Infusing Domain Knowledge to Improve the Detection of Alzheimer's Disease from Everyday Motion Behaviour

摘要

Alzheimer's disease can severely impair the independent lifestyle of a person. Dem@Care is an European research project that conducted a study for timely diagnosis of Alzheimers disease by collecting everyday motion data from couples (or dyads), with one of the person in the couple having AD. Their results suggest that AD can be detected using everyday motion data from accelerometers. They did evaluation based on leave-one-person-out cross-validation. However, this evaluation can introduce bias in the classification results because one of the person from the dyad is present in the training set while the other is being tested. In this paper, we revisit the Dem@Care study and propose a new evaluation method that performs leave one-dyad-out cross-validation to remove the dataset selection bias. We then introduce new domain specific features based on dynamic and static intervals of motions that significantly improves the classification results. We further show increase in performance by combining the proposed features with new time, frequency domain and baseline features used in the Dem@Care study.