The following section presents localization and frailty assessment methods.
Indoor localization is usually handled using RF (Radio Frequency) signals and protocols, such as Bluetooth, WiFi, RFID, UWB (Ultra-Wide Band) and ultrasound [5
]. Bluetooth beacon-based solutions have especially gained attention recently, due to their relative low cost and minimal requirements in signal receivers, since widely used smartphones or smartwatches can readily be used as such. Beacon-based systems are effectively used in commercial applications, such as navigating users in shopping centers, museums, airports, etc., with their main purpose being offering location-related advertisements. However, user localization in such systems often requires the exact floorplan of a facility, a large number of installed beacons, or large training datasets, which increase the total installation cost [7
]. Medical applications can benefit from beacon-based localization. However, in order to achieve wide applicability, low-cost solutions, with minimal training need to be employed.
Theoretically, there are two main categories of beacon-based localization methods: the ones based on trilateration/triangulation and the ones based on fingerprinting. Trilateration
is the process of determining the location of an individual by measuring its distance to at least three fixed points of known positions, such as the positions of three beacons. The position of the individual is computed by solving a least-squares problem [8
]. In case of Bluetooth beacons, distance is usually measured in terms of the RSS (Received Signal Strength) value of each beacon, as measured from a receiver (e.g., smartphone) at the position to be estimated. Current trilateation-based methods focus on increasing the accuracy and efficiency of solving the main least-squares problem, through incorporating long distance shadowing [10
], employing multiple iterations [9
] or utilizing ring-shaped overlapping regions [12
methods, instead of being based on measuring the distance from fixed points, work by computing angles from reference points [8
], based on the time-of-arrival (TOA), the time-difference-of-arrival (TDOA), or the roundtrip-time-of-flight (RTOF) of the radio signals. Triangulation is being used for GPS positioning; however, their application for indoor localization is limited, due to the low capability of smartphones to perform the above types of measurements.
A different concept is adopted by fingerprinting
methods. These methods are based on collecting off-line RSS measurements (fingerprints) from fixed beacons at multiple reference points in the indoor environment. Later, during the on-line localization phase, the current RSS measurements of the tracked device from the fixed beacons are compared to the collected database of fingerprints and the current location is estimated by considering the closest fingerprint matches, as seen in Figure 1
]. Classification techniques are utilized in order to perform fingerprint matching, such as k
-nearest neighbors (kNN), support vector machines (SVM), Naïve Bayes [15
] and neural networks [16
]. Fingerprinting methods have the advantage of not requiring the knowledge of the positions of the fixed beacons. However, for exact localization, a large number of reference training points needs to be collected. Apart from the above broad categories, other methods include fusing measurements from multiple sensors available in smart devices, such as magnetometers, accelerometers, gyroscopes and ambient light sensors, to perform localization [17
]. Measurement fusion is often accomplished through Kalman filters [18
Regarding room-level accuracy localization, in [19
] a beacon threshold-based method is presented, where the dimensions of the rooms are used in addition to RSSI measurements in order to improve accuracy. The proposed method uses the dimensions of the room in order to create thresholds for the RSSI measurements defining if someone is inside or outside of a room. It is not a fingerprint-based method and shows an improvement in the accuracy compared to the situation where only the highest RSSI measurements were used to estimate the room where a person is located. The disadvantage of this method is that the dimensions of each room are needed; moreover the beacons should be placed at the ceiling of each room, making difficult the installation procedure of such a system in multiple houses. In [20
] an In-Room presence detection system is presented. The system uses beacons in the doors in order to recognize when a person enters the room. Specifically, one beacon with motion sensor is placed at the door of a room and the other beacon is placed inside the room close to the entrance. In case of door motion, the system starts to record RSSI measurements from the two beacons, while the user is carrying the smartphone. Examining the differences between the two recorded RSSI time-series, the method is able to recognize when someone enters or leaves the specific room. The proposed system achieves high accuracy, near 99%; however, it assumes that each door is always closed when someone wants to enter or leave a room. In addition, motion sensor beacons and a minimum requirement of two beacons per room are needed, significantly increasing the total cost. The proposed system is more suitable for large indoor environments, such as workplaces, event attendance monitoring, etc. and not for houses, where the doors between rooms remain open for most of the time. In [21
], a Bluetooth fingerprint-based room localization system is presented, where additional signal features are used among the RSSI measurements, such as the Link Quality (LQ) and the Cellular Signal Quality (CSQ), using a combination of Bayesian statistics and Support Vector Machines for classification. Moreover, the proposed combination in [22
] is used in order to improve the accuracy of [21
]. In our case, such additional measurements need specific hardware and cannot be extracted from BLE Bluetooth signals. Table 1
presents a summarized comparison of the previously mentioned room-localization systems with the proposed system. We compared the systems based on the ease of installation, the accuracy, the cost/hardware availability, which refers to the easiness of finding and purchasing the specific hardware, and the system’s suitability for house environments.
Localization-collected data are used in correlation with other health-related data in e-health platforms. In [23
], a medical platform focused on people using wheelchairs is presented. The platform introduces new hardware approaches such as the smart wheelchair with localization and falling sense abilities. The platform monitors, among others, the heart rate or the ECG time-series. Moreover the platform presents a novel social network to improve the efficiency of resource sharing. The FrailSafe project [24
] focuses on better understanding frailty and its relation with other health conditions and indicators. It provides cloud platform functionalities for the monitored person, the doctor and the relatives. Frailsafe monitors older people’s health condition variables, indoor, outdoor and social media activity. The collected data are used in order to find new non-intrusive ways for assessing older people’s frailty and health status such as tablet games. Moreover, it provides automatic suggestions and interventions for individuals. The ACTIVAGE project [25
] is an IoT cloud platform-based project, where daily behavioral activity monitoring of older people is performed, using door and motion sensors. Health indicators such as blood pressure or glucose levels are monitored as well. Alarms and notifications are provided in urgent situations. ACTIVAGE focuses on supporting and extending the independent living of older adults in their living environments, and responding to real needs of caregivers, service providers and public authorities.
2.2. Frailty Monitoring
Activity-related information, such as monitoring indoor movements, are closely related to frailty evaluation of individuals. Although frailty is often assessed through metrics of cognition, function, social health, medication, co-morbidities, etc. [26
], common frailty evaluation measures, such as the Fried index [2
], are also based on physical activity and mobility to determine the frailty level. Activity-related measures, such as the IADL (Instrumental Activities of Daily Living), have been used effectively for frailty assessment [29
]. However, this type of information has mostly been collected through clinical evaluations performed by clinical personnel, which has disadvantages, such as infrequent data collection and increased cost.
Automatic real-time monitoring of the older person’s activity can significantly assist in frailty assessment through continuous monitoring of older people without the need of clinical personnel presence. Monitoring methods have utilized wearable systems [30
] and gait speed recognition [31
] for activity recognition and frailty risk assessment. Recent works have used motion identification through accelerometers and gyroscopes [32
], as well as extraction of activity patterns through ambient sensors for unobtrusive detection of unusual activity [33
], holistic frailty assessment [35
] and rehabilitation [36
]. Mobiliy patterns of the individuals have been usually extracted through motion sensors [33
]; however, localization-based methods such as the ones mentioned in Section II present a promising alternative of lower cost. The discriminating ability of activity and mobility information to detect frailty levels has been assessed through recent studies [37