This study was conducted at Kyung-Hee University from February 2024 to March 2024 after review and approval by the Institutional Review Boards of Kyung-Hee University (No. KHGRIB-24‐040). This study used a randomized crossover (AB/BA) design to compare the effectiveness of 2 dietary assessment methods: the pen-and-paper FD method and the SR method over a 2-week period (Figure 1). Participants were randomly assigned to one of 2 sequences (FD followed by SR, or SR followed by FD) using a computer-generated random sequence created in Microsoft Excel. Each method was used during a nonconsecutive 3-day recording period per week.

Eligible participants were older adults aged 65 years or older of both genders who owned smartphones and were able to complete 3 nonconsecutive days of dietary recording over a 2-week period. Because the SR tool required independent smartphone use, digital competency was considered an integral component of recording capability. Participants were therefore screened for basic digital literacy proficiencies using a tailored survey comprising 19 questions focused on digital familiarity, digital efficacy, and basic digital skills [15]. Inclusion criteria required scores ≥3 for digital familiarity and efficacy and ≥2 for basic digital skills, indicating sufficient digital competency for independent smartphone-based recording.

Participants diagnosed with mild cognitive impairment or dementia, or with significant visual or hearing impairments, were excluded. Individuals were also excluded if, based on researcher judgment, scheduling factors (eg, extended travel or irregular routines) were likely to interfere with accurate dietary recording.

Participants were recruited via posters, word-of-mouth referrals, and social networking platforms over a one-month period. General characteristics, including age, gender, education level, marital status, living status, comorbidities, and financial status, were collected using structured self-reported questionnaires administered with interviewer guidance. A total of 18 individuals were screened; all met the inclusion criteria, and none were excluded. Sample size was determined using G*Power (Heinrich Heine University Düsseldorf) for a 2-tailed paired t test (α=.05; power=0.95), assuming an effect size (Cohen d) of 1.04, based on effect sizes reported in previous dietary validation studies [13]. Under these assumptions, a minimum of 14 participants was required. To account for potential dropout, 18 participants were recruited and randomized to assessment order.

An effect size of 1.04 SDs indicates that the study was powered to detect relatively large differences between methods; for example, given the observed variability in energy intake (SD ≈400-500 kcal), this corresponds to a detectable difference of approximately 400‐500 kcal between FD and SR. Smaller differences would therefore be unlikely to be detected with this sample size.

All analyses were conducted using SPSS (version 29.0; IBM Corp), for macOS. Normality of data distribution was assessed using the Shapiro-Wilk test. Differences in mean nutrient intakes between the SR and FD methods were summarized using mean differences with 95% CIs. An independent t test or Mann-Whitney U test was used as a supplementary analysis, where appropriate, to aid interpretation. The sample size for this study was determined based on the primary aim of the intervention trial rather than for the purpose of method comparison. Consequently, the present validity analysis was conducted with a preexisting sample, and the study may have had limited statistical power to detect systematic differences between the SR and FD methods. Therefore, CIs including zero should not be interpreted as definitive evidence of absent bias, but rather as reflecting uncertainty due to limited power.

Pearson correlation coefficients were calculated to assess the linear association between absolute nutrient intake estimates derived from the 2 dietary assessment methods. 95% CIs for Pearson correlation coefficients were calculated using Fisher z transformation. Agreement between SR and FD was further evaluated using Bland-Altman plots to assess bias and limits of agreement across the range of intake and selected macronutrients, as these variables represent primary outcomes in dietary validation studies, exhibit broader intake distributions, and allow clearer evaluation of proportional bias across intake levels. Micronutrient intakes, in contrast, are characterized by substantial within-person variability relative to between-person variability [20]. When intake varies markedly within individuals across measurement occasions, this contributes to greater variability in the paired differences between methods, resulting in wide limits of agreement in Bland-Altman analysis. In this context, broad limits may largely reflect within-person variation rather than disagreement attributable to measurement methods [21]. Accordingly, Bland-Altman plots for micronutrients were considered less informative for evaluating systematic agreement in this context.

Continuous variables are presented as mean (SD), and categorical variables as counts and percentages.

The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Kyung Hee University (KHGIRB-24-040 on the 21st of February 2024). All participants were fully informed on the purpose and procedures of the study and provided written consent prior to their participation.

General characteristics of the study population. A total of 18 participants (13 women and 5 men) completed the study, each assessed for 2 nonconsecutive 3-day recording periods (Figure 1).

Participants’ ages ranged from 65 to 85 years, with a mean age of 72 (SD 4.89) years (men: mean 71.80, SD 1.79; women: mean 72.30, SD 5.72). The average BMI was 24.92 (SD 2.70) kg/m². Men had higher skeletal muscle mass (mean 28.8, SD 6.4 kg) than women (mean 20.9, SD 2.7 kg). Regarding education, 6 out of 18 (33%) had university-level education, with a higher proportion of men (60%) than women (23%). All participants were married, with most living with their partners (11/18, 61%). The primary source of living expenses was through salary or pensions of the individual or their partner (all men; 11/13, 84% of women). Most participants, 15 out of 18, felt secure in their ability to afford groceries (83%; Table 1).

The MNA scores indicated normal nutritional status for all participants, with a mean score of 27.06 (SD 2.27; men: mean 26.60, SD 2.27; women: mean 27.23, SD 2.34). The NQ-E score was 62.92 (SD 13.96), higher in male participants (mean 70.79, SD 10.68) than female participants (mean 59.89, SD 14.22). Male participants scored higher across NQ-E subscores in balance (mean 62.02, SD 13.79), moderation (mean 60.28, SD 13.61), and practice (mean 66.94, SD 11.75) compared to female participants (balance: mean 60.98, SD 18.15; moderation: mean 59.59, SD 17.81; practice: mean 55.54, SD 14.51).

An assessment was conducted to evaluate the accuracy and reliability of the SR tool used to transcribe speech recordings of participants, capturing details such as date, time, and meal type (eg, breakfast, lunch, dinner, and snacks), as well as the name of the dish and estimated portions using either standard household objects or food scales. The mean accuracy of transcription across participants was calculated to be 95.40% (SD 0.02%), indicating a high level of fidelity in transcribing and discerning dietary speech records. Each participant’s speech transcription accuracy was individually assessed for all 3 days of recorded speech, with each day’s transcription compared to manual transcriptions. This analysis process further reinforced confidence in the reliability of the SR tool for dietary intake assessment, and the findings substantiate the viability of using SR technology as a reliable means for capturing and transcribing detailed dietary intake data among older adults.

Table 2 summarizes the mean daily energy and nutrient intakes obtained from the FD and SR methods, together with the paired mean differences (FD-SR), 95% CIs, and the correlation coefficients. For most nutrients, the mean differences were small relative to the between-person variability, and the 95% CIs were wide and generally included zero. The CIs suggest that any systematic difference between methods, if present, is likely to be small at the group level.

For energy intake, the FD and SR methods yielded mean values of 1949.44 (SD 497.59) kcal and 1941.69 (SD 385.58) kcal, respectively, with a paired mean difference of 38.38 kcal (95% CI −176.63 to 253.40). For carbohydrate intake, the mean difference was 12.05 g (95% CI −26.41 to 50.52 g), while protein and fat showed mean differences of 1.04 g (95% CI −4.86 to 6.93) and 2.56 g (95% CI −4.98 to 10.10). Across micronutrients, mean differences were generally modest in magnitude, although the corresponding CIs were wide, reflecting variability across participants.

Overall, the observed differences are compatible with small overestimation or underestimation by SR relative to FD, rather than indicating a clear systematic shift in one direction. The direction of the paired difference varied across nutrients, with some nutrients showing higher mean intakes with FD and others with SR.

The Pearson correlation coefficient between the 2 methods ranged from 0.178 to 0.907, with corresponding 95% CIs indicating substantial uncertainty for weaker associations (Table 2). Correlations were moderate to high for several macronutrients, including fat (r=0.740), protein (r=0.738), and dietary fiber (r=0.671), as well as for selected micronutrients such as vitamin E (r=0.907) and riboflavin (r=0.871). For other nutrients, including carbohydrate, total sugar, cholesterol, and certain vitamins, correlations were weaker.

Bland-Altman plots for energy and macronutrients (Figure 2) were used to further examine agreement between FD and SR. For energy intake, the mean difference was 38.38 kcal, with 95% limits of agreement ranging from −863.39 to 878.89 kcal. For fat and protein, mean differences were small relative to the limits of agreement, which were wide. These findings indicate substantial individual-level variability in the differences between methods, consistent with the width of the CIs reported in Table 2. For cholesterol, the plots suggested a possible proportional bias, with larger discrepancies between FD and SR at higher average intakes; however, variability remained considerable across the intake range.

The SUS scores were obtained from all 18 participants after they had used both dietary assessment methods. The mean SUS score was 66.25 (SD 11.83) for the pen-and-paper FD method and mean 72.77 (SD 10.77) for the SR tool (0‐100 scale). The paired mean difference (FD-SR) was −6.53 (95% CI −11.75 to −1.30), indicating higher usability ratings for the SR method on average.

Responses to individual SUS items are summarized in Figures 3 and 4. Item-level results are presented descriptively as the proportion of participants selecting each response option. Overall, a higher proportion of participants agreed with positively worded statements for the SR method, whereas responses for the FD method more frequently reflected perceived difficulty or the need for support. No inferential statistical testing was performed for individual SUS items, as these results were intended to illustrate response patterns rather than formally compare item-level differences.

This study evaluated the relative validity and usability of an SR dietary assessment method compared with a traditional pen-and-paper FD among older adults. Overall, SR produced dietary intake estimates that were broadly comparable to FD at a group level. Paired mean differences for energy and most nutrients were small in absolute magnitude, and the corresponding 95% CIs generally included zero; however, this should not be interpreted as evidence of absent bias, as the wide intervals reflect substantial uncertainty and limited statistical power to detect systematic differences between methods. Correlation coefficients ranged from 0.178 to 0.907 depending on the nutrient assessed, with moderate associations observed for energy (r=0.518), protein (r=0.738), fat (r=0.740), and dietary fiber (r=0.671). Bland-Altman analyses show mean differences close to zero for energy and macronutrients, although limits of agreement were wide, indicating substantial individual-level variability. In terms of usability, SR received higher ratings than FD, with a paired mean difference in SUS scores of −6.53 (95% CI −11.75 to −1.30), indicating a modest but consistent improvement in user experience. Together, these findings suggest that SR may serve as a feasible alternative modality for dietary data entry in older adults, while acknowledging variability at the individual level.

While most previous validation studies have focused on web-based 24-hour recalls, online food frequency questionnaires, and artificial intelligence (AI)–based image recognition tools, this research evaluates SR as an alternative modality for dietary data input, with the potential to simplify data recording and reduce respondent burden [5].

To facilitate comparison with prior validation studies that report relative differences, absolute mean differences were also expressed as percentages of the FD mean intake. When calculated in this way, the relative differences for most macronutrients were below 5%, with the relative difference in energy intake below 1%. These values are comparable to those reported in previous systematic reviews of web-based versus conventional dietary assessment methods, which observed mean energy intake differences of approximately 5.3% (range 0.6%‐16.1%) [5]. Similarly, relative differences for protein and fat (1.84% and 0.32%, respectively) were smaller than those reported in the reviewed studies (5.00% and 7.73%, respectively) [22].

According to the classification proposed by Luevano-Contreras et al [23], relative mean differences of 0.0%‐10.9% are considered “good,” 11.0%‐20.0% as “acceptable,” and >20.0% as “poor.” Interpreted within this framework, most nutrients in the present study fell within the “good” agreement range. The direction of the paired differences varied across nutrients, and the magnitude of these differences was small relative to between-person variability, as reflected in the SDs and CIs.

The Bland-Altman analysis indicated mean differences close to zero for most nutrients; however, limits of agreement were wide, reflecting substantial individual-level variability. Carbohydrate intake showed larger dispersion despite modest mean differences, while fat and protein intake differences were small and consistent across intake levels. In contrast, cholesterol demonstrated proportional bias, with increasing divergence between the 2 methods at higher intake levels, reflected in a funnel-shaped distribution. The limits of agreement for cholesterol were also wider than those reported in previous studies, indicating greater variability between the 2 methods at higher consumption levels.

Correlations between FD and SR indicated moderate linear association for total energy intake (r=0.518), essential nutrients (r=0.478), vitamins (r=0.559), and minerals (r=0.589). To aid interpretation, correlation magnitudes were compared with commonly used descriptive thresholds [24]; however, given the small sample size, these values should be interpreted cautiously and viewed as indicative of relative association rather than definitive strength. CIs for several correlation coefficients were wide and, in some cases, crossed zero, reflecting limited precision and statistical uncertainty due to the modest sample size. Certain dietary parameters, including carbohydrates, total sugar, cholesterol, and vitamins A, D, C, and potassium, exhibited weaker correlations, likely reflecting the wide food sources of these nutrients and high day-to-day intake variability [25-28].

Previous studies have similarly reported low correlations for vitamins and minerals in dietary assessments. Kim et al [29] reported correlation of 0.18 for vitamin A and 0.48 for vitamin C using comparable dietary assessment approaches, which aligns with the present findings of −0.146 and 0.416, respectively. These findings highlight the inherent difficulty of capturing micronutrient intake over short recording periods due to underreporting and variability in food consumption [29]. Indeed, Kwon et al [30] estimated that 25‐29 days of dietary records may be required to estimate vitamin intake within 20% of true usual intake with 90% confidence, underscoring a known limitation of short-term dietary recording rather than a method-specific shortcoming.

Usability was assessed using the SUS following completion of both assessment methods. On average, the SR method scored 72.77 (SD 10.77), while the FD method scored 66.25 (SD 11.83). The paired mean difference (FD-SR) was −6.53 (95% CI −11.75 to −1.30), indicating higher usability ratings for SR on average. Although the magnitude of the difference was modest, it is consistent with a perceptible improvement in user experience. According to Bangor et al [31], SUS scores of 68 or above are generally interpreted as “good” and “acceptable” regarding usability, placing SR within this range, while FD fell slightly below this threshold. Previous research evaluating web-based dietary recall systems, such as the Automated Self-Administered 24-Hour Dietary Assessment Tool (ASA24) [32] and EatWellQ8 [33], a food frequency questionnaire system, and other mobile dietary intake applications [34], reported SUS scores ranging from 58 to 75 points, aligning with the SR method’s score.

Patterns of individual SUS item responses suggest that participants perceived SR as easier to learn and more convenient for routine use, whereas FD was more frequently associated with perceived complexity and the need for support. These findings likely reflect the burden of immediate documentation and the cognitive demands of estimating ingredients and quantities in traditional food records [35]. Nonetheless, some participants reported initial difficulty with the SR system, potentially due to unfamiliarity with digital tools, suggesting that training and system refinement remain important for wider implementation.

The randomized crossover design minimized sequence bias and strengthened internal validity by allowing each participant to serve as their own control [36]. Furthermore, the high transcription accuracy of the SR tool underscored its precision in transcribing dietary data, thereby supporting the technical feasibility of the SR system. However, dietary intakes were assessed over only 3 days per method, which limits the precision of micronutrient estimation and may amplify day-to-day intake variability [37,38].

In addition, the small sample size limits statistical power and means that only relatively large differences between methods could be detected. This also precluded subgroup analyses by demographic characteristics (eg, age, sex, education level, or digital literacy), which may influence dietary reporting behavior and usability perceptions. Future studies with larger and more diverse samples are needed to examine these factors in greater detail.

Importantly, this study evaluated relative validity against the traditional pen-and-paper FD method, which itself is subject to recall bias and reporting error. In the absence of an objective gold standard for habitual intake, the findings should be interpreted as evidence of comparability, not absolute accuracy [13]. This SR workflow also required it. Manual postprocessing following transcription, highlighting the need for future fully integrated conversational AI systems capable of real-time prompting and nutrient analysis.

This study evaluated the relative validity and usability of a novel speech-recognition-based dietary assessment tool compared with a traditional pen-and-paper FD among older adult participants. The findings indicate that SR produces dietary intake estimates that are broadly comparable to FD across most nutrients, with small mean differences, moderate correlations, and acceptable agreement in Bland-Altman analyses. Usability ratings were higher for the SR method, with a modest paired mean difference favoring SR.

Rather than demonstrating superiority, these results support the feasibility of SR as an alternative dietary data entry modality for older adults. With further refinement toward fully integrated, conversational dietary assessment systems, speech-based approaches may offer practical advantages for reducing user burden while maintaining acceptable data quality. Future studies should evaluate performance over longer recording periods and examine the capacity of SR-based systems to accurately identify nutritionally vulnerable individuals in real-world settings.

From an implementation perspective, the SR-based dietary assessment evaluated in this study can be applied in aging and aged societies using the available smartphone technology at the time of this writing. Because the method relies on natural speech input rather than written recording, it may reduce documentation burden and support dietary data collection among older adults with visual, literacy, or dexterity limitations. In practice, such an approach could be used for community-based nutrition monitoring, longitudinal dietary surveys, or remote dietary data collection in public health and research settings. With further development toward automated nutrient analysis and integrated conversational prompting, speech-based systems may complement existing dietary assessment tools while maintaining acceptable data quality. However, broader application will require validation in larger and more diverse older adult populations.