Predicting Food Intake from Food Reward and Biometric Responses to Food Cues in Adults with Normal Weight Using Machine Learning

Background Eating behaviors are determined by a complex interplay between behavioral and physiologic signaling occurring before, during, and after eating. Objectives The aim was to explore how selected behavioral and physiologic variables separately and grouped together predicted intake of 8 different foods. Methods One hundred adults with normal weight performed a food preference task combined with biometric measurements (the Steno Biometric Food Preference Task) in the fasting state. The task measured food reward as well as biometric (eye tracking, electrodermal activity, and facial expressions) responses to images of foods varying in fat content and taste. Energy intake from an ad libitum buffet of the same 8 foods as assessed in the preference task was subsequently assessed. A mixed-effects random forest approach was applied to explore how individual and combined measures of food reward and biometric responses predicted energy intake of the 8 single foods. The performance of the different prediction models was compared with the predictions from a linear model including only an intercept (naive model) using bootstrap cross-validation. Results Participants had a median [IQR] intake of 369 kJ [126-472 kJ] per food. Combined or separate measures of food reward or biometric responses did not predict energy intake better than the naive model. Conclusions We did not find that the reward or biometric responses to food cues assessed in a clinical setting were useful in predicting energy intake of single foods. However, this study provides a framework in the field of behavioral nutrition for applying machine learning with a focus on individual predictions. This is necessary on the road toward personalized nutrition and provides great potential for handling complex data with multiple variables. This trial was registered at clinicaltrials.gov as NCT03986619.