Achievement

We designed a method for automating a multimedia prompting system to aid older adults with initiating and completing ADL activities. We train the system using sensor data collected while 260 participants perform ADL activities in a smart apartment. A clinically-trained experimenter watched participants over a web camera and remotely played an audio or video prompt if they observed that the participant was having significant difficulty with the activity. Each sensor event is treated as a sample data point and is labeled by whether a prompt is needed or not. Because the majority of sensor events do not require prompt interventions, this problem presents a challenge of learning from data with skewed class distributions. We investigated a variety of techniques to handle this challenge. Our experiments yield an Area Under the ROC Curve (AUC) value of 0.817. We are currently designing a version of the system that runs on smart phones.