Advertisement

34 pages
9 views

A Keyhole Plan Recognition Model for Alzheimer's Patients: First Results

of 34
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Share
Description
A Keyhole Plan Recognition Model for Alzheimer's Patients: First Results
Transcript
  A Keyhole Plan Recognition Model forAlzheimer’s Patients: First Results Bruno Bouchard 1 , Abdenour Bouzouane 2 , and Sylvain Giroux 1 1 Universit´e de Sherbrooke, 2500, boul. de l’Universit´e,Sherbrooke, Qu´ebec, Canada J1K2R1 { Bruno.Bouchard, Sylvain.Giroux } @usherbrooke.ca 2 Universit´e du Qu´ebec `a Chicoutimi, 555, boul. de l’Universit´e,Chicoutimi, Qu´ebec, Canada G7H2B1 abdenour bouzouane@uqac.ca Abstract.  This paper 3 addresses the problem of recognizing the behav-ior of a person suffering from Alzheimer’s disease at early-intermediatestages. We present a keyhole plan recognition model, based on latticetheory and action description logic, which transforms the recognitionproblem into a classification issue. This approach allows us to formal-ize the plausible incoherent intentions of the patient, resulting from thesymptoms of his cognitive impairment, such as disorientation, memorylapse, etc. An implementation of this model was tested in our smarthome laboratory, by simulating a set of real case scenarios. 1 Introduction The aging of the population in occidental societies has significant consequencesfor healthcare systems, including medical staff shortages for patient home careservices, and an increasing number of people suffering from a category of disor-ders known clinically as  dementias   [42]. The most widespread of these dementiasis Alzheimer’s disease. Currently in Canada, an estimated 280, 000 citizens over65 have Alzheimer’s disease and more than half a million Canadians will havethis disease by 2031 [20]. This dementia is characterized by brain lesions causinga progressive deterioration of thinking (cognitive impairment) and of memory.These symptoms lead to incoherent behavior limiting the patient’s capacity toperform his tasks of everyday life [45].In this context, a growing literature [11][36][58] has begun to explore the process by which cognitive assistance, inside a smart home, can be provided toan occupant suffering from some type of dementia, such Alzheimer’s disease,for the performance of his Activities of Daily Living (ADL). A smart homeis an augmented environment with miniaturized processors, software (agents)communicating between each other, and multi-modal sensors that are embeddedin any kind of common everyday objects, making them effectively invisible to the 3 This paper has been published in Vol. 22 (7) of the prestigious Journal of Applied ArtificialIntelligence (AAI), Taylor & Francis publisher, July 2007.  2 Bruno Bouchard et al. resident [26]. One of the major difficulties inherent in cognitive assistance is toidentify the on-going inhabitant ADL from observed basic actions and from theevents produced by these actions. This difficulty corresponds to the so-called plan recognition   problem that has been well studied in the field of ArtificialIntelligence (AI) [17].The problem of plan recognition can be described by the need “...to takeas input a sequence of actions performed by an actor and to infer the goalpursued by the actor and also to organize the action sequence in terms of a planstructure” [55]. The main objective of this recognition is to predict the behaviorof the observed entity. In the field of cognitive assistance, these predictions areused to identify the various ways a smart home (observer agent) may help itsoccupant (Alzheimer’s patient).Plan recognition can be classified into two main categories, namely  intended  plan recognition and  keyhole   plan recognition [19]. In the first case, one assumesthat the patient knows that he is being observed and is adapting his behaviorin order to make his intentions clear to the observer. Consequently, this form of recognition supposes a cooperative effort on the part of the observed entity. Inthe second case, one supposes that the patient does not know that he is beingobserved or that he is not taking it into account, hence the analogy of someonebeing observed through a keyhole. In this paper, we will deal only with keyholeplan recognition, which is, strictly speaking, more general than intended planrecognition in that it makes no assumption about efforts to cooperate on behalf of the patient [61].The keyhole plan recognition problem can be characterized by whether theobserver has a complete knowledge of the domain or not, and by the possibil-ity that the observed agent may try to perform erroneous plans [47]. Indeed,dealing with a human actor automatically implies the possibility of observingerrors during the performance of a plan. This reality is greatly amplified whenwe are concerned with Alzheimer’s patients [14]. The reason is that healthypeople are usually rational, which means that they adopt a behavior where alltheir performed actions are coherent with their intentions. However, for peo-ple suffering from cognitive impairments, such Alzheimer’s patients, rationalitymight indeed be too strong an assumption. In fact, people with Alzheimer’s dis-ease will usually act incoherently, performing much more erroneous plans thanhealthy people. The assumption of complete domain knowledge is quite usefulin order to limit the scope of the recognition task [35]. Nevertheless, if one mayreasonably assume that the observer agent knows all possible ways to correctlyperform an activity, one cannot assume the same for all the possible erroneousplans that may be carried out. The model described in this paper considersthat the observer agent has an incomplete knowledge base by assuming that theobserved patient, according to his cognitive deficit, will potentially make someerrors, which are not preestablished in the plan library, in the achievement of his everyday tasks.The literature related to keyhole plan recognition may be divided into threemajor streams. The first one comprises works [16][40][61] based on logical ap-  Plan Recognition for Alzheimer’s Patients 3 proaches that stem from Kautz’s contribution [35]. These formal approaches seekto identify, by a series of logical deductions, the set of possible plans that couldexplain a set of observed actions. An important assumption underlying this typeof approach is that the observed agent is rational. However, as pointed out, wecannot assume the rationality of Alzheimer’s patients.The second stream of literature on keyhole plan recognition, including thework of [1][11][18], makes use of probabilistic reasoning without incorporating a learning process. The basic idea is to manually assign a probability to eachpossible plan and to predefine a stochastic model that can update these like-lihoods according to new observations and to the known state of the system.A major strength of these probabilistic plan recognition methods is that theyallow one to capture the fact that certain plans are, a priori, more likely thanothers [35]. Hence, the probabilities are used to model the uncertainty relatedto the recognition task directly in the plan likelihood. The limitation of the non-learning probabilistic methods stems from the set of handcrafted variables usedto compute the plan likelihood (such as a preestablished probabilistic transitionmatrix), which are static and highly context-dependent [17].The third stream of literature comprises the work of [4][6][31][33][60] basedon learning techniques. These techniques try to identify patterns in the behaviorof the observed person and to extract from it a predictive model characteriz-ing his common routines. Some of these learning techniques are also based onprobabilistic approaches [6][33][60]. The main difference between non-learning probabilistic methods and learning probabilistic techniques is that, instead of using a preestablished stochastic model to update the plan likelihood, they keepa trace of their previous observing experiences and use them to dynamicallylearn the parameters of the stochastic model, in order to create a predictivemodel based on the observed agent’s habits. The disadvantage of this kind of probabilistic learning techniques is that they need a large amount of trainingdata to be efficient and also, they are limited in the recognition of unusual ornovel events that did not occur beforehand.We propose, in this paper, a keyhole plan recognition model following theline of Kautz [35] and Wobke [61], which is specifically adapted to Alzheimer’spatients. The objective of this model is to recognize the kind of behavior of aperson suffering from Alzheimer at early-intermediate stages, by formally makingexplicit the plausible behavioral incoherencies induced by his cognitive deficit.For instance, an Alzheimer’s patient may begin to cook some pasta by takingthe pasta box and by putting water on to boil. Then, he may be victim of amemory lapse that makes him, a few minutes later, completely deviate fromhis srcinal intention and use the hot water to make tea instead of pasta [14].This recognition issue is even more complex owing the fact that the patient mayperform multiple coherent tasks in an interleaved way.Our approach tries to address this issue by using lattice theory [54] and anaction model [15] based on Description Logics (DL) [2], which transforms the plan recognition problem into a classification issue. Description logics are a well-known family of knowledge representation formalisms that may be viewed as  4 Bruno Bouchard et al. fragments of first-order logic. The main strength of DL is that they offer consid-erable expressive power going far beyond propositional logic, although reasoningis still decidable [3]. Hence, our DL action model provides an adequate basis todefine algebraic tools used to formalize the inferential process of ADL recogni-tion for Alzheimer’s patients. To summarize, our approach consists of developinga model of minimal interpretation for a set of observed actions, by building aplan lattice structure corresponding to the set of possible plans. In this model,the uncertainty related to the anticipated patient’s behavior is characterized byan intention schema. This schema corresponds to the lower bound of the latticeand is used to extract the anticipated incoherent plans, which are not preestab-lished in the knowledge base, that the patient may potentially carry out as aresult of the symptoms of his disease. Therefore, our approach addresses theissue of completing the plans library, which of course cannot be complete inany domain. Moreover, this approach minimizes the uncertainty of the predic-tions. By minimizing uncertainty, we mean that it bounds the recognition spaceand provides a way to decide when an assisting action might be performed, byhelping the smart home agent to identify the opportunities for assistance. Thiswork can be considered as one brick in a larger research project pursued atthe DOMUS 4 laboratory, aiming to develop an intelligent habitat able to assistAlzheimer’s patients in the performance of their Activities of Daily Living [45].In this project, we are collaborating with researchers studying Alzheimer’s pa-tients at the Research Center of Aging affiliated with the Sherbrooke GeriatricUniversity Institute.The paper is organized as follows. The next section draws an overall picture of Alzheimer’s disease and defines the typical problems encountered by Alzheimer’spatients while performing their tasks of everyday life. The third section presentsour formal keyhole plan recognition model. The fourth section shows how thismodel is implemented in the DOMUS smart home laboratory to concretely ad-dress ADL recognition of an Alzheimer’s patient. The fifth section presents theresults of our first experimentation phase for this implementation, using a set of real case scenarios. The sixth section presents the previous significant works re-lated to plan recognition and most notably, those closer to our approach. Finally,the last section presents our conclusion and future work. 2 Overall picture of Alzheimer’s disease The Senile Dementia of the Alzheimer’s Type (SDAT), which is commonly calledAlzheimer’s disease, is characterized by progressive deterioration of the carrier’sintellectual capacities that evolves during a period of 7 to 10 years on the aver-age [20]. This dementia can be classified in 7 degeneration stages, by referringto the global scale of deterioration stages (GDS) of cognitive functions of an in-dividual [52]. During the first stages (1-2), the symptoms are not very apparent 4 The DOMUS laboratory is sponsored by the Natural Sciences and Engineering Re-search Council of Canada (NSERC) and by the Canadian Foundation for Innovation(CFI).  Plan Recognition for Alzheimer’s Patients 5 and the patient remains autonomous. At these initial stages, the patient oftenhides his cognitive losses by avoiding embarrassing situations [45]. In the finalstages of the disease (6-7), the patient suffer from heavy symptoms, such as ver-bal and communication problems (aphasia), difficulties for identifying peoplesand objects (agnosia), and severe motivation troubles (apraxia). At this point,the patient must be completely taken in charge by a caregiver for his own se-curity, because he is no longer able to meet his primary needs (to feed, washhimself, etc.). The intermediate stages of the disease (3-5) constitute the longestportion of the degeneration process. At these stages, the patient mostly suffersfrom weakening executive functions, sporadic losses of memory, and problems onfocusing his attention on a specific task. Therefore, a distraction (e.g. phone call,unfamiliar noise, etc.) or a memory lapse can lead him to perform actions in thewrong order, to skip some steps of his activity, or to perform actions that are noteven related to his srcinal goal [42]. However, the patient’s capacity to performa simple action (without many steps) remains relatively unaffected [52]. Theseintermediate stages require supervision of the patient and  ad hoc   interventionson the part of an assistant. When continued support is provided to Alzheimer’spatients in the form of cognitive aide, the degeneration process of the diseaseis slowed and he can remain at home longer [11]. Hence, having an intelligentagent able to assist such patients where and when necessary can relieve some of the burden carried by natural and professional caregivers. This technology there-fore has the potential to delay the institutionalization of patients by alleviatingsome caregiver duties while restoring partial autonomy to the care recipient. Inthe longer term, it could also constitute an economically viable solution to theincreasing cost of home care services. 2.1 Characterizing the behavior of Alzheimer’s patients The behavior of Alzheimer’s patients is characterized by the presence of recur-rent incoherencies in their actions, which appear when they perform complexactivities involving cognitive skills [45]. Indeed, people without such impairmentcan act incoherently too. The main difference is that a healthy person is, most of the time, able to recognize his behavioral errors and to correct them by himself.Moreover, healthy people do not act incoherently on a regular basis. In con-trast, a person suffering from Alzheimer’s disease will certainly act incoherently,even while performing familiar tasks, and his behavior will become increasinglyincoherent as the disease evolves.According to  Baum et al.  [8], these behavioral incoherencies of Alzheimer’spatients can be classified in six categories of errors:  initiation  ,  organization  ,  re-alization  ,  sequence  ,  judgment   and  completion  .  Initiation   errors happen when thepatient is, for any reason, unable to begin his task. For example, if a therapistindicates to an Alzheimer’s patient that he must take his medication right now,the patient may answer “OK, I’m going to take it now” but does nothing.  Orga-nization   errors happen when the patient performs some steps of an activity inan inappropriate way. For instance, the patient can use the wrong type of spoon,or even a knife, to mix up the ingredients of a receipt.  Realization   errors happen
Related Documents
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks
SAVE OUR EARTH

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!

x