dWellSense Reveals Critical Factors for the Analysis of Home-Based Sensor Data-Quality of Life Technology Center - Carnegie Mellon University

dWellSense Reveals Critical Factors for the Analysis of Home-Based Sensor Data

Outcome / Accomplishment:

dWellSense - a system developed by QoLT human-system interaction researchers, Anind Dey and Matthew Lee - embeds sensors into everyday home appliances to help detect the potential onset of dementia, physical decline and other changes in the wellness of adults who live alone.

Data collected by Lee (CMU-HCII '12) through the dWellSense system shows that time dimensions are critical factors for sensor-based self-reflection. Mutliple long-term field deployments of dWellSense found that real-time feedback is particularly useful for supporting behavior changes, while longer-term, trended feedback is more useful for supporting a greater awareness of one's individual abilities and identifying deviations from or shifts within observed behavior patterns.


Impact / Benefit:

While ubiquitous sensing systems promise to collect volumes of data about human behavior and activity, there are significant challenges to be faced in making this abundance of highly dimensional data useful. Such challenges extend from when and how often to present collected data, to how the data can be used to support personal or clinical analysis and reflection.

The dWellSense data findings provide a foundation for the design of other personal sensing systems that aim to either assess an individual's abilities or support his or her specific behaviors through the delivery of sensor-based data as objective and timely user feedback.

dWellSense also significantly advances the state-of-the-art for maintaining quality of life and care for older adults and other individuals facing the possibility of cognitive decline. Traditional forms of assessing the functional abilities of individuals tend to be biased, lacking ecological validity, infrequent, or expensive to conduct. In contrast, when tested by a trained clinician, an automated sensor-based approach for assessment matches well with clinician-generated ratings that are objective, frequent, and ecologically valid. The dWellSense team demonstrated that sensors can be used to effectively monitor observations of daily living, even among low-tech populations, and alert users and their medical providers to decline.

Lee presented dWellSense at the Health 2.0 conference in San Francisco on October 8, 2012 in a session entitled, "Big Data Meets Real[time] People"; Dey also gave a talk on the project at the WIRED Health Conference, Living by Numbers, on October 16 in New York City.

Media coverage of the dWellSense conference presentations inspired a First LEGO League robotics team (grades 4 through 8) in Northern Illinois to design a device that would help seniors take their pills. The team noted: "We liked the idea of technology that monitored when pills were taken. Hence, our solution included a computer program that would monitor whether or not [an older adult] has taken their pills. If the [older adult] failed to take their pills, this computer program would call three people and notify them." The team's video presentation can be found online.


Explanation / Background:

Many older adults desire to maintain their quality of life by aging independently in their own homes. But it can be difficult for older adults to notice and track subtle changes in their functional abilities that can have impact on their safety at home. Such changes often occur gradually over long periods of time. Technology - in the form of ubiquitous sensors embedded in typical household objects used in the home - can play a role in keeping track of these functional abilities unobtrusively, objectively, and continuously over time. A person's daily routine provides enough contextual information that it can be instructive in making sense of collected data.

dWellSense specifically explores how a person's routine performance on Instrumental Activities of Daily Living (IADL) tasks can be used as a context informing health status. IADL tasks are frequently performed in the home and require a high level of cognitive functioning. Using a sensing technique called "embedded assessment of wellness," everyday objects in the home that individuals interact with can be used to moniter how well specific IADL tasks that are important for independence are carried out.

The dWellSense system monitors, assesses, and provides feedback about how well individuals take their medications, use the phone, and make coffee. The physical sensor set includes the following:

  • For medication taking, dWellSense uses an ordinary pillbox retrofitted with switch sensors that can detect how people open it and an accelerometer to track how people handle the box.
  • For phone use, it uses a simple circuit that tracks how often people mis-dial the phone.
  • For coffee making, dWellSense uses an ordinary coffeemaker retrofitted with sensors to track the different steps of making coffee like putting in the water, coffee, and turning it on.

Additional infrastructure allows each sensor to transmit, log, and recognize activity data. The full system architecture relies on cloud-based analytics, web-based graphs and real-time feedback displays.

Multiple, long-term field deployments of dWellSense were carried out over a period of 10 months to investigate how data collected from the system were used to support a.) greater self-awareness of abilities; and, b.) intentions to improve in task performance. The average age of the participants was 67.5.

For user feedback, dWellSense employed two data visualization options: 1.) an always-on Android tablet application that functions as an in-home dashboard display providing near real-time feedback (updated once every 30 minutes); and 2.) interactive, web-based graphics to provide delayed feedback on behavior trends over configurable time spans. The two modes allowed for reflection to take place either immediately after a particular instance of a behavior is performed; or, after a short delay, thereby allowing for reflection on trends across multiple instances of behavior.

Overall, the advantage of real-time feedback is that behavior is still fresh, providing a "teachable moment" opportunity to immediately inform behavior changes that may be necessary. Care should be taken to assure real-time feedback does not become intrusive; simple visualizations better enable quick reflection. In contrast, long-term feedback is effective for allowing users to reflect on activity over a specific period of time. By exploring multiple data points, trends or performance patterns can more easily emerge, offering a more convincing case for any problems that may require correction. The drawback of long-term feedback is that some anomalies may be harder to explain if their potential causes are not well remembered.

dWellSense shows that the process of allowing users to reflect on their own behavioral data is powerful enough to encourage behavior change. Although some of the data provided by dWellSense was initially shocking to users, it allowed them to look at their activities and behavior more objectively. The system also provided data points that could prove useful as early indicators for adverse events when shared with clinicians.

dWellSense successfully helped re-orient users' self-awareness. One example of this was found in improvements to the consistency of medication compliance. Four specific measures of pill-taking were explored: adherence (how often pills were taken); correctness (whether pills were taken on a particular day); promptness (whether pills were taken at the correct time); and. variance in the time of day (changes in promptness across days where lower variance was better).

The real-time display helped users improve adherence, promptness, correctness, and variance in time taken; whereas retrospective reflection on trended, long-term data only had a temporary, marginal effect on promptness. Before users were given the real-time dashboard display, pill-taking routines were scattered - with evidence of users taking pills late or missing some altogether. With the real-time display, pill-taking became extremely consistent. The long-term reflection visualization was most successful in significantly increasing the accuracy of users' self-awareness of medication adherence mistakes.

dWellSense shows that an understanding of the differences and trade-offs between real-time and long-term reflection is important for designing feedback systems that avoid overburdening users with information while supporting behavior change. Key observations from the dWellSense study showed that:

  • Real-time feedback helps individuals to initiate and sustain behavior-change.
  • Real-time feedback does NOT increase self-awareness.
  • Long-term feedback increases self-awareness accuracy, but only temporarily.
  • Long-term feedback triggers deeper re-evaluation of personal abilities.

Future sensing-based systems should provide feedback modes that can be combined to provide the greatest amount of self-awareness and complementary support for behavior change. By presenting long-term feedback first, an individual can gain an accurate history and awareness of their behavior and the motivation to set appropriate goals for behavior or attitude change. After long-term reflection, the system should provide real-time feedback to support the individual's ability to achieve and meet their goals and sub-goals. Periodic long-term reflection allows a user to highlight discrepancies within long-term trends and gauge whether or not they are reaching their overall long-term goals.

The dWellSense project was funded in part by the Robert Wood Johnson Foundation's Project HealthDesign program.


NSF Achievement Category:

Research and Technology Advancements



By: Kristen Sabol, ksabol @ cs.cmu.edu