Exclusion criteria include current dieting for weight loss, weight loss ≥15 lbs in prior 3 months, pregnancy, bariatric surgery, medications causing weight loss, genetic weight loss disorders, and active eating disorders or loss-of-control eating. Participants will wear three sensors: (1) NeckSense—a neck-worn device with an infrared proximity sensor (20 Hz) capturing mouth open/close, ambient light, lean-forward angle, and triaxial accelerometer; (2) a Fossil smartwatch on the dominant wrist recording triaxial accelerometer, gyroscope, and photoplethysmography (all at 20 Hz); and (3) an Infrared Activity-Oriented Device (IR-AOD)—a chest-mounted camera with thermal infrared, RGB fisheye lens (180°), and IR-LED for night vision. Participants will complete EMA surveys via the FoodTrck app for each eating episode: a pre-meal 'Decided To' survey (stress, affect, hunger type), a pre-meal 'About To' survey (photo and description), and a post-meal 'After' survey (photo of leftovers, cognitive restraint, uncontrolled eating, overeating, social context, location, activity co-occurrence). Daily 24-hour diet recalls using the Automated Multiple-Pass Method will be administered by trained dietitians starting on day 2, with NDSR software providing precise calorie and nutrient metrics. Overeating is defined as caloric consumption exceeding 1 standard deviation above an individual's mean consumption per eating episode (previously validated with Cohen's Kappa = 0.90 against subjective overeating). An alternative threshold of ≥1000 calories will also be tested. Power analysis using Monte Carlo simulation (ICC = 0.37 from prior study) with 60 participants, 3 eating episodes/day, and 30 episodes/participant achieves 80% power to detect within-subject variability of overeating (effect size 0.5 SD). Planned analyses include correlation-based and wrapper-based feature selection, machine learning classifiers (Gradient Boosting Machines, logistic regression, SVM, random forest, neural networks, Bayesian networks, hidden Markov models) with 60:20:20 train:validation:test split and 10-fold cross-validation, evaluated by ROC-AUC and F-measure. Clustering will use k-means and agglomerative hierarchical clustering with Ward's method, evaluated by purity, NMI, Rand Index, and F-measure.
**Key Results:** This is a study protocol paper; no results are reported. The study is described as ongoing with planned recruitment of 60 participants. Key methodological details include: overeating base rate assumption of 12% without stress exposure and 19% with stress exposure for power calculation; ICC of 0.01 from a prior 20-participant BeYourself study used in one power analysis; and a second power analysis using ICC = 0.37 from a different prior study. The personalized overeating definition showed strong agreement with subjective overeating (Cohen's Kappa = 0.90) in prior work.
**Clinical Implications:** This study will be the first to assess characteristics of eating episodes in situ over multiple weeks with visual confirmation of eating behaviors in a non-dieting population with obesity. By identifying automatically detectable features that predict overeating and clustering episodes into phenotypes (e.g., stress eating, mindless eating, night eating), the findings could enable development of personalized just-in-time adaptive interventions for overeating. The study addresses limitations of prior work by using objective sensing, reducing recall bias through EMAs, and studying a population with obesity (rather than primarily student samples). Limitations include inability to detect night eating (sensors removed during sleep), potential reactivity to measurement (mitigated by 2-week duration and personalized overeating definition), and the assumption that overeating episodes occur at similar rates across individuals. The personalized definition of overeating (1 SD above individual mean) does not account for expected unequal meal sizes throughout the day, which will require future study with larger samples.