Original Article
Quality assurance with an informatics auditing process for Food Composition
Tables
Chi-Ming Chu
a, Meei-Shyuan Lee
a,f,*
, Yi-Hsien Hsu
a, Hsiao-Li Yu
a, Tsai-Yi Wu
a,
Su-Chien Chang
a, Li-Ching Lyu
b, Fang-Ju Chou
c, Yun-Ping Shao
d, Mark L. Wahlqvist
a,e,fa
School of Public Health, National Defense Medical Center, 161 Minchuan East Road, Section 6, Taipei 114, Taiwan, ROC
b
Department of Human Development and Family Studies, National Taiwan Normal University, 162, HePing East Road, Section 1, Taipei 106, Taiwan, ROC
c
Department of Nutrition, Wan Haw Hospital, 6 Chunghaw Road, Section 2, Lane 606, Taipei 108, Taiwan, ROC
dDepartment of Food and Nutrition, Tri-Service General Hospital, National Defense Medical Center, 325 Chenggong Road, Section 2, Taipei 114, Taiwan, ROC e
Institute of Population Health Sciences, National Health Research Institutes, 35 Keyan Road, Zhunan Town, Miaoli County 350, Taiwan, ROC
f
Asia Pacific Health and Nutrition Centre, Monash Asia Institute, 8th Floor, South Wing Menzies Building (11), Monash University, Melbourne 3800, Australia
1. Introduction
Food Composition Tables (FCT) are used to translate data on the nutrients in food for use in labelling, nutrition education, dietary surveys, epidemiological studies, and dietary intervention trials (Bowers, 2002; Buzzard et al., 1990; Deharveng et al., 1999; Dwyer, 1994; Fidanza and Perriello, 2002; Guilland et al., 1993; Leclercq et al., 2001; Matsuda-Inoguchi et al., 2004; Sasaki et al., 1999; Shimbo et al., 1996; Vaask et al., 2004; Zhang et al., 1999). FCT comprise the most frequently used and representative foods in a region, and contain information on energy and generally dozens of nutrients in numerous foods. The quality of food composition information is dependent on the sampling of foods, the analytical procedures used and how data are presented. Errors can have a
profound effect on research findings and food judgements, especially where the information is designed to connect with public health and nutrition education initiatives. Of particular concern is the accuracy of nutrition information in food labelling, which depends on reliable food compositional data (Fabiansson, 2006).
The nature and derivation of nutrient values in food items is often misunderstood, even among specialists, with consequent error in reporting nutrient intakes. For instance, in the Taiwanese FCT, mung bean sprouts have the highest vitamin C per 100 g dry weight among vegetables; yet soybean and mung bean sprouts contain 13 mg and 184 mg vitamin C per 100 g dry weight, respectively (Taiwan Department of Health, 2002c). The difference is 171 mg or 14-fold. Such discordant values between two apparently similar foods create difficulties in the accurate estimation of individual or community nutrient intakes.
Although such discrepancies are clearly recognized, it is not feasible to inspect the quality of FCT data using traditional laboratory methods alone; they are too expensive. This is especially true when assays must be repeated or when this applies to more than 20,000 data points (i.e. assuming 20 nutrients per A R T I C L E I N F O
Article history: Received 7 August 2008
Received in revised form 22 January 2009 Accepted 2 March 2009
Keywords: Quality assurance Auditing process Food composition database Coefficient of variation Food subgroups Compositional similarity Food variety Taiwan Food composition A B S T R A C T
A 6-step auditing process was developed to detect unlikely nutrient values in a Nutrient Composition Data Bank for Foods (NCDBF) in Taiwan. Preference was given to finding errors in the database, rather than to determining significant differences in the biological characteristics of the individual nutrients. There were 239 compositionally similar subgroups categorized within the NCDBF. The coefficient of variation (CV) of nutrient values for each subgroup provided the first-order sorting instrument. Nutrient CVs were ranked in rows for food subgroup (x) and in columns for nutrient type (y) and their product (x,y) in descending order. When the rank was in the top 2 or the product was 20, the Excel ‘‘cell’’ was regarded as a ‘‘hit’’. The ‘‘hit rate’’ (2.6%, 777 hits/29,424 pieces of information) of the computerized analysis was verified through an expert panel review to provide a ‘‘satisfied hit rate (SHR)’’ (agreed errors/ total food group hits). The mean SHR was 14.9% (range: 1.4%–37.6%) for the various food groups. The computerized process performed with a 38-fold increase in likelihood of error detection compared with what manual assessment alone would have produced. This low-cost approach could be applied in various jurisdictions or with other digitized food composition tables.
ß2009 Elsevier Inc. All rights reserved.
* Corresponding author at: School of Public Health, National Defense Medical Center, 161 Minchuan East Road, Section 6, Taipei 114, Taiwan, ROC.
Tel.: +886 2 87910704; fax: +886 2 87910704. E-mail address:[email protected](M.-S. Lee).
Contents lists available atScienceDirect
Journal of Food Composition and Analysis
j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / j f c a
0889-1575/$ – see front matter ß 2009 Elsevier Inc. All rights reserved. doi:10.1016/j.jfca.2009.03.005
1000 food items). A simple and low-cost auditing process for identifying unlikely nutrient values in foods in FCT is needed so as to enable experts to review the values and suggest either more selective revision, or recommend that another laboratory analysis be performed. Such an auditing process would represent sig-nificant savings in time, costs and human resources; it would require computer power which would remove limitation by way of the size of datasets, a situation which might be encountered in any future FCT.
Computers and information technology have become tools to improve efficiency and quality in every profession. Nutritional values can be evaluated through an automated process (Alexander et al., 1993; Dartois et al., 1989; Probst et al., 2004; Samsonov et al., 1986) since informatics applies to the work of nutritionists in public health and individual health (Mccullough et al., 1999; Phillips et al., 2006; Rodriguez et al., 1995; Sharp and Ahmed, 1983; Van Wave and Decker, 2003). A robust approach can efficiently scrutinize nutrient values of food items in several datasets (NATA, 1996; ISO/IEC, 1997a; ISO/IEC, 1997b; Puwastien, 2002). Other approaches include the recognition of extreme medians and normalized inter-quartile ranges by use of the z-score (Institute for Interlaboratory Studies, 2006), where data with z-scores outside of the range of 3 were not acceptable. The International Network of Food Data System (INFOODS) (Rand, 1985; Rand, 1987; Rand et al., 1991; Rand and Young, 1984) and the Association of South East Asian Nations Network of Food Data
Systems (ASEANFOODS) (Puwastien, 2002) use this approach to
approve pooled datasets of FCT, as have the Greenfield and Southgate (Southgate and Greenfield, 1992) guidelines.
Nutritionally similar foods should have similar nutrient values. However, foods may be grouped at different levels, such as fruits then citrus, vegetables then dark green vegetables. For educational purposes, foods are not introduced to the public singly, but usually categorized by their biological similarity, e.g. citrus fruits in the
fruits group and dark green vegetable in the vegetables group are rich in vitamin C and carotene, respectively. Each of the food subgroups, citrus fruits or dark green vegetables, would be expected to be more nutritionally similar than their parent groups, fruits and vegetables, in accord with their known food biology whose features inform dietary guidelines and food guides. Although the conventional manual approach to the evaluation of FCT quality might discover extreme nutrient variation in a food group, it may neglect unlikely nutrient values for a subgroup of similar food items within it. Neglected values are sources of potential error in FCT, and cannot be identified by the z-score for a food group as a whole.
There is a demand for reliable food composition information that can be validated through computerized dietary system tools (Farran-Codina et al., 1994; Shai et al., 2003). We propose an informatics process for data quality audit to identify unlikely nutrient values. The underlying premise is that within a food subgroup, characterized by biological similarity of its food items as basic commodities or by way of recipe or food process involved in preparation, an outlier in nutrient composition might be an error which merits revision.
2. Materials and methods
2.1. Nutrient composition data bank for foods in the Taiwan area The Nutrient Composition Data Bank for Foods (NCDBF) in the Taiwan Area consists of 24 nutrients (including energy and water) in 1226 foods of 18 groups like CEREALS (119 items), POTATOES & STARCHES (17 items), and NUTS & SEEDS (39 items). The number of food items for each food group is shown in the second column of
Table 1(Taiwan Department of Health, 2002a; Taiwan Department of Health, 2002b).
Table 1
Audit of Taiwan Food Composition Tables by statistical ‘‘hits’’ and an Expert Panel to identify unlikely nutrient values in 18 food groups.
Food Group Computerized process Panel process SHR(%)a
All food items Unprocessed food items Sum of
hits
HR/ PER (%)b
Panel found missing or was satisfied Without MH With MH Itemsc Subgroupsd Itemse Subgroupsf MHg SHh CEREALS 119 23 16 6 49 1.72 0 1 2.04 2.04
POTATOES & STARCHES 17 5 11 4 – – – –
NUTS & SEEDS 39 11 25 10 57 6.09 0 11 19.3 19.3
FRUITS 101 22 70 21 72 2.97 0 23 31.9 31.9 VEGETABLES 138 29 67 28 82 2.48 6 27 32.9 30.7 ALGAE 5 1 5 0 – – – – MUSHROOMS 17 7 11 6 – – – – PULSES 69 13 45 12 68 4.11 0 5 7.35 7.35 MEATS 101 25 88 24 89 3.67 4 35 39.3 37.6
FISH & SEAFOODS 103 18 66 17 – – – –
EGGS 31 7 20 6 47 6.32 5 4 8.51 7.69
MILKS 66 13 41 12 66 4.17 1 6 9.09 8.96
FATS & OILS 42 13 40 12 – – – –
SUGARS & SWEETENERS 10 2 2 1 – – – –
BEVERAGES 86 7 15 2 – – – –
SEASONINGS & SPICES 82 12 38 11 70 3.56 1 1 1.43 1.41
CONFECTIONERIES 64 9 0 0 69 4.49 1 1 1.45 1.43
PROCESSED FOODS 136 22 0 0 90 2.76 0 2 2.22 2.22
Total 1226 239 560 172 759 2.58 18 116 15.3 14.9
a
SHR: satisfied hit rate = (satisfied hits/sum of hits) 100%.
b
HR (or Probable error rate, PER): hit rate = (sum of hits/total piece of information) 100%, where total pieces of information are the number of foods multiplied by the number of nutrients, which is 24 in the NCDBF.
cNumber of foods in each food group in all foods.
d Number of subgroups plus one unsorted subgroup in all foods. e
Number of items of unprocessed foods (non-treated or non-prepared).
f
Number of subgroups with items of unprocessed foods.
g
MH: missing hit, unlikely values were identified by the Panel, but not by the computerized process.
h
2.2. Terminologies, abbreviations and algorithms or formulas (1) Coefficient of variation, CV: Statistical measure of dispersion in
a data series around the mean. It is defined as the ratio of the standard deviation to the mean. In this process, the population
standard deviation (
s
) is calculated because there is nosampling. (CV =
s
/mean).(2) CV RANK: Order of descending CVs by nutrients and food subgroups, e.g. it is 1 for the largest CV. (x or y for subgroup or nutrient, respectively).
(3) CV RANK Product, CRP: Product of CV RANK (x,y). (CRP = x y). (4) Outlier: Individual value which is far away from mean, i.e.
mean 1.5 SD (standard deviation) and mean 3.0 SD. (5) Hit: Computerized, process-defined unlikely value.
(6) Hit rate, HR: Hits per total pieces of information in each food group or in each dataset which is a ‘‘probable error rate’’ (PER). HR = (sum of hits/total piece of information) 100%, where total piece(s) of information is the number of foods multiplied by the number of nutrients, namely 24 in the NCDBF.
(7) Missing hits, MH: unlikely values were identified by the Panel, but not by the computerized process.
(8) Satisfied hit, SH: system-defined hit with which the Panel agrees
(9) Satisfied hit rate, SHR: the proportion of Satisfied hits to total hits which is calculated as (Satisfied hits/Sum of hits) 100% 2.3. The audit process
We developed an audit process to identify unlikely nutrient values in food items using the following steps (Fig. 1):
(1) Conversion of nutrient content to a dry weight (DW) basis, i.e. mass of nutrient per 100 g dry food: to adjust for the impact of water on nutrient content and deal with a largely irrelevant source of variability. Thus [original amount of a nutrient per 100 g fresh wt/(100-amount of water in 100 g fresh wt) 100] was used for nutrient content per 100 g dry food.
(2) Assignment of food items to a subgroup: to increase the degree of similarity, particularly for processed food, firstly whether it was ‘‘processed’’ or ‘‘unprocessed’’ and, then the method of
preparation such as pickled, dried or canned to constitute various subgroups. (SeeAppendix Afor examples.)
(3) Calculation and ranking of CVs: to calculate the CVs of nutrient values of food items in one subgroup and to rank CVs, in descending order, by nutrients and food subgroup, respec-tively. (Hereafter referred to as ‘‘CV RANK’’)
(4) Identification of unlikely values: to find outliers by a filter which uses the subgroup nutrient CVs, being the top 2 ranks in each subgroup for either the subgroup (x) or its nutrient (y); or the product of the CV RANKs x and y as a vector (x,y), defined as CRP, with a value less than 20. Each separate vector is referred to as a ‘‘cell’’, which provides the basis for a ‘‘hit’’ (refer toTable 2and Section3.4for the detail of the filter).
(5) Judgment by a food and nutrition expert panel: to verify the findings and to ascertain whether the filtering method missed unlikely values. Experts who had formal training in the food and nutrition sciences and had achieved professional distinc-tion were invited.
(6) Recommendation for further work: to accept, to re-analyse or to revise the nutrient of a given food as required.
Fig. 1. Six-step informatics audit process.
Table 2
Empirical leverage of food (x) or nutrient (y) CV RANKs and their product (x,y) for the filter of the VEGETABLES group. CV RANKs (x or y)a
CV RANK Product (x,y)b
0 5 10 15 20 25 30 35 40 0 SHc 0 8 18 23 26 27 27 27 27 SHR (%)d – 32 36 34.3 32.5 29.7 27 24.8 22.1 1 SH 14 14 19 23 26 27 27 27 27 SHR (%) 37.8 35.0 36.5 33.8 32.1 29.7 27.0 24.8 22.1 2 SH 21 21 21 24 27e 27 27 27 27 SHR (%) 32.8 32.8 32.8 30.8 30.7 28.4 26.5 24.8 22.1 3 SH 24 24 24 24 27 27 27 27 27 SHR (%) 25.3 25.3 25.3 25.3 26.5 25.7 25.0 23.9 21.6 4 SH 27 27 27 27 27 27 27 27 27 SHR (%) 22.3 22.3 22.3 22.3 22.3 22.1 22.0 21.8 20.6 5 SH 28 28 28 28 28 28 28 28 28 SHR (%) 19.6 19.6 19.6 19.6 19.6 19.6 19.6 19.6 19.2
aThe descending rank of coefficient of variation (CV) of either the food (x) or its nutrient (y). b
The product of the rank of CVs of (x) and (y).
c
SH: satisfied hits, system-defined hits which satisfied Expert Panel as legitimate errors worthy of further evaluation.
d
SHR: satisfied hit rate = (satisfied hits/sum of hits) 100%.
e
3. Results
We demonstrate the nature and detail of this six-step audit process by using the NCDBF.
3.1. Conversion of nutrient content to a dry weight basis
Nutrients in foods were compiled into individual Excel work-sheet by groups. The example of the VEGETABLES group is shown in Fig. 2. The columns present the codes of food in ‘‘A’’, the subgroup of food items classified as similar in ‘‘B’’, whether processed or not in ‘‘C’’, the name of food in ‘‘D’’, the energy in ‘‘E’’, the water in ‘‘F’’, and the contents of 22 nutrients from ‘‘G’’ (protein) to ‘‘AA’’ (zinc). The nutrient contents as DW were derived to take water content into account.
3.2. Assignment of food items to a subgroup
The 1226 foods from 18 groups were categorized into 239 subgroups based on their biological similarity by two nutritionists (Lee and Wu), as basic commodities or by way of recipe or food process involved in preparation. The Appendix to this paper provides the detailed food items of each VEGETABLES subgroup.
Fig. 2demonstrates the example of the VEGETABLES group. The cells ‘‘D130’’ and ‘‘D131’’ indicate frozen cabbage and dried cabbage are similar food items sorted into the 4th subgroup in cells
of ‘‘B130’’ and ‘‘B131’’; besides, they belong to processed methods of items coded as 1, which means processed food, in ‘‘C130’’ and ‘‘C131’’. There are 29 subgroups in column B coded as 0 (unsorted), 1 (carrots), 2 (radishes), . . .., 28 (amaranths) and so on.
3.3. Calculation and ranking of CVs
The CVs of subgroups, except the unsorted ones, were calculated for both non-processed and processed foods (Fig. 3). Where all nutrient values are missing or undetectable they have been treated as zero for CVs.
3.4. Identification of unlikely values
The filter chosen comprised the combination of both the ‘‘CV RANK’’ of food and nutrient and also their ‘‘product’’ which leverages the ‘‘hit’’ and ‘‘hit rate’’ (Table 2). There were two possible approaches to set up the cut-offs for filter for the computerized process. The first approach was to choose a filter which could minimize both ‘‘CV RANKs’’ and their product. In the present exercise, we have adopted this method and found the most unlikely values for food varieties and only needed to review the optimized total hits for food groups and
subgroups. InTable 2, among VEGETABLES, we observe that the
2 numbers for most SHs are 27 and 28; and that the ‘‘SHR’’ can reach more than 30%. We then deduce that the optimal
Fig. 2. The Excel worksheet for the VEGETABLES group in the Taiwan Food Composition Tables. The columns are food item code (A), food subgroup (B), whether processed (0: unprocessed, 1: process): (C), food item name (D) and various nutrient values (E–X). All nutrients are expressed per 100 g dry weight. Where nutrient values are missing or un-detectable they are represented by ‘‘0’’. All Chinese characters on the top of the frame are the menu (functional list) of Microsoft Excel.
parameters of ‘‘rank’’ and ‘‘product’’ for the filters are 2 and 20, respectively, inTable 2.
The second approach was to define a certain ‘‘SHR’’ based on resources and capacity to prioritize unlikely values. To choose a smaller ‘‘CV RANK’’ and CRP would result higher ‘‘SHR’’, which would mean more unlikely values, and vice versa.
As shown inFig. 3, we first ranked the food subgroups CVs for each nutrient in descending order (from the 35th row and below). Then, we ranked the nutrient CVs for each food subgroup (from the column ‘‘AE’’ and beyond). Two cut-offs for outliers have been defined as mean 1.5 and 3.0 times the standard deviation calculated, respectively (33rd and 34th row for nutrients and ‘‘AC’’ and ‘‘AD’’ for subgroups). InFig. 4, each ‘‘hit’’ is shown as the two components of a vector (x,y) in parenthesis, with the food subgroup CV RANK as (x) and the nutrient type CV RANK as (y), and the product (CRP) outside the parenthesis. Accordingly, the worksheet ‘‘hit rate’’ is represented as a vector (two CV RANKs, one for food subgroup and one for nutrient, respectively, multiplied). The numeric product of the elements of the vector is shown as in ‘‘E17’’ by ‘‘(1,6) = 6’’.
For the VEGETABLES group, a total of 82 hits were screened by the computerized process. A hit rate (HR) was calculated as (sum of hits/total piece of information) 100%, where total pieces of information are the numbers of foods multiplied by the number of nutrients. The HR for the VEGETABLES group is 2.48%, which is calculated as [82 hits/(138 items 24 nutrients)] 100%. This
computer process can only assess a food subgroup which comprises two or more foods. Some performance parameters have been shown in the collective food group worksheets for ‘‘hits’’, as inFig. 4for the VEGETABLES group.Table 1lists sum of hits and HR for all 18 food groups.
3.5. Judgment by a food and nutrition expert panel
The Expert Panel was able to review each hit of a screened subgroup or nutrient for unlikely values. Suggestions made for further investigation were ‘‘recheck in laboratory’’, ‘‘definitely describes food item/species’’, ‘‘too high/low nutrient content’’ and so on.
‘‘Satisfied hits’’ (SH) were system-defined hits which the Expert Panel agreed were legitimate errors worthy of further evaluation, as highlighted inFig. 4. The satisfied hit rate (SHR) was calculated as (Satisfied hits/sum of hits) 100%. ‘‘Missing hits’’ (MH) were those identified by the Expert Panel to avoid false negatives generated by the third and fourth steps, but not by the computerized process. In this example, the cell ‘‘C31’’ inFig. 4displayed the sum of hits as 82, cell ‘‘C32’’ displayed SH as 27, cell ‘‘C34’’ displayed the few MH (e.g. ‘‘****’’ on cell I22) as 6, and cell ‘‘C33’’ displayed the SHR without MH as 32.93%.
In all cases the Expert Panel took into account FCT from not only
Taiwan, but also China (China Institute of Nutrition and Food
Fig. 3. The coefficients of variation (CVs) for the subgroups of the VEGETABLES group in the Taiwan Food Composition Tables. For the purpose of overview, this worksheet has omitted columns, such as D, I, J, etc... Values in cells E3 to AA30 are the CVs of subgroups. ‘‘*’’ means non-calculable CV. All CVs were ranked in descending order for food subgroups (from the column ‘‘AE’’ and beyond) and for each nutrient (from the 35th row and below). Two cut-offs for outliers have been defined as mean +/- 1.5 and 3.0 times standard deviation calculated (33rd and 34th row for nutrients and ‘‘AC’’ and ‘‘AD’’ for subgroup). ‘‘#NUM!’’ indicates ‘‘un-rankable’’. Highlight in this worksheet are for Panel discussion. All Chinese characters on the top of the frame are the menu (functional list) of Microsoft Excel.
Safety, 2002) and Japan (Kagawa, 1999) in their review and evaluation of ‘‘hits’’ (see below).
3.6. Methodological gains
The audit process screened for potentially unlikely values in 18 food groups, by way of 239 subgroups of 1226 food items and 24 nutrients in the Taiwan FCT and the Expert Panel reviewed the values to determine likely validity or need for revision.Table 1is a summary of the performance of the process. This system detected 777 (759 from the computerized process and 18 from the Panel process) unlikely nutrient values in the FCT (comprising 29,424 (1226 items 24 nutrients) pieces of information). For this FCT, the HR (or probable error rate, PER) from the auditing process was 2.6% (777/29,424). The process alone performed nearly 38 times (29424/777) better than the manual only method. Even though there were 18 missing nutrient errors identified by the Panel, and not by the computerized process, they were readily identifiable in the subgroups screened by changing the CRP from 20 to 30 (Table 1). However, in this case, the HR would decrease due to a false positive as indicated inTable 2.
The Expert Panel reviewed values in relation to the FCT of Taiwan, China and Japan and recommendations were made to regard 116 items as genuinely unlikely (SH) and to require further laboratory analysis, so that the FCT could be revised. The overall SHR was 14.9% (ranged 1.4–37.6) for the various food groups. 3.7. Accuracy and applicability of the method
There were high SHR in several food groups, 19.3%, 32.0%, 30.7%, and 37.6% in the NUTS & SEEDS, FRUITS, VEGETABLES, and MEATS groups, respectively (Table 1). In food groups with many processed food items, few unlikely nutrient values were assigned by the Panel; these were the CEREALS, PULSES, EGGS, MILKS, SEASON-INGS & SPICES, CONFECTIONERIES, and PROCESSED FOODS groups with SHR ranging from 1.4% to 9.0%.
Some food items, mostly processed and treated, were not well subgrouped for reliable assessment of variation: they had no hit in this system. These were the POTATOES & STARCHES, ALGAE, MUSHROOMS, FISH AND SEAFOOD, FATS & OILS, SUGARS &
SWEETENERS, and BEVERAGES groups. For example, many shrimps and fish items were not generic and could not be subgrouped reliably. This also applied to foods that had been used to prepare other items as with certain fats, oils, sugars, sweeteners, potatoes, and starches. Additionally, foods of low nutrient density like FATS & OILS and SUGARS & SWEETENERS have few nutrients to detect, except for fat-soluble vitamins A and E in FATS & OILS; there were no CVs or vectors to calculate or assess. For ALGAE, there were only 5 items and too few to create subgroups. For MUSHROOMS, even though there were 6 subgroups, there were too few food items to allocate subgroups for reliable variation.
3.8. Methodological limitations
The hits which satisfied the Panel as legitimate errors were greater when there were enough food items that could be correctly assigned to a subgroup, especially when there were ‘‘unprocessed foods’’. This data audit performed better for unprocessed foods, which could be confidently subgrouped.
In terms of number of food items needed in a subgroup, theoretically, a population CV can be calculated with an n of 2. Although the meaning and utility of the CVs from small subgroups may be questioned, we have demonstrated that this approach still yields useful information. For example, in the bean sprouts subgroup, with 2 items only, we have identified that mung bean sprouts, but not soy bean sprouts, have an unlikely high vitamin C content.
For calculating CVs, we have had to treat all missing information as undetectable values and zero. This system cannot distinguish between true zero and missing information. Foods of low nutrient density (e.g. SUGAR) have too few nutrients to calculate CVs. 3.9. External validity
To exemplify, the FRUITS in the Taiwan FCT for the subgroups
GRAPES and WATERMELONS are shown in Table 3. They are
juxtapositioned alongside those from the Chinese and Japanese FCT (China Institute of Nutrition and Food Safety, 2002; Kagawa, 1999) to assist in the recognition of unlikely values. In the GRAPES subgroup of the Taiwan FCT, there were 3 items (California grape (D019400), White grape (D020400) and Grape (D021400)) which Fig. 4. The ‘‘satisfied hit rate’’ (SHR) derived from food subgroups and nutrient type for the VEGETABLES group of the Taiwan Food Composition Tables. For the purpose of overview, this worksheet has omitted columns J, K, and L. Satisfied hits (Row 32) were system-defined hits which the Expert Panel agreed were legitimate errors worthy of further evaluation but without missing hits. SHR (Row 33) is the satisfied hit rate (C33 = C32/C31) without missing hits. The Panel detected unlikely values which are shown as ‘‘****’’. Highlights in this worksheet are Panel agreed. All Chinese characters on the top of the frame are the menu (functional list) of Microsoft Excel.
the Panel regarded as ‘‘too variable in vitamin A potency’’ when considered in conjunction with similar food items for China and Japan (China Institute of Nutrition and Food Safety, 2002; Kagawa, 1999). It was also concluded that: ‘‘The analytical laboratory should indicate explicitly what cultivar the D019400 is, by its scientific name (Vitis spp. L.) as in the Japanese FCT (Japan Science and Technology Agency, 2005) or by description as in the Encyclopaedia Britannica and Wikipedia (Wikipedia, 2006).’’
Since the D020400 (White grape) contained 1410 RE vitamin A activity per 100 g DW, markedly greater than the ones (0, 22.9, 38.5, 70.8, 76.9, and 86.2 RE) from the China FCT and the one (32.1 RE) from the Japan FCT. The vitamin A activity of the subgroup GRAPES was unlikely.
Likewise, in the WATERMELONS subgroup (D027400, Water-melon) contained 1810 RE vitamin A activity per 100 g DW that was much greater than the ones (148, 494, 530, and 1119 RE) from the China FCT and the one (707 RE) from the Japan FCT. This meant that vitamin A activity in the WATERMELONS subgroup was unlikely. 4. Discussion
We present an informatics process to identify extreme nutrient values for foods in the Taiwan FCT from a data audit of coefficients of variation (CVs). These included minerals among NUTS & SEEDS, vitamin A among FRUITS, crude fat and carbohydrate and vitamins among VEGETABLES, and all nutrients among MEATS. This is in close agreement with the experience of nutritionists and suggests that there are needs for laboratory re-checks and revision of various food composition tables.
Nutritionists often recommend intakes from particular food subgroups in order to achieve specific nutrient adequacy. Citrus fruits are considered to be vitamin C rich fruits; pro-vitamin A is abundant in dark green or yellow fruits and vegetables. Intuitively, this informatics process for data quality audit of the FCT is based on the idea that similar foods should consist of similar levels of nutrients; this is thought to agree with common sense and current theory which underpins dietary guidelines and food guides. At the same time, the evaluation of nutrient variety represented by a diversity of food groups, with defined nutritional characteristics, can assist with the universal dietary guideline that encourages variety and for which there is growing evidence (Wahlqvist and Lee, 2007).
We used CVs for the outliers, being those ranked highest either for food subgroups or for particular nutrients, as well as the least products of these 2 ranks to find the most suspect pieces of nutrient information – they were marked as unlikely values and constituted ‘‘hits’’. This identification of unlikely values acted as a filter to evaluate the entire food composition table. The ranks for CVs (CV RANK) and their product (CRP) could provide a threshold and minimize the need for manual review of each piece of nutrient information.
The filtered and unlikely nutrient value was far from the expected, if similar foods have like levels of nutrients. The Panel reviewed and discussed the unlikely values in the Taiwan FCT as ‘‘system-defined hits’’ (SH) and also considered them in
conjunc-tion with the China and Japan Food Composiconjunc-tion Tables (China
Institute of Nutrition and Food Safety, 2002; Kagawa, 1999). Several extreme food nutrient values were identified. This made it possible to reappraise them and to comment about the appro-priateness of analytical techniques for some foods.
The variability of analytical data in FCT is high for micro-nutrients (Leclercq et al., 2001; Salvini, 1997; Stryker et al., 1991) and is moreover increased by taking other factors into con-sideration: sources of plants, animal husbandry, storage, trans-port and marketing. Food processing methods vary by ingredient content and composition with changes in formulation and production. There are also problems from artefacts introduced though inadequacy of sampling, differences in analytical meth-ods, lack of use of quality assurance programs, and differences in deriving protein and other macronutrient values with the compilation of FCT (Leclercq et al., 2001). The inappropriate compilation of FCT has caused gross errors in the assessment of nutrient intake levels in European countries. These have included time trends and between-country estimates for food and nutrient intakes. Improper food composition data limit the understanding of the relationships among nutrients, health and disease (Dwyer, 1994).
Our informatics approach, with more efficient use of an Expert Panel, has provided a way to check whether data might derive from incompatible variation. ‘‘Hits’’ for extreme variation of nutrient values have provided an audit for FCT. The approach has enabled the compilation of more correct sources of data for FCT by verifying the definitions, analytical methods and modes of expression originally used. At the same time, where the intra-subgroup Table 3
The subgroups ‘‘grapes’’ and ‘‘watermelons’’ in the FRUITS group of Taiwan Food Composition Tables: a comparison with Chinese and Japanese data for vitamin A activity.
Code Food items Suggestions Vitamin A activity (RE)
Subgroup of Grapes
D019400 California grape - Indicate plant breed 37.5
D020400 White grape - Too variable for vitamin A activity 1410
D021400 Grape 0 Chinese FCT Grape 70.8 Hongmeigui# 0 Kyoho 38.5 Menindee 76.9 Muscat 22.9 Dark-skinned 86.2 Japanese FCT Grape 32.1 Subgroup of Watermelons D026400 Xiaoyu#
; yellow pulp - Too variable for vitamin A activity 3.33
D027400 Watermelon; red pulp 1810
Chinese FCT Watermelon 1119 Jinxin# 1 148 Zhenzhou#3 530 Zhongyu# 6 494 Japanese FCT Watermelon 707 #
variation is real, the approach may provide opportunities to refine nutritional advice, based on reliable FCT.
The Panel was convened to clarify food similarity when audit process hits were discovered. The clearer and more certain identity of food items, by scientific names and illustration, is recommended for FCT (Poortvliet et al., 1992; Puwastien, 2002; Rand et al., 1991; Southgate and Greenfield, 1992). Scientific names would enable foods within a subgroup to be more biologically and nutritionally similar and would allow a more convincing subgroup to be created in the first place. This process would be strengthened by having foods labelled in accordance with a scientific classification, as well as secular names, as has been the objective in the INFOODs project through the FAO, United Nations University and IUNS (Interna-tional Union of Nutri(Interna-tional Sciences). Illustrations (photos) of food items, appropriately referenced, are also recommended. INFOODs recommends constructing a complete internet homepage of foods throughout the world to assist when food items are difficult to
identify and name (Masson, 1999; Rand and Young, 1984). Food
illustrations, the various secular, local or traditional food names, scientific names and adequate food descriptions would help practitioners with their dietary assessments when utilizing FCT and help consumers and manufacturers to implement recommen-dations, especially on Internet-based food information.
Another challenge in the assessment of the utility and validity of FCT is the overall food culture (whether by geography or ethnicity), which may have more or fewer food items in a particular food group or subgroup. Examples would be the few or many types of cheese in European subgroups of the dairy food group. Another would be the spices group which could comprise few or many items depending on the food culture in question, and which could lead to an enormous potential variability in nutrient contents.
In the case of the spices group, a particular difficulty arises: this is partly because of the small quantities that may have disproportionate biological effects, partly because the bio-active food components are likely to extend to many different phyto-nutrients, and partly because of different potential effects dependent on background diet. This example typifies how the demands for an ‘‘informatics process’’ applied to FCT are likely to grow rapidly beyond the present requirements for macro- and micronutrients.
The ‘‘informatics’’ approach in the present paper acknowledges the possibility that an outlier nutrient content in a food subgroup may not be the same from one food culture to another. This may be because the range of food species of apparently biologically similar foods may vary considerably. In turn, the median nutrient value and the CVs may be skewed. However, there is no reason not to test different cut-off points, FCT by FCT, and make the criteria known. From a practical point of view, where those responsible for FCT accuracy in various settings wish to use our method, there will be translational requirements. One has to do with differences in numbers of food categories and items and in number and type of nutrients. The cut-off points (CV RANKS and CRP) might be affected by the number of nutrients in a database because the variations and precisions of macro- and micronutrients are influential in laboratory processes and methods.
In conclusion, an informatics process for data audit is able to identify potentially unlikely nutrient values in apparently similar foods. It can help improve the accuracy and utility of FCT in the Taiwan area. It should also improve FCT in other countries and be responsive to new demands on FCT as long as they become more and more digitized. The demands on quality assurance (QA) programs for FCT as databases expand, with growing numbers of foods and bio-active food components subject to measurement
Appendix A
Subgroup assignments for the VEGETABLES group.a
Code Subgroup Name (Food code)b
0 Unsorted (Un-proc
): Burdock (E001400), Lotus root (E003400), Dasheen stalk (E007400), Chinese water-chestnut (E010400), Green asparagus juice (E016502), Shallot (E020400), Onion (E023400), Scallion (E024400), Leek (E025400), Pakchoi (E026400), Basil (E027400), Chinese knotweed (E029400), Sweet potato leaves (E031400), Leaf mustard (E034400), Coriander (E036400), Edible rape (E038400), Ching- chiang pakchio (E039400), Garland chrysanthemum (E042400), Shannon coriander (E043400), Crown green (E045400), Fern bud (E053400), Solanum nigrum (E054400), Chayote stem shoot (E055400), Day-lily flower (E060400), Vegetable sponge flower (E061400), Rape flower (E062400), Eggplant (E067400), Pumpkin (E069400), Bell pepper (E070400), Chayote stem gourd (E071400), Yellow okra (E072400), Chili (E076400), Mountain potato (E086500), Yam (E088600), White mugwort (E091600), Green water cress (E092600), Feather cockscomb E094600, Nodding burnweed (E095600), Red stem pearl vegetable (E096600), Fresh Mastixia (E097600), Cedrus (E098600), Coral flower (E101600), Houttuynia (E105600), Guogou vegetable fern, (E106600), Garbo melon (E107600), Dragon grass (E111600), Mint (E112600), Chinese onion (E113600), Notoginseng (E114600), Aloe (E115600); (Prod
): Chili sauce (E076401), Pickled mustard leaves (E078400), Pickled leaf mustard (E079400), Oriental pickled melon (E081400), Pickled mustard stem (E082400), Pickled cucumber in soy sauce (E083300)
1 Carrots (Un-pro):Carrot (E002400); (Pro): Frozen carrot (E002601)
2 Radishes (Un-pro): Radish (E004400);
(Pro): Pickled radish (E004501), Dried radish (E004502)
3 Bamboo shoots (Un-pro): Bamboo shoot (E005400) Betel-nut palm stem (E006400) Water-bamboo (E009400) Ma bamboo shoot (E012400); (Pro): Dried bamboo shoots (E005503) Frozen bamboo shoot (E005601) Canned bamboo shoot (E005602)
4 Cabbages (Un-pro): Kohlrabi (E011400), Cabbage (E030400), Cabbage (E030401), Purple cabbage (E050400), Baby cabbage (E100600); (Pro): Frozen cabbage (E030602), Dried cabbage (E030603)
5 Gingers (Un-pro): Young ginger (E014400), Ginger (E017400);
(Pro): Pickled young ginger (E014501)
6 Asparagus (Un-pro): Green asparagus (E016400), Asparagus (E019400); (Pro): Canned green asparagus (E016601)
7 Stem lettuce (Un-pro): Stem lettuce (E018400), Snow pea stem (E047400); (Pro): Pickled stem lettuce (E018501)
8 Chinese chives Chinese chives (E021400), Yellow Chinese chives (E022400), Chinese leek flowerbud (E059400) 9 Bean sprouts Soy bean sprouts (E013400), Mung bean sprouts (E015400)
and nutritional evaluation, will be great. The present approach should allow QA commensurate with this trend.
Acknowledgements
This study was supported by a grant (DOH94-TD-F-113-067-(2)) from the Research Fund of Bureau of Food Sanitation, Department of Health, Taiwan, Republic of China.
References
Alexander, J.T., Cheung, W.K., Dietz, C.B., Leibowitz, S.F., 1993. Meal patterns and macronutrient intake after peripheral and PVN injections of the alpha 2-receptor antagonist idazoxan. Physiol. Behav. 53, 623–630.
Bowers, L., 2002. An audit of referrals of children with autistic spectrum disorder to the dietetic service. J. Hum. Nutr. Diet. 15, 141–144.
Buzzard, I.M., Asp, E.H., Chlebowski, R.T., Boyar, A.P., Jeffery, R.W., Nixon, D.W., Blackburn, G.L., Jochimsen, P.R., Scanlon, E.F., Insull, W.J.R., et al., 1990. Diet intervention methods to reduce fat intake: nutrient and food group composi-tion of self-selected low-fat diets. J. Am. Diet. Assoc. 90 (42-50), 53. China Institute of Nutrition and Food Safety, 2002. Chinese Food Composition. China
CDC, Beijing (in Chinese).
Dartois, A.M., Ducamp, S., Decaux, F., Broyer, M., 1989. Computer-assisted diet therapy in pediatric kidney diseases. Pediatrie 44, 197–202 (in French). Deharveng, G., Charrondiere, U.R., Slimani, N., Southgate, D.A., Riboli, E., 1999.
Comparison of nutrients in the food composition tables available in the nine European countries participating in EPIC. European Prospective Investigation into Cancer and Nutrition. Eur. J. Clin. Nutr. 53, 60–79.
Dwyer, J.T., 1994. Future directions in food composition studies. J. Nutr. 124, 1783S– 1788S.
Fabiansson, S.U., 2006. Precision in nutritional information declarations on food labels in Australia. Asia Pac. J. Clin. Nutr. 15, 451–458.
Farran-Codina, A., Boatella-Riera, J., Serra-Majem, L., Ribas, L., Raffcas-Martinez, M., Codony-Salcedo, R., 1994. General criteria for the development and use of data tables and systems on food composition. Rev. Sanid. Hig. Publica. (Madr) 68, 427–441.
Fidanza, F., Perriello, G., 2002. Validation of the Italian food composition database of the European institute of oncology. Eur. J. Clin. Nutr. 56, 1004–1010. Guilland, J.C., Aubert, R., Lhuissier, M., Peres, G., Montagnon, B., Fuchs, F., Merlet, N.,
Astorg, P.O., 1993. Computerized analysis of food records: role of coding and food composition database. Eur. J. Clin. Nutr. 47, 445–453.
Institute for Interlaboratory Studies, 2006. Statistics in Porficiency Tests, Vol. 2006. ISO/IEC, 1997a. ISO/IEC Guide 43-1—Part 1: Development and operation of profi-ciency testing schemes. In: Profiprofi-ciency testing by interlaboratory comparisons. ISO/IEC.
ISO/IEC, 1997b. ISO/IEC Guide 43-2—Part 2: Selection and use of proficiency testing schemes by laboratory accreditation bodies. In: Proficiency testing by inter-laboratory comparisons. ISO/IEC.
Japan Science and Technology Agency, 2005. Food Composition Database (in Japanese).
Kagawa, Y. (Ed.), 1999. Standard Tables of Food Composition in Japan. Women’s Nutrition College Publishing Department, Tokyo (in Japanese).
Leclercq, C., Valsta, L.M., Turrini, A., 2001. Food composition issues—implications for the development of food-based dietary guidelines. Public Health Nutr. 4, 677– 682.
Masson, L., 1999. LATINFOODS and its role in the generation and compilation of data for Latin America. Arch Latinoam Nutr. 49, 89S–91S (in Spanish).
Matsuda-Inoguchi, N., Shimbo, S., Nakatsuka, H., Watanabe, T., Higashikawa, K., Ikeda, M., 2004. Effects of revision of Japanese food composition tables on estimation of nutrient intakes, with reference to age-dependent differences. Public Health Nutr. 7, 901–909.
Mccullough, M.L., Karanja, N.M., Lin, P.H., Obarzanek, E., Phillips, K.M., Laws, R.L., Vollmer, W.M., O’Connor, E.A., Champagne, C.M., Windhauser, M.M., 1999. Comparison of 4 nutrient databases with chemical composition data from the Dietary Approaches to Stop Hypertension trial. DASH Collaborative Research Group. J. Am. Diet. Assoc. 99, S45–S53.
NATA, 1996. New Statistics for Proficiency Testing Programs. National Association of Testing Athorities (NATA), Australia.
Phillips, K.M., Patterson, K.Y., Rasor, A.S., Exler, J., Haytowitz, D.B., Holden, J.M., Pehrsson, P.R., 2006. Quality-control materials in the USDA National Food and Nutrient Analysis Program (NFNAP). Anal. Bioanal. Chem. 384, 1341– 1355.
Poortvliet, E.J., Klensin, J.C., Kohlmeier, L., 1992. Rationale document for the Euro-code 2 food coding system (version 91/2). Eur. J. Clin. Nutr. 46 (Suppl. 5), S9– S24.
Probst, Y.C., Krnavek, C., Lockyer, L., Tapsell, L.C., 2004. Developing a self-adminis-tered computer assisted dietary assessment tool for use in primary healthcare practice: perceptions of nutrition and computers in older adults with T2DM. Asia Pac. J. Clin. Nutr. 13, S136.
Puwastien, P., 2002. Issues in the development and use of food composition databases. Public Health Nutr. 5, 991–999.
Rand, W.H., 1985. Food composition data: problems and plans. J. Am. Diet. Assoc. 85, 1081–1083.
Rand, W.M., 1987. INFOODS and food composition data. Arch. Latinoam. Nutr. 37, 609–617 (in Spanish).
Rand, W.M., Pennington, J.A.T., Murphy, S.P., Klensin, J.C., 1991. Compiling Data for Food Composition Data Bases. The United Nations University, Tokyo, Japan, p. 77.
Rand, W.M., Young, V.R., 1984. Report of a planning conference concerning an international network of food data systems (INFOODS). Am. J. Clin. Nutr. 39, 144–151.
Rodriguez, M.C., Coupeau, I., Larralde, J., Martinez, A., 1995. Development of a computer program to assess nutritional status of infants and preschool chil-dren. Arch. Latinoam. Nutr. 45, 274–280 (in Spanish).
Appendix A (Continued )
Code Subgroup Name (Food code)b
10 Spinaches (Un-pro): Spinach (E049400); (Pro): Frozen spinach (E049601)
11 Cauliflower (Un-pro): Cauliflower (E057400);
(Pro): Frozen cauliflower (E057401)
12 Wax gourd (Un-pro): Wax gourd (E063400);
(Pro): Pickled wax gourd (E063601)
13 Baby corns (Un-pro): Baby corn (E064400);(Pro): Canned baby corn (E064601) 14 Cucumbers (Un-pro): Small cucumber (E065400), Cucumber (E068400);
(Pro): Pickled cucumber (E065401)
15 Bitter gourds Un-pro): Bitter gourd (E066400), Wild bitter gourd (E102600), Young wild bitter gourd (E103600); (Pro): Pickled bitter gourd (E066401)
16 Luffa Luffa (E073400), Angled luffa (E077400)
17 Mustard tuber (Pro): Canned hot pickled mustard tuber (E080401), Hot pickled mustard tuber (E080500)
18 Tomatoes (Un-pro): Tomato (E084400);
(Pro): Canned tomato (E084402), Tomato juice (E084501) 19 Broccoli (Un-pro): Broccoli (E058400); (Pro): Frozen broccoli (E058601)
20 Chinese cabbages San-tong cabbage (E028400), Chinese cabbage (E032400), Cuiyu cabbage (E109600), Baby Cuiyu cabbage (E110600) 21 Celery Chinese celery (E033400), Celery (E040400), Hill celery (E087600)
22 Water convolvulus Water convolvulus (E037400), Water cress (E041400)
23 Lettuce Head lettuce (E051400), Lettuce leaves (E052400), Leaf lettuce (E056400) 24 Bottle gourd Bottle gourd (E074400), Bottle gourd (E075400)
25 Snake gourd White snake gourd (E089600), Green snake gourd (E108600)
26 Sprouts Alfalfa sprouts (E008400), Chinese kale (E035400), Chinese kale sprouts (E093600)
27 Gynura Red gynura (E044400), White gynura (E090600)
Salvini, S., 1997. A food composition database for epidemiological studies in Italy. Cancer Lett. 114, 299–300.
Samsonov, M.A., Meshcheriakova, V.A., Shadur, S.S., Shekhter, E.G., 1986. A com-puter-assisted method of evaluating the actual nutrition of patients with ischemic heart disease during their clinical observation. Vopr. Pitan. 18–21 (in Russian).
Sasaki, S., Kobayashi, M., Tsugane, S., 1999. Development of substituted fatty acid food composition table for the use in nutritional epidemiologic studies for Japanese populations: its methodological backgrounds and the evaluation. J. Epidemiol. 9, 190–207.
Shai, I., Vardi, H., Shahar, D.R., Azrad, A.B., Fraser, D., 2003. Adaptation of interna-tional nutrition databases and data-entry system tools to a specific population. Public Health Nutr. 6, 401–406.
Sharp, M.M., Ahmed, K., 1983. A computer application for dietary analysis in clinical nutrition. J. Can. Diet. Assoc. 44, 228–234.
Shimbo, S., Hayase, A., Murakami, M., Hatai, I., Higashikawa, K., Moon, C.S., Zhang, Z.W., Watanabe, T., Iguchi, H., Ikeda, M., 1996. Use of a food composition database to estimate daily dietary intake of nutrient or trace elements in Japan, with reference to its limitation. Food Addit. Contam. 13, 775–786.
Southgate, D.A., Greenfield, H., 1992. Principles for the preparation of nutritional data bases and food composition tables. World Rev. Nutr. Diet. 68, 27–48.
Stryker, W.S., Salvini, S., Stampfer, M.J., Sampson, L., Colditz, G.A., Willett, W.C., 1991. Contributions of specific foods to absolute intake and between-person variation of nutrient consumption. J. Am. Diet. Assoc. 91, 172–178.
Taiwan Department of Health, 2002a. Taiwan Food Composition Database: basic food composition nutrient tables (in Chinese) Retrieved 2004-10-14:http:// www.doh.gov.tw/FoodAnalysis/ingredients.htm.
Taiwan Department of Health, 2002b. Taiwan Food Composition Database: prin-ciple of classification and coding system (in Chinese) Retrieved 2004-10-14:
http://www.doh.gov.tw/FoodAnalysis/account.htm.
Taiwan Department of Health, 2002c. Taiwan Food Composition
Database: Vegetable (in Chinese). Retrieved 2004-10-14:
.
Vaask, S., Pomerleau, J., Pudule, I., Grinberga, D., Abaravicius, A., Robertson, A., Mckee, M., 2004. Comparison of the Micro-Nutrica Nutritional Analysis pro-gram and the Russian Food Composition Database using data from the Baltic Nutrition Surveys. Eur. J. Clin. Nutr. 58, 573–579.
Van Wave, T.W., Decker, M., 2003. Secondary analysis of a marketing research database reveals patterns in dairy product purchases over time. J. Am. Diet. Assoc. 103, 445–453.
Wahlqvist, M.L., Lee, M.S., 2007. Regional food culture and development. Asia. Pac. J. Clin. Nutr. 16, 2–7.
Wikipedia, 2006. Grape. Retrieved 2006-06-21: http://en.wikipedia.org/wiki/ Grape.
Zhang, Z.W., Shimbo, S., Miyake, K., Watanabe, T., Nakatsuka, H., Matsuda-Inoguchi, N., Moon, C.S., Higashikawa, K., Ikeda, M., 1999. Estimates of mineral intakes using food composition tables vs measures by inductively-coupled plasma mass spectrometry: Part 1. calcium, phosphorus and iron. Eur. J. Clin. Nutr. 53, 226–232.
