Submitted to: International Blaise Users Conference
Publication Type: Proceedings
Publication Acceptance Date: August 6, 2004
Publication Date: September 14, 2004
Citation: Anderson, E., Seinfeldt, L. 2004. Blaise instrument design for automated food coding. 9th International Blaise Users Conference Proceedings, p. 129-136. Available: http://www.blaiseusers.org/IBUCPDFS/2004/10.pdf Interpretive Summary: The Automated Multiple Pass Method (AMPM) Blaise instrument, developed by the Food Surveys Research Group, Agricultural Research Service, U.S. Department of Agriculture, collects 24-hour dietary recall data. During the interview, respondents recall the foods and beverages that were consumed the day before the interview. Details describing the types of foods and beverages are collected as well as an estimate of the portions consumed. These details are used to assign food codes and calculate gram weights and nutrients for the reported foods. There are more than 2,500 questions and more than 21,000 responses in the AMPM instrument. While the AMPM Blaise instrument was designed to facilitate the food intake interview, also considered was how the food detail data stored in the Blaise database would be extracted for manual and automated coding. Each question is represented by a unique field name. Open-ended responses are limited by the use of over 90 look-up tables, resulting in standardized responses. The unique combination of field names and responses creates a pathway that describe a particular food. The number of possible pathways through all the food detail questions is estimated to be over 400,000. To identify the most frequently reported pathways, we analyzed approximately 147,000 foods that had been reported using the AMPM instrument and coded manually by trained food coders. From this data set, a database of approximately 12,000 pathways linked to food codes was created. Pathways were grouped by food category and assigned a unique path number. The most common pathways were reviewed by a nutritionist, ensuring that the correct food code was linked to the pathway. Once the pathways are approved as correct, they are incorporated into the Post Interview Processing System (PIPS) where they are used to automatically code food intake data collected with the AMPM. Currently, PIPS is able to autocode approximately 56% of reported foods using 2,788 reviewed pathways linked to over 1,300 foods.
Technical Abstract: The Automated Multiple Pass Method (AMPM) Blaise instrument, developed by the Food Surveys Research Group, Agricultural Research Service, U.S. Department of Agriculture, collects 24-hour dietary recall data. During the interview, individuals recall the foods and beverages that were consumed the day before the interview. Details about each food and beverage are collected as well as an estimate of the amount consumed. AMPM contains more than 2,500 questions and more than 21,000 responses. Ninety-five percent of the questions are about specific food details, including the amount of food eaten. The large number of questions and responses produce an even larger number of skip patterns because the questions asked about a food depends on the responses to the previous questions. The number of possible paths through the food detail questions is roughly estimated to be over four hundred thousand. The food detail information is used to assign food codes and calculate gram weights and nutrients for the foods reported consumed. During the design of the instrument, careful attention was paid to how the food detail data would be stored in the Blaise database to insure that it could be extracted and used for manual and automated food coding. Foods were grouped into categories and the fields within each category were given a standard prefix. Across categories the same question, for example, "What kind was it?" was given a standard suffix. Responses were written to be understandable to the food coder as well as the interviewer. Open-ended responses were limited by the use of over 90 look-up tables. The use of standardized responses allows the creation of food coding paths. Food coding paths are unique sets of field names and responses reported for a specific food. Common paths could be used for automated food coding, while infrequently reported paths would be left to manual coding. Initially, 10 commonly reported foods (e.g. milk) were linked with 28 food coding paths. These 10 foods produced an initial automated food coding rate of 11% of reported foods. Next, approximately 147,000 coded and approved foods were analyzed to determine find commonly reported paths that could be linked to a single food code. The most common paths were selected, reviewed and approved by a nutritionist, and then incorporated into the Post Interview Processing System (PIPS). Currently 2,788 pathways linked to over 1,300 foods are automatically coding approximately 56% of the foods reported.