Affiliation:
Funding Period: 2005 - 2006
Funding Source: Intel BAIC
View related topics
Abstract
This project will develop and apply the statistical inference algorithms required to use unobtrusively-gathered longitudinal data to predict and identify cognitive health changes. These algorithms will address the varied challenges inherent in human behavioral data including variability within and between individuals, non-stationarity in both healthy and deteriorating individuals, instrumentation noise, sensor drift, and data drop-out. The algorithms will be crafted by combining exploratory data analysis with modern statistical learning technology.
The ORCATECH BRP and BAIC projects will provide longitudinal data on various activities. These include sleep patterns, medication adherence, computer use, gait patterns, daily motion through the home (including patterns of room visitation, walking speed, and overall activity and its variability), and standard clinical tests. Data modalities include contact (door) sensors; IR motion sensors; RFID identifiers; bed mats; the pillbox MedTracker; low-resolution video; and computer mouse, keyboard, and voice interface events.
Using these varied sensor modalities and the monitored activities to quickly and accurately detect emerging changes in the individual poses a significant inference challenge. Furthermore, we need to distinguish changes that reflect cognitive decline from changes that reflect normal individual variability, healthy aging, or non-cognitive illnesses.
The overall aim is to develop and apply inference models that accrue information over time to rapidly and robustly identify changes in the individual. These algorithms will also distinguish between changes related to cognitive deterioration and those related to normal aging or to non-cognitive illness. Although the specific activities and modalities addressed will depend on data availability, we expect to focus our efforts on motion and computer use data from the BRP and its allied BAIC pilot studies.
Our hypotheses regarding effective detection of cognitive decline are based on our experience designing statistical detectors of gradually emerging phenomena that have high variability across individual events, but few examples of the onset. Although we will explore them early on to understand the efficacy of discriminative features derived from the data, we expect that naïve application of traditional classifiers (decision trees, neural networks, support vector machines) will fail due to large variation between individuals and a dearth of examples of individuals transitioning from health to cognitive impairment. We will also explore outlier detection approaches, in which the behavior of healthy individuals only is modeled.