The experimental study reported in this paper, however, includes 12 participants with a median age of 89 years. Three participants were later excluded due to unexpected life events. We used a statistical method to classify participant behavior into normal and abnormal (anomalies). This was based on whether the amount of time spent on an activity in each room was low or high in duration. The experimental setup worked according to routines collected during the interviews. The main focus was the meal activities.

The iVO architecture is a layered architecture with horizontal connectivity of different sensors and service providers. iVO architecture is built using FIWARE [18], connecting off-the-shelf sensors and IoT devices via an IoT platform. The platform is called societal development through secure IoT and open data (SSiO) [19]. The SSiO platform was designed and implemented for different IoT applications and services within a smart city, a detailed description of the iVO architecture being given in Saguna et al [20]. The installed sensors are connected via gateways to a service provider (an iVO project partner) to push the sensor data into the SSiO platform, as shown in Figure 1a. The iVO analytics component is shown in Figure 1b.

The project’s first pilot study for the selected age group has been described by Saguna et al [20]. We identified targeted activities through the interview process. The interviews also provided information on the needs of participants and their relatives and activities of interest. The interviews were conducted with the participants along with their relatives who answered questions on the activities they would like the model to recognize, their general opinion of home monitoring, their expectations of implementation, and details about their daily routines. All interviews were documented and recorded by researchers. All participants gave their written consent to use their data.

Most of the participants and their relatives expressed their interest in monitoring meal activities in the kitchen. A matrix of the activities of each participant was constructed based on the environmental setups in each household, on the number of sensors, available kitchen appliances, and the interview responses. The interviewees expressed an interest in receiving both negative and positive notifications. Negative notifications notify that anomalies have been detected in behavioral patterns and positive notifications notify that the activity behavior is normal. This is shown in Table 1.

The ethical principles raised by and applied to the project were considered in collaboration with the department of homecare at Skellefteå municipality and were approved by the regional ethical committee. The participants gave their consent for the use of their data and the installation of in-home sensors. The project was, overall, in compliance with the European Union’s General Data Protection Regulation guidelines [21]. The data collection and processing included in this study were approved by the Regional Ethical Board in Umea, Sweden (diary no. 2018-189/31).

iVO smart homes use a wide range of off-the-shelf IoT devices and sensors, these systems also being referred to as dense-sensing network technologies [22]. This study, however, only looked at motion sensors, wall plugs, and smart water meters. The sensor installation and floor plan of an older person’s home are shown in Figure 2. All nine apartments have a similar floor plan. Data cleaning is an essential part of the first phase of the study implementation. It is unavoidable that sensors will fail, sensor readings will be lost, and sensor data sets are duplicated, leading to vagueness and imprecision, false positives, and false alarms [23]. This is of even greater importance in older person care health monitoring systems [24]. A reliable monitoring system, therefore, needs to be built before the feature engineering phase is begun. Data cleaning applies, in particular, to motion sensors, redundant data being removed, and missing reading values being identified. Outliers due to visitors or home care visits were excluded.

In this module, we calculated daily duration and visit frequencies for all rooms to extract patterns from and analyze user behaviors. Daily durations are further processed into hourly-based durations. Our analysis used historical data sets from the 2-year data sets. Scalability and re-usability are, however, notable challenges to building individualized activity models [25]. Features engineering is the first step in this module, the time duration feature being the length of time the person spent moving or not moving in the room until transiting to another room/location. The frequency of visits is the number of times the person transits to another room and returns. We used a fixed time window, which is specified by the start and end times of the routines described in the interviews. These time windows, described in Table 1, are used to classify data into normal and abnormal activities.

Hawkins has defined an outlier as “an observation that deviates so much from other observations as to arouse suspicions that it was generated by a different mechanism” [26]. Detecting anomalies in data has been studied in the statistics community since the 19th century [27]. There are different approaches to detecting anomalies, including the mining-based approach, the logic-based (rule-based) approach, the ontology-based approach [28], and the statistical-based approach [16]. A statistical parametric model can be a simple approach to anomaly detection problems, assuming that the data is normally distributed, fits certain distributions, and that the value of these parameters is unknown. These must be estimated from the given data. Selecting the correct statistical tool for anomaly detection, however, requires the validation of a normality assumption [29].

Normally distributed data can often be tested using histograms. A histogram may not, however, reveal the shape of the distribution. The selection of the normality test tool was based on a comparative study of different normality tests. This study showed that the Shapiro-Wilk test was the most suitable tool for data sets of sample sizes of between 50 and 2000 [30], the null hypothesis in this test being that the sample is a normal distribution. This hypothesis is rejected if P values are less than .05 (95% CI). Results of the tests can be found in Table 2.

The normality test results show that 12 out of 13 (92%) tested meal activities in our study do not show a normal distribution. We, therefore, selected Chebyshev’s inequality theorem, a nonparametric statistical method, to detect anomalies [31] in user daily activities based on time duration. Chebyshev’s inequality constructs the upper and lower interval for the percentage of the data that falls outside of k standard deviations from the mean. It holds no assumptions about the distribution of the data and can be used in situations where at least 75% of the data is within 2 standard deviations of the mean. This can be more than 75% in some cases.

The inequality in Equation (1) calculates an upper bound ¼ for the probability of random values exceeding (k) 2 standard deviations from the mean. We, therefore, define an outlier as a data point of a time duration in minutes or hours that exceeds the expected duration by 2 standard deviations [13].

P(|X-μ| ≥ kσ) ≤ 1/k²(1)

X is the random variable, μ is the mean, and σ is the standard deviation.

A value for p, the significance level of the intervals, is 0.25. This determines which data are potential outliers. To find k using Equation (2):

k = 1/√p (2)

The significance level is 0.25 with a confidence level of 75% and a k value of 2. Equation (3) can be used to determine the probability that a randomly selected value is in the interval, around 75% of observations falling within 2 standard deviations of the mean.

μ ± kσ (3)

Durations that lie between the lower and upper thresholds are considered to be normal. Data from wall plug sensors was also processed and added to the notifications as additional context. We, however, restricted the model to the duration feature, wall plug sensors, and visit frequencies not being used in the anomaly detection classification process. Our observations showed duration to be the most important feature, duration showing how long a participant remained in a room to perform an activity.

Types of anomalies with example situations:

Our implementation delivers positive or negative notifications to relatives and caregivers. Positive notifications represent normal days, and negative notifications represent anomalous days. The notifications were sent via SMS [15]. Negative notifications are sent when activities deviate from the observed normal behavior, which is based on analyzed historical data. Positive notifications are sent when the user’s behavior is normal and when the relative expresses an interest in receiving such notifications. The design of the notification service process in Figure 3 starts by reading a configuration file, which automates a schedule and starts the service process.

We contacted the relatives before running the system to inform them of the activities they would be notified about. They were shown a notification format and the information that would be included. They were optimistic about the possibilities that the system could provide, the system providing comfort through being able to remotely check in on their parents daily. A consensus was reached with participant relatives on the content of the notification and the type of communication to be used (ie, SMS).

The notification structure included the number of minutes spent on the meal activity, across how many visits, whether they used any appliances, and the normal duration of the activity for that participant based on their identified patterns. An example of the structure and information for one positive notification sent for apartment 1 is shown below:

The relatives were given the option to reply to every notification. This response included feedback when there was anomalous behavior.

Our study demonstrates, through nine cases, how a data-driven approach and longitudinal data from interviews can convert large amounts of sensor data into knowledge. Our approach can detect anomalies in ADLs and utilize notifications to alert relatives of these anomalies. An important aspect of this approach is that it can facilitate the interpretation of data from real-world smart homes and use this in real-time monitoring systems to identify activities that deviate from the normal patterns of older persons. This enhances our understanding of personalized setups and different individual daily routine preferences. The results support the use of data from off-the-shelf sensors and IoT devices (installed in real homes) and the improvement of health care services by feedback to caregivers in near real time. This further enables older persons to live independently in their homes for longer.

We used, in our research, statistical methods to detect anomalies, an approach that has been used in other studies [13,16,32]. The statistical methods we used to detect anomalous behavior in ADLs are in line with previous studies. We, however, in the interests of accurate interpretation, tested the assumptions of data normality. Konios et al [32] recruited 30 volunteers aged 30 to 45 years to conduct a study of the recognition of low-level activities such as steps in preparing and drinking tea/coffee. They used the mean ±1.5 standard deviations as the threshold for classifying normal and anomalous events. Another similar study [16] used low-level activities from the annotated CASAS public data set [33]. The mean and ±1 standard deviation were used to classify activities as normal and abnormal behavior. The mean ±2 standard deviations were, however, used in [13] to detect anomalous behavior at the 95% CI, thresholds being calculated using 2 months of history data of students’ working activity. Our method uses Chebyshev’s inequality, thresholds being based on intervals in which only 75% of the data is within 2 standard deviations, 95% being within 2 standard deviations in a normally distributed data set.

Our results are distinct in that they were collected in an uncontrolled environment and used real-world data of older participants in their homes. Konios et al [32], in contrast, used a lab setting; Jakkula et al [13] used synthetic data and one day of real data from a lab environment to validate their approach. The work in Paudel et al[16] was implemented using annotated public data sets.

Our overall approach builds on earlier work [12,14,34], all being real-world implementations. The data collection process, sensor setup, the method used for pattern identification, and the behavior of older persons, however, differ. Beunk et al [12] aimed to visualize sensor information, duration, and start time from log data, notifications being sent to caregivers, participants, and relatives using real-world monitored activities of 5 participants aged ≥ 65 years. Kasteren et al [14] used 180 days of real-world data obtained from power usage, motion sensors, and interviews to carry out a behavioral analysis of the 3 participants. The distribution of daily activities aggregated over multiple days was visualized using radar plots. Interviews and motion sensors were used to identify residents' sleep and daily movements. Another longitudinal study [34] evaluated the usage of unobtrusive technologies in detecting a change in activities and cognitive decline by statistically analyzing 200 days of data on the daytime and nighttime activities of 233 senior participants with a mean age of 83 years.

Our study proposes a solution that detects anomalies in different types of activities in real-time. Our approach is validated by comparing longitudinal data collected from interviews with observed patterns collected from historical data (80-355 days of sensor data). This was processed to classify activities into normal and anomalous behavior and so allow notifications to be generated.

Our findings show that our system can benefit relatives and can be used by formal health care providers in Skellefteå municipality. Relatives and caregivers received timely notifications of older participants’ activities. The system was also able to distinguish between normal and anomalous behavior, which can be used to detect long-term changes in routines and which can signal the early stages of cognitive decline. We believe that this type of system can have a direct impact on the enabling of older persons to live independently for longer. Feedback collected on our system (Table 5) was mainly positive and reflects a high level of user satisfaction with the iVO service. Some participants showed an interest in continuing the service after the trial period ended.

Our approach is limited to the use of duration to identify anomalies. The root cause of the anomalies can, however, only be known via feedback from relatives. A direct method for establishing the root causes of anomalies would help us to understand anomalies better. This approach would, however, make our system too intrusive, interrupt the day-to-day lives of participants, and have an impact on their normal routines. Our preference is that relatives communicate with participants about anomalous behavior and that feedback to our system is then provided by relatives. The use of Chebyshev’s inequality is another limitation. Thresholds are based on loose intervals with only 75% within 2 standard deviations. This can be compared with a normally distributed data set where 95% is within 2 standard deviations. These loose intervals mean the thresholds are wider apart and can therefore lead to fewer anomalies.

We developed an SMS-based notification service based on information provided in interviews on the needs of participants and their relatives, on the activities they were interested in, and on data from the off-the-shelf sensors and IoT devices installed in homes. We also conducted user evaluations. This service acts as an extension of the municipality health/social care services and helps older persons to live in their homes independently. We proposed, developed, and implemented an anomaly detection framework for the recognition of anomalous daily activities of older persons living in single resident smart homes, using the real-life uncontrolled setting of 9 older participants. This paper proposed a probabilistic approach to the temporal analysis of ADL of nine older participants in a real-world environment. The method introduces a way of indicating whether the probability of a performed activity is considered to be normal or anomalous behavior using duration.

Our system observed approximately 1000 meals and bathroom activities. Notifications were also sent to 9 apartments between July and August 2020. Four hundred twenty-one notifications (59%) out of the observed meals activities (716) were sent to 8 apartments on each meal activity, with our approach considering more than 75% (537) of observed meal activities to be normal. This figure was, however, in reality, 93% of meal activities, these falling within 2 standard deviations of the mean and so corresponding to their normal behavior. The behavior patterns derived from the historical processed sensor data closely match the routines participants reported in the interviews. We received positive user experience feedback on the service from 5 out of 9 participants’ relatives (55%) and an average of 4 points on a 1-5 satisfaction scale. The results ultimately support the use of IoT devices in homes as an extension of health/social care services, which can, in turn, increase the opportunity to age in the home independently.

We will, in the next phase of this project, use advanced data analytics methods to further investigate the data. We would include other types of anomaly classes in this, such as transitions or visits of unusually short durations, which can indicate unrest and the detection of an unusual presence. Other types of context sensors and features also need to be considered in the analysis. Contextual information such as weather conditions and holidays could, for example, improve results and reduce false positives in detected routines [45]. Future research can furthermore test algorithms that monitor real-time ADLs using data for that specific week, which will allow for weekly variations in routines. Data based on a monthly moving window would also allow for seasonal variations in routines. We are working towards using reinforcement learning as a multi-armed bandit problem (MAB). We plan to conduct follow-up interviews for the study as part of a workshop to communicate our results and understand further needs. Direct feedback from older persons is another important aspect. This would, however, increase the cognitive load upon them. There was no appropriate communication channel for feedback on the alerts or activity updates sent to relatives by older persons, which was not implemented due to the increased burden of daily feedback on older persons. This type of feedback was mainly gathered via the relatives, who communicated with their older relatives about an anomalous notification. This will be addressed further in our future research.