Sensor-Based Analysis of Object-Use Patterns for the Automatic Assessment of Cognitive Status

Indications of cognitive impairments such as dementia and traumatic brain injury (TBI) are often subtle and may be frequently missed by primary care physicians. These impairments are not uncommon: approximately 0.46% of Americans are hospitalized for brain injury each year and it is estimated that by 2050 up to 13 million Americans may have Alzheimer's disease—the most common form of dementia—quadruple the number that did in 2002. This dissertation proposes and investigates ways in which machine inference and wireless sensors can be used to support the assessment of cognitive functioning. The central hypothesis is that object-usage data collected from wireless sensors during the performance of daily activities are sufficient to assess cognitive impairment. I first investigate the ability to recognize individuals based on their sensed object-usage patterns during a simple task. This experiment constitutes an initial step in understanding how well object-use patterns can be automatically observed and analyzed. A preliminary study, using the simple task of preparing a cup of coffee, demonstrated the ability to correctly recognize ten individuals with 77% accuracy based on nine trials from each individual as training data. The dissertation then directly addresses assessment of cognitive impairment with a study in which individuals with TBI made a pot of coffee. Four features were derived from the sensed data and compared to the subjects' scores on standard neuropsychological evaluations. A key result is that Edit Distance, the most knowledge-rich feature, significantly correlates with an apparent indicator of general neuropsychological integrity, namely the first principal component of the neuropsychological assessments. Since cognitive impairments are measured along many dimensions, suggestive correlations between the computed features and individual assessments are also presented. Lastly, I present a preliminary study that investigates the possibility of differentiating impaired individuals from unimpaired individuals. Data was collected from five subjects with TBI and five matched, unimpaired subjects; analysis was done using the same set of computed features. Although the study is preliminary, it is interesting that Edit Distance is able to perfectly differentiate the impaired subjects from the unimpaired and that full results are consistent with those from the assessment study.