Handling data imperfection-False data inputs in applications for Alzheimer's patients

摘要

Handling data imperfection is a crucial issue in many application domains. This is particularly true when handling imperfect data inputs in applications for Alzheimer's patients. In this paper we first propose a typology of imperfection for data entered by Alzheimer's patients or their caregivers in the context of these applications (mainly due to the memory discordance caused by the disease). This topology includes nine direct and three indirect imperfection types. The direct ones are deduced from the data inputs e.g. uncertainty and uselessness. The indirect imperfection types are deduced from the direct ones, e.g. the redundancy. We then propose an approach, called DBE_ALZ, that handles false data entry by estimating the believability of each data input. Based on the proposed typology, the falsity of these data is related to five imperfection types: uncertainty, confusion, typing error, wrong knowledge and inconsistency. DBE_ALZ includes a believability model that defines a set of dimensions and sub-dimensions allowing a qualitative estimation of the believability of a given data input. It is estimated based on its reasonableness and the reliability of its author. Compared to related work, the data input reasonableness is measured not only based on common-sense standard, but also based on a set of personalized assertions. The reliability of the patient is estimated based on the progression of the disease and the state of his memory at the moment of entry. However, the reliability of the caregiver is estimated based on his age and his knowledge about the data input's field. Based on the believability model, we estimate quantitatively the believability of the data input by defining a set of metrics associated to the proposed dimensions and sub dimensions. The measurement methods rely on probability and fuzzy set theories to reason about uncertain and imprecise knowledge (Bayesian networks and Mamdani fuzzy inference systems). Three languages are supported: English, French and Arabic. Based on the generated believability degrees, a set of decisive actions are proposed to guarantee the quality of the data inputs e.g., inferring or not based on a given data. We illustrate the usefulness of our approach in the context of the Captain Memo memory prosthesis. Finally, we discuss the encouraging results derived from the evaluation step.