ابزار مداخله سیستم ضبط اتوماتیک وعده غذایی مبتنی بر کارت هوشمند برای تجزیه و تحلیل با استفاده از روش داده کاوی

The Smart-card-based Automatic Meal Record system for company cafeterias (AutoMealRecord system) was recently developed and used to monitor employee eating habits. The system could be a unique nutrition assessment tool for automatically monitoring the meal purchases of all employees, although it only focuses on company cafeterias and has never been validated. Before starting an interventional study, we tested the reliability of the data collected by the system using the data mining approach. The AutoMealRecord data were examined to determine if it could predict current obesity. All data used in this study (n = 899) were collected by a major electric company based in Tokyo, which has been operating the AutoMealRecord system for several years. We analyzed dietary patterns by principal component analysis using data from the system and extracted 5 major dietary patterns: healthy, traditional Japanese, Chinese, Japanese noodles, and pasta. The ability to predict current body mass index (BMI) with dietary preference was assessed with multiple linear regression analyses, and in the current study, BMI was positively correlated with male gender, preference for “Japanese noodles,” mean energy intake, protein content, and frequency of body measurement at a body measurement booth in the cafeteria. There was a negative correlation with age, dietary fiber, and lunchtime cafeteria use (R2 = 0.22). This regression model predicted “would-be obese” participants (BMI ≥ 23) with 68.8% accuracy by leave-one-out cross validation. This shows that there was sufficient predictability of BMI based on data from the AutoMealRecord System. We conclude that the AutoMealRecord system is valuable for further consideration as a health care intervention tool.

Obesity is a significant issue in Western countries [1]. Similarly in Japan, the prevalence of obesity in males has been increasing for the past 20 years and is now more than 30% for males in their 40s to 60s [2]. Moreover, a longitudinal analysis at the individual level demonstrated that the prevalence of obesity increased among middle-aged Japanese participants [3]. To prevent lifestyle-related diseases, especially visceral fat obesity, the Ministry of Health, Labor and Welfare, Tokyo, Japan, has issued an act that regulates health insurance unions. Under the act, these unions are to recommend an annual medical checkup for insured individuals between the ages of 40 and 75 and to conduct health guidance for those who are diagnosed with or at risk for metabolic syndrome [4]. Because most people at risk for metabolic syndrome are of working age, some companies have started to create environments that aid employees in improving their lifestyles [5], [6] and [7]. The company cafeteria plays an important role in the diet of employees and has come under the spotlight. Although there are several methods for assessing intake of foods/nutrients, including weighed diet records, 24-hour recall, and food frequency questionnaires, few companies have attempted to monitor eating habits of all employees because available methods require tremendous amounts of time, effort, and money. Recently, smart cards (pocket-sized cards with embedded integrated circuits that can process data) have become common as employee ID cards. An innovative system for company cafeterias has been developed to monitor employees' eating records using a smart card with an electronic wallet function. The smart-card-based Automatic Meal Record (AutoMealRecord system) relates point-of-sale purchase data to nutritional information per serving. For employees who registered, it also provides their nutritional records through email and its corresponding Web site [8]. The system also interfaces with body composition scales that have smart card readers. Employees can use their own dietary history and body composition records to improve their health. The AutoMealRecord system was originally developed by an electric company as a commercially available health care service to make the smart-card-based system pervasive. Because of their withdrawal from the health care business, the AutoMealRecord system has only been operated in-house and received less attention, even within the company, for several years. Although the AutoMealRecord system only targets meals from company cafeterias, it is a unique nutrition assessment tool for automatically monitoring the meal purchases and body composition of all employees. One of the authors (SZ) had a chance encounter with the implementer of the AutoMealRecord system and saw massive potential in the system to be a powerful tool for health promotion and lifestyle disease prevention. So far however, the system has not received any validation as a nutrition assessment tool as it has only been used for providing a weekly nutrition summary to registered employees. We decided to assess the potential of the AutoMealRecord system as a preventive measure against lifestyle diseases. Before starting an interventional study though, it must be tested whether the data accumulated by the AutoMealRecord system is reliable as a diet record. We applied the data mining approach, which is commonly used in a wide range of profiling practices such as marketing and surveillance, to extract important patterns from large amounts of data [9]. We hypothesized that the AutoMealRecord system could explain current obesity status if the data were reliable as a diet record, and so, we explored whether data previously collected by the AutoMealRecord system could predict current obesity in this study.

This exploratory data analysis has shown that the data accumulated in the AutoMealRecord system could explain, to some extent, current BMI. Although further evaluation will be required, we can assume that the AutoMealRecord system is reliable and valuable as a tool for health promotion and lifestyle disease prevention. We observed 5 major dietary patterns, but there was not a “Western” pattern. This makes it difficult to compare our results with other Western studies [20], [21] and [22]. The “healthy” and “traditional Japanese” patterns in this study were nearly identical to the healthy and Japanese traditional patterns, by Okubo et al [23]. The “healthy” pattern, which had a high loading of vegetables, fruits, fish, and soy products, was associated with lower BMI, whereas the “Japanese traditional” pattern, which had a high loading of rice, miso soup, and soy products, was associated with higher BMI among Japanese female students [23]. Although we could not see the direct correlation between these patterns and BMI, it is suggested that the “healthy” and “traditional Japanese” patterns are the major dietary patterns in a wide age range of Japanese people. In our study, BMI was correlated only with the “pasta” preference, calculated as annual proportions of the “pasta” pattern for each person. Higher “pasta” preference was associated with lower BMI (r = −0.15; P < .001). Those who preferred the “pasta” pattern tended to be female (r = 0.29; P < .001) and younger (r = −0.13; P < .001). These data indicate that the association between “pasta” and BMI was not a causal association but a self-selection bias. In Japan, pasta is regarded as fashionable and healthy food. This might explain why the “pasta” pattern includes grilled seafood, which is relatively lean rather than high-fat dish such as meat or fried dish. The multiple linear regression analysis illustrated that higher BMI was associated with higher frequency of the “Japanese noodles” pattern, which is likely due to the combination of Japanese noodles (soba or udon) with side dishes such as tofu, natto, or fruit cups. This association may be partly understood by the role of dietary glycemic index (GI) and glycemic load (GL). A positive association of dietary GI and GL with BMI among American and Japanese people has been reported [23], [24] and [25]. Udon (wheat noodles) are a major high GI/GL food in Japan and soba (buckwheat noodles) have a relatively low GI/GL. Furthermore, those who like to eat Japanese noodles may have little time to enjoy their meal because they can be quickly prepared and eaten. Eating quickly has been shown to be associated with being overweight in Japanese men and women [26]. Most Japanese regard Japanese noodles as “daddy” food as there are many stand-up-eating soba shops all over Japan filled with working-age men. Although we cannot know the type of noodles and speed of eating from the data, the association of BMI and preference of Japanese noodles seems to support previous studies. The obesity prediction model for all participants indicated that males and younger participants tended to have a higher BMI. This is consistent with the National Health and Nutrition Survey [2]. However, preference of dietary patterns was not equally distributed among sex and age-groups. An example of this is the “pasta” pattern mentioned above. Thus, we constructed BMI prediction models by sex and age-groups. More than half of the models contained the annual mean of total energy and frequency of body measurement as explained variables. Rolls et al [27] has shown that large food portion sizes leads to excess energy intake. The company cafeteria we observed served 2 portion sizes (normal and small) for almost every main dish, and the high energy intake was probably the result of choice as there was an indication that a tendency to choose high energy content foods was associated with higher BMI. Frequency of body measurement seems to reflect body consciousness. Among females in their 20s to 40s and males in their 30s, more frequent measurement was associated with higher BMI. It is possible that these participants had already been conscious of their body shape and kept assessing themselves. On the other hand, less frequent measurement was associated with higher BMI among males in their 50s. This group is the central target of the annual medical checkup act mentioned above. It would appear that body conscious men had tried to keep or improve their shape, although this point regarding body consciousness and satisfaction needs to be examined in more detail in future studies. The models for males and females in their 20s and females in their 40s and 50s showed relatively higher multiple regression coefficients. The more variables that were selected, the higher multiple regression coefficients tended to be, and so we did not think it was appropriate to use multiple regression coefficients to compare the models. We examined the model fit by AIC and BIC as these criteria are independent of the number of variables [12], [13], [14] and [15]. Compared with the whole model, AIC and BIC values were relatively smaller in the sex and age-group models, which means that these models were better than the whole model. The data fitness of models for males in their 30s and 40s was relatively lower, despite their frequent cafeteria use. It was suggested that lifestyle outside the cafeteria had a greater impact on these age-groups. On the model validation, we used BMI of 23 kg·m−2 or greater as the cutoff value instead of the Japanese standard obesity criteria, BMI of 25 kg·m−2 or greater [17]. We did this for 3 reasons. First, this was in accordance with a previous study on the AutoMealRecord system by Ishida [18]. That study showed that a dietary education program had more effect on people with a BMI between 23 and 25 kg·m−2. The result suggested that obese people more than 25 kg·m−2 could not easily change their dietary lifestyle and that preventive intervention for nonobese people would have a greater impact on public health. Secondly, in one study [3] founded on a longitudinal analysis of data from a population-based cohort study, it was reported that Japanese men and women aged 40 to 49 years tended to gain their weight for a period of 10 years. In the same study, nearly 10% of men aged 40 to 49 years who had a BMI of less than 25 kg·m−2 at baseline became obese (BMI ≥ 25 kg·m−2) during the 10-year period. Therefore, we decided to focus on prediction of “would-be obese” people with a BMI more than 23 kg·m−2 for model validation. The third reason we used BMI of 23 kg·m−2 or greater as the cutoff value is that our study participants were a relatively lean population compared to the general Japanese. Few obese people whose BMI were more than 25 kg·m−2 were in the female subgroups (n = 0-4), which might destabilize the prediction capability. The models tended to predict lower BMI, and “would-be obese” (BMI ≥ 23 kg·m−2) prediction accuracy was better among females and young males whose measured BMI was relatively low (prediction accuracy, 70.9%-100.0%). The models for males in their 30s to 50s were less well fitted, as indicated by AIC and BIC. We must consider lifestyles outside the company cafeteria to check their health status more precisely. It should be noted that our study has several limitations because of the nature of the exploratory data analysis. First of all, the data we used were not collected for this study. We only “mined” the data accumulated in the AutoMealRecord system database, which were used for the system users to check the feedback on their dietary record on a weekly basis. We could not make the dish classification or nutrient calculation. Second, the dataset included only the purchased data from a company cafeteria. We cannot know their whole dietary life and what the participants did or did not eat for certain. In addition, the participants were not randomly selected but were voluntary registrants from employees of a major electric company. They may have higher incomes, higher education, or better health consciousness. Although high education had no association with dietary consciousness in the study by Carrera [28], it is possible that our study participants were more health conscious or at least healthier because of their lower BMI than the national average in Japan. We must be cautious about generalization of the results. Also, we had no way of knowing the disease history of the participants and could not eliminate effects on dietary life from any previous or concurrent disease. The company we studied had employees who were at risk for lifestyle diseases according to dietary education programs using the AutoMealRecord system and personal comments from dietitians. The at-risk employees were identified by system ID. We believe that almost everyone who registered in September 2008 was healthy. To overcome those limitations, we are conducting a Web-based questionnaire survey asking socioeconomic status, lifestyle, disease history, and whole nutrition using a validated food frequency questionnaire. In addition, we used self-reported height and weight that are said to be subject to substantial measurement error [29]. However, Stommel et al [30] recently suggested that nonobese people are not likely to overreport their weight according to data from the continuous National Health and Nutrition Examination Survey. It is reported that 90.4% of people with “normal weight,” whose BMI was between 18.5 and 25 using physically measured weight, were classified correctly as “normal weight” using self-reported weight [30]. It was also suggested that high-income respondents tend to report their weight more accurately [31]. In our study, participants were relatively lean (85.0% [539/634] of them had BMI < 25) and most of them might be college educated according to our recent study on the AutoMealRecord system users (data not shown). It cannot be denied that self-reporting bias exists within our data, but we can assume it is minimally suppressed. Furthermore, our dietary pattern analysis has some limitations. We conducted a principal component analysis to reduce the number of purchase-related variables and make easy-to-understand dietary patterns. Principal component analysis itself is mathematically clear-cut as it does not depend on subjective decisions but is based on a simple mathematical rotation of axes. However, we made a subjective decision to use 5 major principal components for analytical simplicity and interpretability. An exploratory factor analysis with 5 factors had shown a similar pattern to the result of principal component analysis, and we confirmed the reliability and robustness of the dietary patterns beyond analytical methods. Beyond these limitations, it is notable that we could show the ability of the AutoMealRecord system to predict would-be obese people accurately by only using the data in the system. Although further evaluation will be required, we can assume that the data were reliable and valuable as tools for lifestyle disease prevention. We are now conducting a Web-based questionnaire survey and a messaging intervention using e-mail for the AutoMealRecord system users. Segmentation and message tailoring are the most important methods for effectively changing health behavior by health communications [32]. Previous dietary education for adults has paid little attention to age and sex but has focused more upon BMI or health consciousness. This study indicates that tailoring the health message, with consideration given to age and sex, is also important for improving the effectiveness of dietary education to prevent lifestyle diseases. In conclusion, we are justified in believing that the AutoMealRecord system is valuable for further consideration as a health care intervention tool by analyzing the data with data mining approach. With the spread of smart cards, the AutoMealRecord system could be a powerful infrastructure to maintain a healthy dietary lifestyle not only for company employees but also for the general public in the future.