Models of Individual Dietary Behavior Based on Smartphone Data: The Influence of Routine, Physical Activity, Emotion, and Food Environment.

INTRODUCTION: Smartphone applications (apps) facilitate the collection of data on multiple aspects of behavior that are useful for characterizing baseline patterns and for monitoring progress in interventions aimed at promoting healthier lifestyles. Individual-based models can be used to examine whether behavior, such as diet, corresponds to certain typological patterns. The objectives of this paper are to demonstrate individual-based modeling methods relevant to a person's eating behavior, and the value of such approach compared to typical regression models. METHOD: Using a mobile app, 2 weeks of physical activity and ecological momentary assessment (EMA) data, and 6 days of diet data were collected from 12 university students recruited from a university in Kunming, a rapidly developing city in southwest China. Phone GPS data were collected for the entire 2-week period, from which exposure to various food environments along each subject's activity space was determined. Physical activity was measured using phone accelerometry. Mobile phone EMA was used to assess self-reported emotion/feelings. The portion size of meals and food groups was determined from voice-annotated videos of meals. Individual-based regression models were used to characterize subjects as following one of 4 diet typologies: those with a routine portion sizes determined by time of day, those with portion sizes that balance physical activity (energy balance), those with portion sizes influenced by emotion, and those with portion sizes associated with food environments. RESULTS: Ample compliance with the phone-based behavioral assessment was observed for all participants. Across all individuals, 868 consumed food items were recorded, with fruits, grains and dairy foods dominating the portion sizes. On average, 218 hours of accelerometry and 35 EMA responses were recorded for each participant. For some subjects, the routine model was able to explain up to 47% of the variation in portion sizes, and the energy balance model was able to explain over 88% of the variation in portion sizes. Across all our subjects, the food environment was an important predictor of eating patterns. Generally, grouping all subjects into a pooled model performed worse than modeling each individual separately. CONCLUSION: A typological modeling approach was useful in understanding individual dietary behaviors in our cohort. This approach may be applicable to the study of other human behaviors, particularly those that collect repeated measures on individuals, and those involving smartphone-based behavioral measurement.

The Kunming CalFit study: modeling dietary behavioral patterns using smartphone data.

Human behavioral interventions aimed at improving health can benefit from objective wearable sensor data and mathematical models. Smartphone-based sensing is particularly practical for monitoring behavioral patterns because smartphones are fairly common, are carried by individuals throughout their daily lives, offer a variety of sensing modalities, and can facilitate various forms of user feedback for intervention studies. We describe our findings from a smartphone-based study, in which an Android-based application we developed called CalFit was used to collect information related to young adults' dietary behaviors. In addition to monitoring dietary patterns, we were interested in understanding contextual factors related to when and where an individual eats, as well as how their dietary intake relates to physical activity (which creates energy demand) and psychosocial stress. 12 participants were asked to use CalFit to record videos of their meals over two 1-week periods, which were translated into nutrient intake by trained dietitians. During this same period, triaxial accelerometry was used to assess each subject's energy expenditure, and GPS was used to record time-location patterns. Ecological momentary assessment was also used to prompt subjects to respond to questions on their phone about their psychological state. The GPS data were processed through a web service we developed called Foodscoremap that is based on the Google Places API to characterize food environments that subjects were exposed to, which may explain and influence dietary patterns. Furthermore, we describe a modeling framework that incorporates all of these information to dynamically infer behavioral patterns that may be used for future intervention studies.