Modeling elderly behavioral patterns in single-person households

This paper proposes and describes an unsupervised computational model that monitors an elderly person who lives alone and issues alarms when a risk to the elderly person’s well-being is identified. This model is based on data extracted exclusively from passive infrared motion sensors connected to a ZigBee wireless network. The proposed monitoring system and model is non-intrusive, does not capture any images, and does not require any interaction with the monitored person. Thus, it is more likely to be adopted by members of the elderly population who might reject other more intrusive or complex types of technology. The developed computational model for activity discovery employs a kernel estimator and local outlier factor calculation, which are reliable and have a low computational cost. This model was tested with data collected over a period of 25 days from two elderly volunteers who live alone and have fairly different routines. The results demonstrate the model’s ability to learn relevant behaviors, as well as identify and issue alarms for atypical activities that can be suggestive of health problems. This low-cost, minimalistic sensor network approach is especially suited to the reality of underdeveloped (and developing) countries where assisted living communities are not available and low cost and ease of use are paramount.