Comparison of Energy Intake Determined by a Natural Spoken Language Application with 24-h Recall

Self-monitoring daily dietary intake is recommended for weight loss and weight loss maintenance. However, current online platforms and applications are often burdensome, which may limit use. We conducted a pilot study to evaluate the accuracy of a new application designed to self-monitor dietary intake using natural spoken language (COCO; The Conversational Calorie Counter). A total of 35 participants were enrolled in this pilot study. Participants were asked to record daily dietary intake using the COCO application for a period of at least five days. Two 24-hour dietary recalls were conducted during this time, between day three and day five, and served as the reference method for evaluating total energy intake (TEI; measured in kcal). Mean two-day energy intake was calculated for each assessment method for the days when the 24-hr recall and COCO data were collected. Self-reported TEI from COCO were compared to estimates obtained from the 24-hour dietary recalls by a paired samples t-test and a Pearson’s correlation coefficient. On average, participants consumed three meals a day and recorded six days of food intake days with COCO (range: 4 to 10 days). The mean TEI was not significantly different between the two methods (1902 ± 621 kcal by 24-hour dietary recall and 1988 ± 1033 kcal by COCO, P = 0.59). There was a significant correlation between mean TEI measured with the two methods (r = 0.45; P = 0.006). In addition, a strong correlation was observed between the number of food items logged in COCO and those recalled in the 24-hour diet recalls (r = 0.82; P >0.0001). Completion of the exit survey by 28 participants indicated that 43% would definitely or probably use the application again. These results suggest that natural spoken language technology may have utility in applications to self-monitor food intake. Additional research is required to fully elucidate the validity of COCO in estimating dietary intake. This research was supported by the NIH Grant # 1R21HL118347–01 (SBR and JG), Quanta Computing, Inc., and the National Defense Science and Engineering Graduate fellowship.