This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR mhealth and uhealth, is properly cited. The complete bibliographic information, a link to the original publication on http://mhealth.jmir.org/, as well as this copyright and license information must be included.
An ongoing challenge for smart homes research for aging-in-place is how to make sense of the large amounts of data from in-home sensors to facilitate real-time monitoring and develop reliable alerts.
The objective of our study was to explore the usefulness of a routine-based approach for making sense of smart home data for the elderly.
Maximum variation sampling was used to select three cases for an in-depth mixed methods exploration of the daily routines of three elderly participants in a smart home trial using 180 days of power use and motion sensor data and longitudinal interview data.
Sensor data accurately matched self-reported routines. By comparing daily movement data with personal routines, it was possible to identify changes in routine that signaled illness, recovery from bereavement, and gradual deterioration of sleep quality and daily movement. Interview and sensor data also identified changes in routine with variations in temperature and daylight hours.
The findings demonstrated that a routine-based approach makes interpreting sensor data easy, intuitive, and transparent. They highlighted the importance of understanding and accounting for individual differences in preferences for routinization and the influence of the cyclical nature of daily routines, social or cultural rhythms, and seasonal changes in temperature and daylight hours when interpreting information based on sensor data. This research has demonstrated the usefulness of a routine-based approach for making sense of smart home data, which has furthered the understanding of the challenges that need to be addressed in order to make real-time monitoring and effective alerts a reality.
An aging population is a challenge to current models of health care delivery and engagement [
Real-time alerts and visual presentation of data from smart homes sensors are a work in progress. In a 3-month UK pilot study, Sixsmith [
Visual presentation is also important for making smart homes a reality. The amount of data generated by continuous monitoring systems can result in information overload [
In the Sixsmith pilot study [
Routines are described as “strategically designed behavioral patterns (conscious and subconscious) used to organize and coordinate activities along the axes of time, duration, social and physical contexts, sequence and order” [
This study presented results from a pilot of Smarter Safer Homes testing ubiquitous home monitoring for the elderly, as described in Zhang et al [
In-home sensors for residents.
Sensors | Trigger for sensor firing | Place of installation | Sensor data upload | Data type |
Passive infrared motion sensors | Motion within 5 m | Wall (near ceiling) in all rooms | Sends ad hoc as status change | Binary |
Current draw of appliances | Wall power outlets | Pushes 1-minute data every 5 minutes | KwH to binary | |
Circuit meter | Current draw of stove or oven | Switchboard | Pushes 1-minute data every 5 minutes | KwH to binary |
Accelerometer | Movement | Under bed | Sends ad hoc as status change | Binary |
Reed switches | On breaking of circuit | Exit doors. Kitchen or bedroom doors | Sends ad hoc as status change | Binary |
Acoustic sensor | Water flow | Kitchen | Sends ad hoc as status change | Binary |
Environmental sensors | Continuous data collection | Kitchen, bathroom, laundry | Pushes 1 reading every minute | Temperature and humidity |
Electronic |
Daily recording of temperaturea | Indoor usage | Sends ad hoc after measurement | Temperature |
Glucometer of blood | Daily recording of blood pressurea | Indoor usage | Sends ad hoc after measurement | Glucose |
Electronic scalesb | Daily recording of weighta | Indoor usage | Sends ad hoc after measurement | Weight |
iPad and Web portal | Residents were given an iPad for their personal use and to consult an app on which summary sensor data appeared and diary for health measurements |
aResidents would vary in the regularity and consistency of use of these devices.
bInformation sent directly to database without the need for user data entry.
The data used in this study were collected from passive infrared motion sensors and power use sensors located in participants’ lounge, kitchen, bedroom, and bathroom. These sensors were chosen because they were the most reliable and provided the bulk of the data. Motion sensors send data whenever motion is detected. Motion sensors fire on detection of movement, but cannot distinguish between types of movements. Also, movements have to be sufficiently large for sensors to fire. Power use sensors were connected to all key appliances in the kitchen and lounge. Power use sensors send data (KwH) every 5 minutes.
Sensors were installed and maintained by a local technician. The technician was very friendly, and his visits were appreciated by all residents. The most common complaints involved battery replacement and flickering lights on motion and power sensors. Duct tape was used to hide the light on motion sensors, and residents were shown how to switch the light off on power sensors.
Time (GMT) and date stamped data from motion and power sensors were relayed to a Web-based database and downloaded to Microsoft Access. Each entry identified the residence, room, sensor type (motion or power), and description (eg, kettle). Time was converted into local time, which comprised 16 days of AEST (GMT+10) and 175 days AEDT (GMT+11).
Data were analyzed using Microsoft Access. Prior to analysis, data were cleaned. Standby power was removed for appliances such as televisions and microwaves (based on low wattage) and then converted into binary data (in use or not in use). Missing data because of sensor battery failure or absence from home were identified. Absences over 24 hours were noted, but did not negatively affect daily data or cumulative data. Battery failure was distinguishable because of normal movement in other rooms. For some motion sensors, a notable increase in sensor firing prior to battery failure caused high outliers. High outliers could also occur on days residents received guests or visitors. Mean replacement was used for high outliers (±3×IQR).
Continuous streamed data 24/7 for 3 residents over 181 days resulted in 345,470 data entries. Data were presented as 24-hour radar plots, which reflected daily patterns of movement and power based on cumulative frequency of sensor firing. Data were grouped by hour (± 30 minutes). Radar plots reflected the build-up and change in pattern of movement or power over time. Sharp spikes in the outlines indicated strict adherence to time schedules.
Residents of an aged care facility living in independent units self-selected for participation in the pilot. To be eligible to participate in the pilot, participants had to be aged over 70 years and have no home care arrangements. Participants with cognitive difficulties were also excluded. Of those who self-selected (N=23), 17 signed consent forms; however, 3 residents withdrew before the sensors were installed. Retention of participants in longitudinal trials with the elderly was problematic for reasons including morbidity, mortality, relocation, or other. Over the course of the 180 days of the trial, there were further withdrawals. We collected 7 complete sets of data (sensor and interview data at 3 time points); however, only 5 residents were eligible for this study. Participants living in dual occupancy were excluded because multiple occupancy was problematic for interpreting data from the motion sensor used in this trial [
Three interviews were conducted 2, 6, and 8 months after sensor installation. Interviews were recorded and transcribed. Relatives or friends were present at most interviews and contributed to the discussion. Relatives of case study participants were also interviewed separately.
Participants answered a brief questionnaire on preference for routinization using a shortened version of Reich & Zautra’s scale [
Preference for routinization and outings based on qualitative data and scores for routinization.
Pseudonym | Scorea | Description | Activity | Regular outings | ||
Family | Care facility | Community | ||||
Rupert | n/ab | Set daily routines that vary little from day to day | Fairly sedentary or cerebral | Low but regular | High | Low |
Elizabeth | 100% | Routine varies by day of week | Active or always busy | Med | Med | High |
Jacqui | 63% | Flexible daily routines that also vary by day of week. Travels a lot. | Very active or restless energy | High (daily) | Low | Low |
aShortened scale of 8 items [
bOnly answered true to 1 of the 5 statements.
Although daily activities comprise many common elements such as eating and sleeping, the way in which activities are organized in time reflects the uniqueness of individual routines [
Jacqui, a widow and former secretary, travels a lot and is frequently away from home for as many as 14 days at a time. She is also not often at home during the day, and her routine is irregular. She scored 5 out of 8 items (63%), indicating a low preference for routinization. Her grandson said: “she’s got her routines but sometimes she just doesn't do them.”
Results comparing routines constructed from sensor data closely matched resident’s self-reported routines, as illustrated in
Elizabeth: Radar plot showing cumulative motion in bedroom by time of day (24/7 over 181 days) as indicated by the time of firing of motion sensors. Data is grouped by hour (±30 minutes).
Elizabeth: Radar plot showing kitchen appliance power use (kettle solid line, microwave dotted line) by time of day based on cumulative frequency (24/7 over 181 days). Text in italics reflects the routine activity as per Elizabeth's self-reported routine.
Times series data are traditionally analyzed by examining the component parts: cyclical, seasonal, trend, and irregular variations. In these data, cyclical variation shows as recurring weekly patterns by day of week. Seasonal variation in data includes annual recurrent patterns of temperature and timing of sunset and sunrise by time of year. Trend data is longitudinal change over time, which is important for the detection of slow-onset gradual decline in physical or cognitive abilities over and above cyclical and seasonal data. Finally, irregular variation can explain ad hoc disruption to routines such as illness or death of a spouse.
Cyclical variation results from the organization of days of the week into work and rest days as well as the annual cycle of holidays. These cultural rhythms appear in the routines of the elderly as a continuation of past habits or indirectly through their contact with relatives. Jacqui’s routine absences from home are dependent on the schedule of her relatives. The most direct impact of cultural rhythms on the elderly is the scheduling of community events such as church and bridge club.
Routine absence from home can be clearly identified on radar charts. Monday through Saturday, Rupert has lunch at the residential home restaurant. This regular absence from home is clear from
Rupert: Radar plot using cumulative data to show absence of lounge movement Monday to Saturday. Absence is highlighted using a reverse scale such that no movement appears at the outer edge of the diagram and frequent movement appears at the centre.
Rupert: Radar plot using cumulative data to show absence of lounge movement on Sundays. Absence is highlighted using a reverse scale such that no movement appears at the outer edge of the diagram and frequent movement appears at the centre.
Routines vary according to the seasons and weather. All residents spoke of changes to routines that were related to the temperature. For example, Elizabeth explained changes in hygiene practice:
(Shower time) can vary... in the winter time it will be before an evening meal... if you have your shower... just when the sun’s going on down, the room is still a bit warm, because (the sun is) on that side of the house... If you leave it until nine o’clock on a frosty night, it’s a lot colder in (the bathroom)... In the summer time I shower every day, but in the cold weather I get lazy and I think... I am not going out tomorrow I won’t shower tonight.
In winter, Elizabeth also dresses later in the day. As she stays in her warm dressing gown longer, she also has hot porridge rather than muesli for breakfast. Temperature also affects her level of activity. The results show changes in the frequency of motion associated with temperature. Independent-samples
Above average high or low temperatures can lead to temporary changes in routine. Interview data indicated that Jacqui and Rupert changed their walking routine on hot days, either forgoing walks or changing the time of day. Hot weather can also affect the amount of movement in the home. Elizabeth is more likely to take an afternoon nap: “No, I’m not one to sleep very much, but over the hot weather... I have gone to sleep twice.”
Furthermore, the number of hours of daylight varies with seasons. Participants reported timing certain activities with sunset.
I mean, I – I walk regularly ... this time of the year I walk every evening just as the sun’s going down, for 20 minutes... I go for the walk according to the sun. At the moment I’m going about a quarter past seven ... and then (in winter) it will get to the point where I’ll go before dinner.
Daylight savings also have to be accounted for in the data. In this study, because data showed clock time, motion and power use moved backward 1 hour as a result of the change. Clock shifts maintained scheduled activity at the same clock time, but caused an artificial shift in the time of sunrise and sunset, which must be allowed for in interpreting data and establishing alarms.
These cyclical annual changes affect routines and, consequently, the interpretation of sensor data. Factoring in an understanding of changes in routines that vary annually based on weather and seasonality will further help reduce false positives.
Analysis of trend data exposed a gradual change. Using linear regression, and controlling for changes in temperature based on daily maximum temperature, the results showed a deterioration in Rupert’s sleep quality over the period of the trial. There was a significant increase in bedroom movement at night (beta
After cyclical and seasonal changes have been accounted for and preference for routinization and adherence to routine taken into account, irregular variance in data is most likely the result of unexpected events, including illness, acute medical events, and falls. For ad hoc alerts to be effective in early detection and intervention, they need to be reliable. Analysis of the data showed the effect of variability of routines and looked at ways to establish customized thresholds to reduce false alerts.
To set up effective ad hoc alerts based on changes to routines, 2 conditions need to be met: activities should occur at the same time every day (consistency of timing), and routines should be a regular daily occurrence (adherence to routine). Examination of the data in this sample revealed that motion sensor data was more diffuse in time and therefore less reliable than power use data. Residents’ radar plots showed that kitchen power data and specifically breakfast power data were the most reliable, both in consistency of timing and in adherence to routine. Lunch and evening meals were less reliable indicators because residents reported regularly dining out as well as varying the time of evening meals according to what was on television. Knowing a person’s routine provides valuable insight into understanding variance in data, and together with radar plot, choosing appropriate triggers for alerts; for example, Jacqui revealed that:
On the weekend ... I’ll either have, bacon, eggs and ... during the week I usually have cereal, fruit, yoghurt, yeah and toast, always toast.
Routines used to trigger ad hoc alerts should also be a regular daily occurrence. Adherence to routine was calculated based on the number of days when key appliance sensors fired within ±30 minutes of the median time of use (
Adherence to breakfast routines.
Description | Jacqui | Rupert | Elizabeth | |||
Kettle | Kitchen | Microwave | Kitchen | Kettle | Kitchen | |
Days absent | 47b | 45b | 1 | 1 | 11 | 10 |
Days with dataa | 98 | 119 | 156 | 173 | 145 | 145 |
Days breakfast not taken | 36 | 17 | 8 | 3 | 25 | 8 |
Total days with data | 181 | 181 | 165c | 177c | 181 | 166 |
Adherence to scheduled, n (%) | 62 (63.2) | 102 (85.7) | 148 (94.8) | 170 (98.3) | 120 (82.8) | 137 (94.5) |
a± 0.5 hours of time of the median calculated on all power use between 5 am and 9 am .b50 days absent, but on 4 of the days, breakfast was prepared on day of departure.cData missing because sensor not installed till after start date.
dDaily adherence to schedule based on median ± half an hour.
The case studies also show 2 examples of ad hoc events that can be clearly traced by changes to routine, illness, and death of spouse. Someone who is ill may stay in bed, stay seated for longer, and forgo regular meals. These changes are detectable with motion and power sensors. Additionally, self-reported information from residents or relatives can narrow down how residents respond to being unwell. Rupert’s daughter-in-law, who has remote access to his daily movement data via a secure website, was able to identify when he was ill because of the reduced movement in the kitchen:
I look at the kitchen and I know when he’s sick because it’s down, the usage (motion data) ... And I know definitely when he’s sick he doesn’t cook for himself.
The first example is an example of illness. On Wednesday and Thursday of Week 3, it appeared that Rupert was unwell. Rupert’s normal routine for the same 2 days of the preceding week is shown in
The second example is a more persistent change in daily routine. Elizabeth lost her spouse around the time the sensors were first installed. Traumatic events can result in a temporary loss of routine [
The main challenges for using sensor data as ad hoc alerts are individual’s preference for routinization, adherence to routine, and consistency in timing. Understanding these variations and the cyclical and seasonal variation in timing of routines is important in determining the thresholds for alerts. Having a two-sensor alert (power and motion) could be additional security against false positives and technical issues. Additionally, it is possible to calculate the longest time spent in the room such as the bathroom and the kitchen, rooms, that have very specific purpose, and to set up an alert when maximum length of stay is exceeded as another opportunity for cross-checking alerts related to change in routines.
Breakfast preparation activity by participant. Appliance power use frequency of firing by time between 5:00 and 9:00. Only the appliance that is most consistently used for breakfast is illustrated. Boxes show first to third interquartile range (IQR), with the line of separation indicating the median and the diamond the mean. Whiskers indicate the minimum and maximum and outliers are represented by asterisk for values that fall between 1.5 and 3 IQRs and circles represent outliers that fall outside the 3 IQR.
Rupert: Normal household movement Tue & Wed Week 3.
Rupert: Changes in motion due to (assumed) illness. Unusual houshold movement attributed to illness Tues & Wed Week 4.
Elizabeth: Changes in lounge movements on bridge club days (Mondays, Thursdays, and Saturdays) 1, 2, and 3 months after the death of her husband showing a return to routine (resumption of bridge club) around Month 2.
This study demonstrated how a routine-based approach could help make sense of the large amounts of data from in-home sensors in such a way as to make the interpretation of data easy, intuitive, and transparent. The routine-based approach can be used to detect both change in sensor data and absence of sensor data for routines that take place outside the home, facilitating the use of alerts for both ad hoc events and gradual change over time. Importantly this study advances understanding of how sensor data can be used in real-world applications for real-time monitoring through an improved understanding of normal variation in routines and behavior (cyclical, seasonal) and a better understanding of the need for customized setups to allow for inter- and intraindividual differences in preference for routinization and in timing of and adherence to different routines. Overall these results support the use of in-home sensors as an extension of existing health care prevention approaches, one which, because of the 24/7 nature of the data, has the potential to facilitate aging in place.
This study builds on earlier work by Kaye et al [
Routines are an appropriate framework for understanding sensor data because they account for the cyclical nature of everyday living and because changes to routine, including the timing, are important in understanding both physical and psychological well-being [
This research also reflected on the challenges of interpreting sensor data for the development of practical applications based on single-point comparisons. Viewing data as cyclical increases the reliability of single-point comparisons by facilitating comparisons of like for like, that is, comparing routines on any given day with the matching activities of the same day in previous weeks. Challenges for interpreting data include allowing for predictable and unpredictable events arising from the local context. First, cultural routines and seasonal change are contextual factors that shape activity over time. Allowance needs to be made for cultural routines such as weekends, holidays, religious practice, as well as seasonal changes of temperature and daylight [
Importantly, prior knowledge of routines and variations in routine personalizes information that is otherwise just numbers. This is especially important if the sensor data are to be used by community services providers who can better know the individuals they care for through the everyday occupations and routines that give structure and meaning to life [
Future research should look at testing algorithms to monitor real-time daily activity using data from the same day of the week to allow for normal weekly variation in routine and, based on a 4- to 6-week moving window, to allow for changes in seasonal movement and activities. Daily weather data would need to be incorporated, such as profiles to allow for unseasonal above or below average temperatures as well as precipitation, to account for normal weather-based changes to routine. Data from different sensors should be correlated, for example, kitchen appliance use and motion in kitchen to detect battery failure or technical faults and minimize false alerts. Incorporating room layout and location of exits can additionally allow for interpretation of movement around the home. A two-level alert system, low and high, based on deviation from normal routine could be established. Alerts should be accompanied by a summary of the issues: routine versus actual and a time line of activities. Changes in behavior over time for comparable seasons can look for changing trends.
The user interface should display the 24-hour clock showing movement for key rooms (bedroom, kitchen, bathroom, lounge).
Screenshot of user interface showing daily bathroom movement over a period of 4 months using a 24-hour clock format.
In-home sensors may be able to facilitate aging in place and improve community health care services through the provision of alerts and notifications for early intervention as well as longitudinal health data for improved decision making. They could also potentially complement and reduce the direct costs of care in the community by facilitating routine monitoring. For smart homes to be a reality, it is important to strike a balance between the privacy and independence of the elderly and the effectiveness of monitoring systems. Furthermore, it is important that in-home monitoring provides data that allow carers and relatives to engage with and understand how these technological solutions can enhance care for the elderly, and that these systems are not viewed as a substitute for face-to-face health care, but as a means of improving the effectiveness of current interventions.
Although research based on case studies is not generalizable, the results nonetheless provide rich insight into how routines can be used to monitor data in smart homes. There also are limits to the interpretation of data, especially from motion sensors. Most problematic is dual occupancy because sensors cannot distinguish between individuals, and neither can they separate out the activity of visitors to the home. However, new motion sensors under development may overcome this problem. Sensors cannot distinguish between types of motion. Therefore, matching specific daily activities such as getting dressed or doing daily exercises can only be inferred by time of day and room use. Equally because lack of motion could be due to absence from home or a fall, accurate and early fall detection is problematic for motion sensors [
Concerns about ethics with respect to informed consent for technology-based research in the elderly remain problematic. There were some erroneous perceptions by residents who did not fully understand the technology, despite repeated explanations. This was mitigated by involving the relatives of residents in the process. However, it is imperative for smart homes pilots and installations for the elderly to continue to clearly communicate what the technology can do and what it cannot do, to better allay concerns and manage expectations.
This study demonstrated the usefulness of a routine-based approach for making sense of smart home data and furthered the understanding of the challenges that need to be addressed in order to make real-time monitoring a reality, through improved visualization of data and a better understanding of variation in routine, which could improve the effectiveness of alerts and notifications. Future research needs to explore larger longitudinal datasets to test the potential of routines for alerts in order to minimize false positives and negatives and still be able to deliver reliable alerts to the satisfaction of persons being monitored and their carers.
Animation showing 4 months of daily bathroom movement data on a 24-hour clock format.
Australian Eastern Daylight Time
Australian Eastern Standard Time
Greenwich Mean Time
Kilowatt hour
None declared.