Speech Recognition-Based Dietary Assessment Tool for Older Adults: Validation and Usability Study

As populations age globally, accurate and feasible dietary assessment for older adults has become increasingly important. South Korea has already become an "aged society," with over 14.2% of its population being aged 65 years and older, and is projected to become one of the world's most rapidly super-aged societies by 2050, with more than 40% of its population in this age group. Similarly, the Asia-Pacific region is experiencing accelerated population aging, with 10 countries classified as "aging societies" (>7% aged ≥65 years), 5 as "aged societies" (>14%), and 11 as "super-aged societies" (>21%) in 2025. Despite the growing need for accurate dietary monitoring in this demographic, nutritional assessment remains challenging due to limitations of conventional methods, compounded by cognitive burden, functional decline, and low literacy. Although various technology-based solutions, including web-based, scanner-based, and mobile tools, have been introduced, challenges related to usability, accuracy, and cost remain unresolved.

This study aims to evaluate the validity and usability of a speech-based dietary assessment tool as a potential alternative to a traditional pen-and-paper food diary among older adults.

In a randomized crossover study, adults aged ≥65 years (n=18) completed 2 nonconsecutive 3-day dietary assessment periods using both a speech-recording (SR) method and a food diary (FD). Mean daily nutrient intakes were compared between methods, and agreement was examined using Bland-Altman plots. Usability was assessed after each method using the System Usability Scale (SUS).

Mean daily intakes estimated by SR and FD were similar in magnitude across most nutrients. The paired mean difference in energy intake (FD-SR) was 38.38 (95% CI -176.63 to 253.40) kcal. Across nutrients, mean differences were generally small, and most 95% CIs included zero, indicating limited evidence of a large systematic difference at the group level. The correlation coefficient between SR and FD ranged from 0.178 to 0.907 depending on the nutrient assessed, though CIs were often wide (eg, energy intake; Pearson r=0.52, 95% CI 0.05-0.81). Bland-Altman analyses for energy and macronutrients show mean differences close to zero, although substantial individual-level variability was observed. Cholesterol demonstrated greater dispersion and possible proportional bias at higher intake levels. Mean SUS scores were 66.25 for FD and 72.78 for SR, with a paired mean difference (FD-SR) of -6.53 (95% CI -11.75 to -1.30), indicating higher usability ratings for the SR method on average.

These findings suggest that in older adults, the speech recognition method-produced dietary intake estimates may be a feasible approach with broadly comparable nutrient intake levels to written food diaries and modestly higher usability ratings. Further work is needed in larger samples to compare approaches with increased precision.