146x Filetype PDF File size 0.13 MB Source: www.aaai.org
Dealing with Malnutrition: A Meal Planning System for Elderly JohanAberg Department of Computer and Information Science ¨ Linkopings universitet ¨ 581 83 Linkoping, Sweden {johab}@ida.liu.se Abstract of our ageing population, the problem of malnutrition must Malnutrition is a serious problem among people of old age. be dealt with. Toovercomemalnutrition, a change of food consumption be- There are several causes of malnutrition, see e.g. the dis- haviour is necessary, which needs to be based on specialist cussion in (McCormack 1997). There can be specific age- advicefromhealth-careprofessionals. Changingfood-related related causes (e.g. optimal nutrient intake is affected by in- behaviour, however, is known to be difficult. Our approach to dividual rates of change in physiologic function, or by dis- this problem is to provide an intelligent meal planning sys- easesordrugtherapies),economiccauses(e.g.financiallim- tem to be used by the elderly person in his or her home. The itations lead to a down-prioritisation of nutritious food), so- system provides recommendations of suitable food recipes, cial causes (e.g. loss of spouse, causing a loss of appetite taking into account the advice of the care givers (e.g. in terms due to depression, or simply not knowing how to cook, or of dietary restrictions, suitable energy and fat levels, etc). We what constitutes a nutritious meal), and limited dietary di- describe the requirements, design, algorithms, and user inter- versity (Kant et al. 1993). Hence, the role of health-care face of the system, and discuss ongoing and future work. professionals is to educate and motivate the elderly patient to change his or her food consumption behaviour. Introduction Changing habits of food consumption is known to be dif- The world’s population is ageing. Due to societal improve- ficult, and may require continual supervision and education ments in health-care, living standards, and socioeconomic (Maciuszek, Aberg, & Shahmehri 2005). However, such status, more and more people are living to old age. The pro- support is not always available due to shortages in care re- portionoftheworld’spopulationaged65oroverisexpected sources. Thus, as an aid to changing food-consumption be- to increase from 11% in 1998, to 16% in 2025 (U.S. Bureau haviour weproposeanintelligent food support system, to be oftheCensus1998). Thiscausesamajorpublichealthissue, usedbytheelderlypersoninhisorherhome,capableofpro- aswithincreasedagethereisanincreasedriskofdeveloping viding informed and individualised suggestions about what a number of age-related diseases. to eat. The system takes several important variables into There is scientific evidence that many of the biologi- account in the suggestions, such as taste, cost, preparation cal changes and risks for chronic disease which have tra- difficulty, dietary diversity, dietary restrictions, nutritional ditionally been attributed to ageing are in fact caused by needs and properties, and available food items. Hence, a malnutrition (sub-optimal diets and nutrient intakes) (Beck- health-care provider’s suggestions for the user can be incor- man & Ames 1998; Blumberg 1994; Chandra et al. 1982; poratedintothesystemasindividualconstraints. Suchasys- ´ tem, if used properly, has the potential of limiting the prob- Mowe, Bøhmer, & Kindt 1994; Potter et al. 1995; Vellas lemofmalnutrition. Forexample,forelderlywitheconomic et al. 1997). While some nutritional surveys of the elderly constraints, low cost meals with good nutritional properties have shown relatively low prevalence of frank nutrient de- canbesuggested,optimisingtheuseofavailablefooditems, ficiencies, there is a clear increase in risk of malnutrition whilestill taking the taste of the user into account, and main- ¨ ¨ (Blumberg 1997; Sjogren, Osterberg, & Steen 1994), and a taining dietary diversity. Note that our system should not be high prevalence of malnutrition of elderly patients admitted regarded as a finished product, but rather as a tool for further to different clinical settings has been reported in the litera- investigations into the malnutrition problem of the elderly ´ ture (Larsson et al. 1990; Mowe, Bøhmer, & Kindt 1994; and how artificial intelligence can make a difference. Volkert et al. 1992). It has also been shown that hospitali- The rest of this paper is organised as follows. In the sation as such has a negative influence on nutritional status next section, we describe the problem of changing food con- ˚ of geriatric patients (Elmstahl et al. 1997; Larsson et al. sumption behaviour in more depth. After that we describe 1990). Hence, to solve the challenges of improving quality the meal planning system in terms of the requirements, the of life and preventing or reducing disability and dependency design, the algorithms, and the graphical user interface. We c thendiscussongoingworkandfuturedirections. Finally,we Copyright ° 2006, American Association for Artificial Intelli- gence (www.aaai.org). All rights reserved. conclude the paper. Attitude AMealPlanningSystem toward the Our approach to helping users change their food consump- behaviour tion behaviour is a system that recommends meal plans. As such, our system is a recommender system, which is a class of decision aids, where the aim is to provide users with in- Subjective Intention Behaviour dividualised recommendations on objects from some partic- norm ular domain (Montaner, Lopez, & de la Rosa 2003). Rec- ommendersystemshavesofarbeenofgreatimportance for e-commerce (Schafer, Konstan, & Riedl 2001), and also of Perceived valueforotherimportanttaskssuchasinformationsearchon behavioural the Internet (Montaner, Lopez, & de la Rosa 2003). How- control ever, to the best of our knowledge, our system is the first recommendersystemtobeappliedtoahealthcareproblem. Figure 1: Theory of planned behaviour Requirements The factors influencing a person’s food choice have been ChangingFoodConsumptionBehaviour studied to a fairly large extent in the science of food and nutrition. Shepherd (Shepherd 1989) described several at- The problem of malnutrition is caused by sub-optimal eat- tempts to identify factors influencing food choice, and went ing. Hence, in order to come to terms with a malnutrition on to propose the use of the theory of reasoned action as a problem, a person must change the food consumption be- general model for food choice. However, this model is com- haviour, and eat food that better fits his or her body’s current pletely based on user’s attitudes, and does not seem suitable needs. However,changingfood-relatedbehavioursisknown as a normative framework. After all, we are not really in- to be difficult, and continual support is commonly needed. terested in predicting a user’s food choice, but to persuade According to one of the dominant theories in social psy- the user of choosing optimal food, weighing in the relevant chology, the theory of planned behaviour (Ajzen 1991) factors. Hence, we have taken the approach of gathering 3 factors from all the models presented in (which is based on the theory of reasoned action), human the most feasible behaviour is determined by specific considerations (see Fig- (Shepherd 1989). This means that our system is required to ure 1). Behavioural beliefs refer to the outcome of a be- represent and reason about the following information: haviour and the evaluation of the outcomes, and lead to an ² Dietary restrictions, e.g. ingredients that the user is aller- attitude toward the behaviour. Normative beliefs refer to gic to, or must not eat for other medical reasons. the perceived expectations of others and the motivation to ² Nutritional values, e.g. amount of fat or protein contained live up to these expectations, and lead to a subjective norm. in a recipe, or required by a user. Control beliefs refer to factors that can help or hinder per- formanceofthebehaviourandtheirrelativeimportance,and ² Preparation time of a meal. lead to perceived behavioural control. Together, the attitude ² Preparation difficulty of a meal. towardthebehaviour,thesubjectivenorm,andtheperceived ² Cost of a meal, i.e. the cost of the needed ingredients. behavioural control lead to a behavioural intention. Finally, given an intention to perform a behaviour and the perceived ² Availability of ingredients for a meal, e.g. to what extent behavioural control, a person is expected to succeed in per- does the needed ingredients match the ingredients avail- forming the behaviour (assuming the perceived behavioural able to the user at home. control is close to the actual behavioural control). ² Variation with respect to other meals in the plan, in terms Our approach to help elderly people change their food of used ingredients and the category of a meal. consumption behaviour (and thus deal with malnutrition ² The user’s food taste, i.e. how the user rates a recipe on a problems), is to provide them with a tool for meal plan- 1. Connecting to the theory of taste scale. ning to be used in their homes planned behaviour, we expect this would raise the perceived Design 2, in the sense that they feel behavioural control of the users that they have all the knowledge and resources needed for To be able to take these requirements into account, the sys- actually changing their behaviour and prepare and consume tem has a hybrid design in the sense that it makes use of meals suitable for them. This meal planning system is de- both collaborative filtering and a content-based approach. scribed next. The collaborative filtering is used for predicting a user’s taste opinion of a certain recipe that he or she has not yet 1This obviously raises questions of whether the users would be rated, based on the user’s other ratings and the ratings of able and willing to use the system. See our discussion on user other users. For the content-based approach we make use of studies below. 2Ourfocusonincreasingauser’sperceivedbehaviouralcontrol 3By feasible we mean that a factor should be feasible to make does not mean that we neglect the other factors influencing inten- use of in the system, with respect to practical knowledge engineer- tion and behaviour, it simply means that we must start somewhere. ing and reasoning issues. a specially designed XML-based mark-up language for food whereausercanselectthetimeperiodforwhichthesystem recipes, that allow us to represent the needed content infor- will recommend meals. Note that this is just a part of the 4. Ourapproachtocon- mationfortherecipesinthedatabase settings that a user can perform. Among other things, a user struct optimal meal plans according to the factors presented can also select required intervals for energy, fat, cholesterol, above uses constraint satisfaction techniques. More details etc. Such settings are absolutely crucial for our purpose of onthealgorithmsemployedinourmealplanningsystemare helpingelderlypeopleavoidmalnutrition,andtheactualset- provided next. tings should be done in collaboration with the user’s care givers. The user can also select ingredients to avoid, select Algorithms preference levels for cost, preparation time, etc, and mark Wemodeltheconstraint-satisfaction problem with a mix of ingredients as currently available. 6 shows an example of a recommended meal plan weighted soft constraints and traditional hard constraints, Figure 3 similar to the approach in (Torrens 2002). We have exper- for a certain time period. Note that the user can switch be- imented with two different ways of modelling the problem. tween the top-5 meal plans, and give taste ratings on sug- In our parameter-based approach, variables are used for the gested recipes (on a scale from 1 to 5) and re-plan to take parameters of a recipe, such as time, cost, energy, protein, the new ratings into account, or create special settings for a etc, and the variable domains are based on the existing val- certain meal, such as allowing a greater cost and preparation ues in the recipe database. There is also a special hard con- time for the Sunday meal. straint requiring a complete variable assignment to match OngoingWork only existing recipes in the database. The other, recipe- based, model is simpler, and has only one variable per meal Algorithms in the plan, with the set of recipes as value domain. In our ongoing work, we are investigating the trade-offs Forbothmodels,weemployasetofadditionalconstraints for the two constraint-satisfaction models we have imple- to take the user’s needs and preferences into account. Such mented. The main aspect we are looking into is the com- constraints include hard constraints, e.g. for ingredients to putation time required for solving the constraint-satisfaction avoid, and soft constraints, e.g. for variation between meals problem. Based on our two alternative constraint models of (a recipe with many ingredients in common with a recipe for the meal planning problem, we are experimenting with the a previous meal gets a penalty) and for taste (recipes with following parameters: high rating or predicted rating get low penalty). A collabo- rative filtering approach is used to predict ratings for unrated ² Numberofrecipes in the database. recipes. We have implemented a version of the item-based ² Numberofmealsintherequested plan (e.g. the length of algorithm (Sarwar et al. 2001), with adjusted cosine simi- the time period). larity, and weighted sum predictions. ² Thenumberofuserstoplanfor(e.g.thesizeofthefamily, For solving the constraint-satisfaction problem we base or the number of persons in the assisted living facility). our approach on the well-known depth-first branch and bound algorithm. We have also been experimenting with Based on a set of 50 real food recipes and a much larger a set of forward-checking approaches and variable order- set of randomly generated recipes (the random generation ing heuristics. Our current implementation uses depth-first is based on parameters from the real recipes) we have con- branch and bound with partial forward checking. ducted several simulation experiments. So far, the results User Interface indicate that the two models have complementary charac- teristics. The parameter-based model performs very well The user interface of the system has been designed partic- with small recipe collections and can make plans for sev- ularly for elderly users. The current user interface design eral meals with just a few seconds response time. How- is the result of an in-depth exploration of the design space ever, this model only provides reasonable response times for (by means of the QOC framework (MacLean et al. 1991)), data sets of a maximum of roughly 500 recipes. The recipe- taking existing literature on universal access and user inter- based model on the other hand scales well with increasing face design for elderly into account as evaluation criteria for riod”). Theareatotherightofthecalendarshowstheuser’scurrent the explored design options. Two separate prototype designs calendar choices as additional feedback. The rightmost area of the were created as paper prototypes and evaluated empirically screen contains a help text for this particular settings page. This with elderly users. Based on these user studies the current help text can be toggled on and off, but is on by default. user interface was designed and implemented, in an attempt 6This screen shot illustrates the meal plan menu generated by to use the best features from each of the two earlier proto- the system. Currently the first, and best, alternative is displayed. 5 shows a part of the settings management, types. Figure 2 The user can toggle between different alternatives with the top- 4 most buttons. Note that the recipe names are shown in English in Wehavealsodeveloped a semi-automatic tool to facilitate the this example. To the right of each recipe name is a slider for chang- extraction of information from food recipes in text format. ing the taste rating, and to the right of this slider is the current rat- 5The user interface is designed in Swedish, so some additional ing, written in text. In this example all recipes have previously been explanations may be needed for most readers of this paper. The rated by the user, so no predicted ratings are displayed. Above the menu to the left is used for reaching different settings pages. In area for the recommended recipes are buttons for changing settings this screen shot we are at the settings for the time period (“Tidspe- and for replanning based on new taste ratings. Figure 2: Meal planning system settings: selecting the time period for the meal plan recipe collections, but is limited with respect to the number through all the test tasks, all four users have been able to of meals to plan for. use all the main functions of the system, despite minor flaws and initial orientation problems. This is an important result, User Studies illustrating the potential of the system. As discussed previously, our main aim with the meal plan- Future Directions ning system is to put the elderly person in charge of chang- ing his or her food consumption behaviour. By providing Faster Algorithms the user with recommendations of suitable recipes that take into account important parameters such as dietary restric- Given the preliminary results reported on earlier, there is tions, cost, and the preparation skills of the user, we hope to a clear need to further investigate means to speed up meal increase the user’s behavioural control, which is critical for planning algorithms. Ideally we would like the system to be changing behaviour. However, it is of course a prerequisite able to make plans for at least a week at the time, and with that the user accepts the system and really uses it. Although large databases containing up to 10,000 recipes. Even if our the user interface has been designed based on user studies of system is working well with our current small databases, we paper prototypes involving potential future users, an impor- are not there yet, and we won’t be there in a few years ei- tant question is whether the implemented system is usable ther if we extrapolate using Moore’s law. Hence, we plan and acceptable for elderly users. To answer this question to continue our efforts to improve the algorithms and the weare currently conducting a user study with several older constraint models. As a starting point we intend to examine adults. In the study, after a very brief oral description of the how the complementing characteristics of the two present system,theusersareassignedasetoftaskstobesolvedwith constraint models can be exploited. We will also explore the system. At the time of writing we have data from four the possibilities of terminating the search before the whole users (with age ranging from 70 to 82, and with varying de- search tree has been explored. The rationale for this ap- grees of previous computer experience). Our observations proach comes from our empirical results showing that the and interviews have highlighted several problems with the time spent searching for the last few complete assignments current interface that will need to be adressed in the next make up the great majority of the total time spent on the version. However, and more importantly, after having gone search, while the reduction in upper bound that these last as-
no reviews yet
Please Login to review.