In this section we provide an overview of state-of-the-art assistive technologies focusing on three main directions: objective monitoring using sensors devices, data analysis and machine learning for behavior assessment and finally delivering interventions using social robots. In the end, we discuss for each type of technology the limits, challenges and adoption from the perspective of older adults.
2.1. Monitoring Daily Life Activities
The development and deployment of sensors and smart devices for monitoring older adults’ activities is a dynamic research and innovation area currently triggered by the advent of big data and machine learning technologies. On top of the monitoring devices, applications are developed for assisting the older adults’ healthcare, improving the social communication and inclusion or for managing various activities targeting to prolong independent living. In [10
] the most important dimensions of older adults’ lifestyle that need to be properly monitored are identified as: the physiological parameters and the body functions, the changes in the daily life activities (DLAs) and daily routines, the social factors and the environmental factors. In [11
] and [12
] the data to be monitored and collected for older adults’ support and intervention is classified in behavior, health, occupancy, lifestyle, home automation, environmental and personal safety. Similarly, in [13
] the most common data types to be monitored and used in AAL systems are identified as: behavioral habit data, physiological information, healthcare information and environmental data.
On other dimension, there are several types of sensors which can be used to monitor and analyze different aspects of the older adults’ healthcare and well-being, most frequently being classified in: (a) physical sensors, (b) virtual sensors and (c) logical sensors.
Physical sensors are the most frequently used type of sensors in collecting data about the older adults’ healthcare and well-being. Table 1
presents an overview of various AAL physical sensors and the type of monitored data that was identified in the research literature [14
]. These research approaches are classified and discussed below.
Wearable physical sensors are used to collect the data about (i) the physiological signs of a person (e.g., heart rate, blood pressure or temperature, blood cholesterol, blood oxygen saturation, respiratory rate, blood sugar level, etc.), (ii) the posture, gait and movement, (iii) the social interaction or (iv) the activities performed by a person during a day. All these data are analyzed to assess the health/wellbeing status of a person [15
], the sleep quality [20
] and the level of stress and cognitive decline [24
] or to detect the falls [27
]. Currently, the most important wearable sensors that can collect this type of data are embedded into wristbands, smart clothes, smartphones or smartwatches, arm bands or chest straps.
Monitoring the physiological signs and physical position provides valuable information regarding the health status of a person. Some of the main benefits provided by the wearable sensors in monitoring the health status of a person are the following [28
]: (i) allow to continuously monitor the vulnerable patients, (ii) allow the medical specialist to have a better view on the patient’s health issues and to make the most accurate diagnosis and (iii) offers the patients the opportunity to evaluate their health condition and consult a doctor when appropriate. In [15
] the data collected to evaluate the health status of a patient is: physiological signs and physical position. The types of sensors used for collecting the data are airflow sensor, galvanic sensor, body temperature sensor, blood oxygen sensor, electroencephalogram (EEG)/ heart rate sensor and digital accelerometer. An Internet of Things (IoT)-based system for monitoring the health status of Cardiovascular disease(CVD) patients consisting of a sensing layer responsible for collecting the physiological signs of the patient (e.g., blood pressure, electrocardiogram (ECG), amount of oxygen in the blood (SpO2)) and the patient’s location is proposed in [16
]. Custom wristbands that integrate specific sensors (i.e., piezoelectret sensor) that can detect heartbeats are also proposed [29
]. The piezoelectret sensor can help detecting the pulse waveform, which is similar to the one provided by an electrocardiogram. Similarly, arm-wearable ECG sensors are researched for monitoring the heart rate that integrates the main components of an ECG [17
]. The sensor provides two functions, namely a monitoring function in which the signal is acquired and a Holter function in which the signal is stored in the internal memory. Authors of [18
] develop a smart t-shirt for recording the ECG signals which integrates a set of active electrodes and an ECG Portable recorder. The accuracy of ECG signals recorded with the developed device was comparatively evaluated with a standard 12-lead Holter. A system for monitoring in real time the CVD patients’ heart rate, blood pressure and body temperature which of multiple wearable sensors and collects medical information about patients with CVD, is connected to a web portal for displaying the acquired data from patients in [19
Sleep quality is an important factor for a person’s stage of health because it reduces the risk of developing chronic diseases and mental disorders. Most state-of-the-art approaches for monitoring sleep quality focus on comparative analyses and evaluation of the commercial devices available on the market [21
], while only some of them present academic research solutions [20
]. For example, in [21
] commercial devices for monitoring the sleep quality are analyzed, the Up Move Jawbone (U) and the Withings Pulse accelerometers (monitor sleep duration, how many times s/he wakes up at night, etc.), the Bodymedia SenseWear Pro Armband actigraph (monitors the sleep quality by estimating the sleep duration and efficiency) and the home-polysomnography (collects data about thoracic and abdominal movements, airflow, etc.). Similarly, authors of [22
] evaluate five wearable devices for recording the sleep quality, namely the Basis Health Tracker, the Misfit Shine, the Fitbit Flex, the Withings Pulse O2 and the Actiwatch Spectrum actigraph. The Basis Health Tracker is an actigraph embedded in a wristwatch, while the Misfit Shine is a sleep-tracking device provided with a strap to be worn on the hand. The Withings Pulse O2 is also a sleep-tracking device worn on the wrist, while the Actiwatch Spectrum is a wristwatch with an embedded accelerometer. In [30
] the accuracy of a commercial device, namely, Fitbit Charge 2, compared to polysomnography, in measuring the sleep and the wake states is analyzed. This device can detect, besides the sleep and wake states, the time spent awake, in light or in deep sleep as well as to track the level of daily activities. In [23
] it is proposed a wearable device for monitoring the abdominal and thoracic respiration that can be used to detect the obstructive sleep apnea by monitoring the breathing status. Chest-worn sensors collect data about the posture (during the day or the sleep), the position changes and body acceleration of a person in order to detect the sleeping and the waking periods to detect the quality of the sleep during the night and the time spend by the person in bed [20
]. In [31
] wearable sensors positioned on arms and chest are used to monitor the sleep quality. The sleep quality is evaluated based on the sleep posture and the sleep stages (awake, rapid eye movement (REM) sleep and non-REM sleep cycles).
Monitoring the level of stress is very important because, on long term, it has a negative impact on the cognitive functions of a person. Also, by monitoring the cognitive function, valuable information is obtained that can be used to improve the daily/detect early cognitive decline in the case of older adults. In [24
] the physical attributes (e.g., galvanic skin response and skin temperature, EEG heart rate, respiration rate and voice data) are monitored and collected in order to assess in real time the emotional, physical and mental stress of a person. Data are acquired with a set of wearable biosensors integrated in smart clothing and with a microphone. In [25
] a smartwatch which integrates a Global Positioning System (GPS) sensor, an ambient light sensor and an acceleration sensor, is proposed as a solution for monitoring wandering risk for older adults suffering from dementia. The developed smartwatch acquires the following types of data that are used to avoid the risk of the patient losing their way when they experience wandering episodes: the older adult position and the route in which s/he moves that is necessary to identify if the older adult moves outside the area considered to be safe, the sunlight exposure duration and the number of older adult’s steps. In [26
] the benefits of the physical activity on the cognitive functions of the older adults are inferred based on (i) the data (e.g., light/ moderate to vigorous physical activity) collected with an accelerometer and (ii) a set of questionnaires containing information about the cognitive ability of older adults. A platform for monitoring older adults at home in order to detect the early cognitive decline is researched in [32
]. It integrates Infrared (IR) motion sensors with magnetic contact sensors and sensors for monitoring user computer interactions (e.g., mouse movements, computer usage time, etc.) and uses all the collected data to build the profile of changes in the activities performed by older adults, which are than analyzed in order to detect possible cognitive decline.
In the case of fall detection, most approaches from the research literature are based on accelerometers and gyroscopes. Accelerometers can provide information about the motion data during daily living activities (i.e., walking, running, stepping and falling), as well as the instability that can appear during these activities and can provide valuable information in predicting the risk of falling during walking [27
]. Usually a wearable sensor feds data to a fall detection algorithm to identify the fall risk within the daily activities of a person [34
]. In [35
] the data about acceleration signals, postural instability and falls are recorded with a tri-axial seismic acceleration sensor placed on a belt while [36
] presents an approach for detecting the fall, based on a smartphone which integrates a tri-axial accelerometer and a tri-axis gyroscope and is used to record the acceleration patterns, in the case of older adults, for the following types of daily activities: sitting, lying, jumping, running, walking and hitting the sensor. A fall monitoring system which integrates a portable sensor that can be placed on the person’s shoulder, waist or foot and a mobile phone is proposed in [37
]. The portable sensor records the data with a triaxis accelerometer, a triaxis gyroscope and a triaxis magnetometer and sends it to the mobile phone. The collected data are used to compute the acceleration and Euler angle which are provided as input to a fall detection algorithm.
In contrast with wearable sensors that need to be placed on the body of the person, ambient (non-wearable) sensors are positioned in different places of the older adult’s home and are used to collect data regarding the behavior, the occupancy, the lifestyle, the environmental safety and the personal safety of a person. The non-wearable sensors can be used to detect the well-being/health status of a person as well as to monitor the sleep quality.
Ambient sensors can be used for monitoring the daily living activities performed by older adults as well as their daily routine and to detect behavior changes caused by the deterioration of the health state. In [14
] the data that are monitored and collected in order to evaluate the wellness of the older adult in a controlled environment (i.e., his/her house) are: behavior, occupancy, lifestyle, environmental safety and personal safety. The monitored infrastructure integrates the following types of non-wearable heterogeneous sensing units installed in the older adult’s houses: (i) movement sensors, (ii) electronics and electrical devices monitoring sensors placed on the electrical and electronic appliances, (iii) toilet sensors, (iv) contact sensors to monitor the opening/ closing of a door, office desk or and self-grooming table and (v) force sensors placed on sofa /bed/ armchair to monitor the sleeping or sitting activities. In [38
] the solution proposed for detecting the deviation from the daily activity routine of older adults is based on a non-intrusive monitoring infrastructure consisting of low-cost sensors such as passive infrared (PIR) motion detectors and magnetic door contacts sensors located in each rooms of the house. The monitored data contains information about the older adults’ lifestyle and is used to detect unusual activities performed by an older adult in a day or deviations from his previous routine.
Smart beds as well as cameras are an alternative to the wearable commercial devices for monitoring the sleep quality, beside the polysomnography and the videosomnography that can be used in clinical evaluations. Authors of [39
] propose a non-intrusive monitoring infrastructure consisting of pressure sensors embedded in bed which is used for detecting sleep quality. The sleep quality is assessed based on the sleep position and the sleep stages. In [40
] a microbend fiber optic mat embedded in the mattress of the bed is used to monitor the sleep quality. The sleep quality is evaluated based on the following types of data acquired from sensor: sleep duration, movements during the sleep, heart rate, awake stage duration, time spend in bed as well as the respiration rate. Similarly, in [41
] a sensor integrating an Emfit foil electrode embedded in the mattress of the bed is used to measure the sleep quality. The sleep quality is evaluated based on the movement activity and the heart-rate fluctuations. Piezoelectric film sensors can be embedded in the bed mattress to monitor the sleep quality, based on the variation of the heart and respiration rate and the binary actigram [42
]. Microsoft Kinect camera can be employed to detect the human shape and the body movements and a sensor tag that provides information about the sleep environment such as the temperature and humidity [43
], while Near-Infrared camera can be used to analyze the sleep behavior, based on the collected videos/images [44
Non-wearable sensors for fall detection approaches that are based on ambient sensors use cameras, infrared sensors, acoustic sensors or force sensors installed in the home environment to detect the fall. In [45
], an approach for detecting the fall that is based on low-cost fall detector that integrates Raspberry Pi 2 and an universal serial bus (USB) camera and several algorithms (e.g., background subtraction, Kalman filtering optical flow) is proposed. The detector is developed as a portable device that can be moved from a room to another of the older adult’s house and can detect the following positions: walking, standing, sitting and falling. Similarly, in [46
] is presented a method of detecting the fall in the case of older adults. The method is based on a Microsoft Kinect sensor for collecting the images combined with an algorithm for background subtraction and an ensemble of decision trees. In [47
] is proposed an approach for detecting the fall in the case of older adults that is based on a ground sensors network consisting of accelerometers and force sensors installed on the tiles from each room of the older adults’ apartment while [48
] proposes a classification method which is able to detect between human fall from non-fall, based on the acoustic waves transmitted on the ground collected with a floor acoustic sensor. A method to detect the fall in the case of older adults which combines acoustic sensors that collect the sound signals of the footsteps performed by older adults during the daily activities and a support vector machine algorithm is used to make distinction between fall and non-fall sounds is proposed in [49
]. Authors of [50
] develop a fall detection system that uses a circular array of 8—microphones to detect the older adult’s fall in real time and sends alerts to the caregivers.
Wearable sensors can be combined with ambient sensors in complex monitoring infrastructures that can be used to obtain more accurate results in monitoring the heath/well-being status or sleep quality. In [51
] the data that are monitored and collected in order to evaluate the wellness of the older adult in terms of functional and cognitive capacity can be classified in: vital sign, social interaction and physical activity. The monitoring infrastructure consists of a wireless sensor network integrating a collection of wearable and non-wearable heterogeneous sensors. eWall [52
] is a holistic monitoring platform for home, which collects the following type of data in the case of older adults suffering of chronic obstructive Pulmonary disease, mild cognitive impairments or others age-related impairments: (i) health data (e.g., SpO2, pulse, heart rate), (ii) well-being state data (physical activity, sleep, mood), (iii) environmental data (e.g., temperature, luminosity). In [53
] the sleep quality is monitored with a wearable three-axis accelerometer and a pressure sensor installed in bed. The accelerometer is used to determine the sleep pose and the sleep state (e.g., REM sleep and non-REM sleep cycles), while the pressure sensors are used to detect heart/respiration rate. Based on the collected data a novel algorithm is develop that is able to detect the sleep quality, based on the number of apneic episodes, the duration of sleep and the depth of sleep.
Virtual sensors are a source of data coming from software applications or services. For example, it is possible to determine an older adult’s location not only by using tracking systems (physical sensors) but also by using a by browsing an electronic calendar, a travel-booking system, emails etc., for location information. Other attributes that can be sensed by virtual sensors include, e.g., the user’s activity by checking for mouse-movement and keyboard input. Logical sensors are an extension of the virtual ones and make use of a couple of information sources and combine physical and virtual sensors with additional information from databases or various other sources in order to solve higher tasks. Pain Care is a healthcare app developed for iOS and Android device that allows older adults to keep a medical journal in which stores information about the medication that s/he take, specific symptoms/pains or side effects of medicine that could occur during the treatment [54
]. Based on this medical journal, the causes of pain and treatment efficacy is estimated. All these data are transmitted to medical specialist that can adjusted the medical treatment according to the patient personal profile and his medical journal. Researchers have investigated the benefits of using computer or mobile applications (e.g., memory training application or brain training applications) in maintaining/ improving the cognitive functions of older adults [55
]. In [55
] specific smartphone applications for improving the cognitive function of older adults are comparatively analyzed. The attention and the working memory are improved by engaging the older adult in training tasks with different difficulty levels for concentration, speed, memory, visual and logic. HealtheBrain, is another smartphone application in which physical activities are combined with memorizing activities in order to improve the cognitive functions of older adults with and without mild cognitive impairment [57
]. Similarly, in [58
] the results of using the cognitive training game in improving the cognitive functions of older adults is analyzed. The game includes a set of tasks to be performed by the older adults which aim to train various cognitive functions of the older adult, such as attention, memory, visuospatial or language function. There are also studies that demonstrate the efficiency of using computer/smartphone applications in dementia treatment. For example, in [59
] three smartphone applications are analyzed: EVO which is a cognitive training application that reduces the symptoms of depression, iPST which acts as a psychotherapy for depression and health tips which assists in the treatment control. In addition, in [60
] the effect of playing action video games in increasing the cognitive ability in the case of the people suffering of with dementia is analyzed. Reference [61
] presents SONOPA, a framework which combines ambient sensors with social networks to create social connection between older adults, based on their hobbies, localization or activity levels. The activity levels and the number of persons in a house (i.e., the level of occupancy) are determined based on the data collected with PIR and visual sensors, while the socialization interaction level of an older adult in a day is computed based on the level of the occupancy of the house and the information collected from the social network (the message communication, the visualized photograph, etc.). In addition, a matching algorithm is proposed that can identify new social connections for the older adults with low social level based on the data collected from sensors and from the social network (i.e., person profile, socialization level, etc.).
Considering the above, we have mapped the nowadays available sensing devices onto different type of assessments they enact aiming to determine their potential usage as reliable source of data for deciding on robot-based interventions (see Table 2
2.2. ML for Behavior Assessment
Machine learning (ML) techniques are an important component for building smart environments and associated technologies. Using the gathered monitored data, these techniques can be used for identifying and assessing certain situations in the older adults’ behavior, situations that usually require personalized intervention from the caregivers or health professionals [62
]. At the same time, the automatic recognition and classification of various daily life activities has the potential of reducing the costs associated with the healthcare of the older adults significantly, especially in the case of those cognitively challenged. By using sensors and advanced analytics over collected data, it is possible to extract information in real time about the monitored person and thus it is possible to detect anomalies and patterns that may indicate wellbeing, social and cognitive decline as well as healthcare problems that require intervention.
In this section we analyzed a selection of representative research articles which consider the application of ML techniques for various cases of behavior assessment for older adults. For each representative research article, the following characteristics are considered: the machine learning techniques used, their advantages and disadvantages and type of behavior assessment targeted. We have identified and classified the most important ML techniques that are used for building such smart AAL technologies:
Classification techniques—The state-of-the-art literature features several methods based on different types of classifiers for monitored data streams out of which ensemble learning methods are considered the best techniques for the classification of the data streams. There are still a lot of challenges posed by the data streams in the case of the ensemble learning algorithms such as the temporal dependencies [63
], the concept drifts [64
] and the feature drifts [65
] and those challenges may appear especially in the monitoring of the daily living activities that are situation-aware where similar monitored data can correspond to related activities such as ascending stairs or descending stairs;
Regression techniques—The application of regression techniques for daily living activities recognition in context-aware AAL systems [66
] is challenging because the identification of the activities should be performed after the beginning of the activities as soon as possible. A part of the limitations of the current approaches are the recognition of the activities after they are completed and the training of the models using offline historical data, a machine learning phase that leads to models which cannot predict the ongoing activities in a timely manner;
Clustering techniques—The clustering of the data streams should be adaptable due to the fact that the underlying data streams may change and evolve significantly in time, like in the case of data that results from the monitoring of the older adults while they perform different types of daily living activities. In [67
] are addressed in more details challenges regarding the clustering, the labeling and the interpretation of the IoT data streams dynamically, challenges that exist especially in those AAL systems that monitor the daily behavior of the older adults;
Other ML techniques—This category considers techniques such as discovery of association rules, patterns detection, anomalies detection, etc. The abnormal human activities are very diverse [68
] in nature due to a variety of aspects such as the way in which the anomalies are defined, the feature representations of the anomalies and the characteristics of the daily living activities data. The detection of the anomalies using various ML algorithms was approached in the research literature in a few studies such as the one presented in [69
] where the analysis of the anomalies is not considered as the main subject of the study, but in relation with the recognition of the daily living activities, the discovery of the behavioral patterns and the decision support.
There are three types of older adult behaviors considered that is usually addressed in the research literature through ML techniques, namely daily activities behavior, agitated and aggressive behavior and medication adherence behavior.
Assessing the daily activity behavior of the older adults is relevant for the detection of the abnormal situations [70
]. In [71
] the authors propose a generic architecture for the monitoring of the activities in smart homes and approach a large variety of ML techniques such as the classification of the activities, the prediction and the reminding of various activities and the detection of the anomalous patterns. However, that review is focused on the basic daily living activities behavior of the older adults and it considers only partially or not at all other types of behavior such as the medication adherence behavior or dementia specific behavior. ML challenges related to the classification of frequent daily living activities (i.e., eating and drinking) are described in [72
]. The authors consider data from three datasets collected using various types of monitoring sensors such as power meter sensors and motion, contact and audio sensors. The tested classifiers were support vector machines (SVM), random forest (RF), Fisher kernel learning (FKL) and hidden Markov model (HMM). A different classification approach in [73
] considers classification techniques for the remote acoustic monitoring of the older adults in AAL scenarios of residential scale. The data were collected from wireless acoustic sensors and the decision to trigger or not an alarm is taken after the running of a two-stage ANN-based classification process for audio events. One ML usage in AAL systems is the prediction of the daily living activities of the older adults [74
] ranging from regression techniques related to the prediction of the CO2
consumed inside a room using an Artificial Neural Network (ANN) Levenberg–Marquardt (LMA) prediction model to other ML techniques related to the detection of the daily living activities from data collected by humidity, temperature or microphones [75
]. The authors of [14
] consider a smart aging system for distinguishing the variations from the baseline. The proposed ML method consists of preprocessing, segmentation, feature extraction, classification, pattern recognition and anomaly forecasting. Moreover, this approach also considers the medication activity with respect to the food intake activity. However, the performance results are affected by noise that can be generated by faulty sensors or by the presence of a visitor. Another approach, [76
], proposes a method based on a deep convolutional neural network (DCNN) in order to classify ten types of activities. Even though that approach is based on one of the most preferred solutions for daily living activities monitoring, namely the unobtrusive activity recognition and returns a very high F1 score for eight out of ten activities that are monitored it may be more expensive in terms of computational resources than other approaches. The recognition of the activities was approached in [77
] using data collected from public available datasets which are characteristic to smart home scenarios. In addition to classical daily living activities such as bathing, sleeping and eating, that approach also considers the medication intake activity. The proposed classifier is a long short-term memory (LSTM) model and compared to other existing machine learning models it returns better results while the performance of the approach is affected by various dimensions such as the number of residents, the number and the types of activities and the test days duration. The authors of [78
] propose a fog-based deep learning fall detection system using data collected by a tri-axial accelerometer. The proposed architecture is based on three layers, namely a medical devices layer, a fog layer and a cloud layer. The results presented in that article show that the deep learning methods such as LSTM and gate recurrent unit (GRU) are better than the classical machine learning methods such as SVM and k-nearest neighbors (K-NN) for falls detection.
Agitated and aggressive behavior is one of the most challenging symptoms of dementia [79
] and its automated detection using sensors is useful for the caregivers that can act quickly in these kinds of situations. In [80
] the authors consider challenges related to the agitated behavior in the case of the people with dementia. The automatic detection of the agitation is approached considering data from various sensors placed around the bodies of the monitored subjects that collect information about the skin temperature, the skin response and the heart rate and that data are further analyzed using a SVM classifier. One drawback is that the approach requires the sensors to be placed on the bodies of the monitored older adults. The authors of [81
] consider the application of novel technologies for the early Alzheimer disease (AD) detection. Some associated symptoms are the aggression, the anxiety, the aberrant motor behavior and the irritability. The data from the subjects was analyzed using an ANN with two classes (control subjects and AD), the approach is low cost and it does not have any side effects. Related to prediction of the agitative behavior for the patients with dementia an important challenge is overcoming weakly labeled and sparse data [82
]. The goal is to infer proficiently the agitation episodes from data collected by wearable sensors using multiple-instance learning (MIL) models. However, that data comes only from 10 residential deployments, each with a duration of 30 days. In [83
] the data collected from various sensors is used in order to detect the challenging behaviors in the advanced stages of dementia. K-means clustering was applied in order to cluster the residents in two groups, one group characterized by challenging behavior with more passive features, and the other group characterized by challenging behavior with more active features. However, the possibility to detect the disorientation using accelerometer data in different solutions based on assistive technologies is considered as a future research direction. The detection of the agitation behavior of the people with dementia using wearable devices is also approached in [84
]. The combination of data from multiple sensors leads to better results than in the case when data from a single sensor is used. The applied classifiers are RF and SVM and the results show that the multi-model sensors are feasible for detecting agitation in the case of the people with dementia.
Medication adherence behavior of the older adults to specific medication plans is critical, especially when the older adults have special conditions such as dementia. The application of pervasive technologies [85
] for the monitoring of the daily medication behavior may predict when the medication prompting is effective. The approach presented in [86
] considers clustering challenges related to the medication adherence. Since the medication adherence behavior is correlated with other daily living activities such as drinking and eating, that approach also considers other activities. The proposed clustering approach is a k-means that is improved using fuzzy set. The authors of [87
] consider the application of reminder-based interventions for the individuals with dementia as the memory limitations often lead to activities that are incomplete or not initiated. The article considers the application of the sensor technologies in combination with machine learning technologies in order to address those challenges. The classifiers considered in that approach are various such as decision trees (DT), K-NN, naïve Bayes (NB), SVM and logistic regression (LR). The approach presented in [88
] assesses the medication adherence of 38 older adults with a mean age of approximately 87 years. The two measures of adherence proposed in that article are the percent of the days when the medications were missed and the spread in time when the medications were taken. The data were analyzed using three linear regressions and the medication intake habit was monitored continuously using a MedTracker 7-day pillbox. The medication adherence is approached in [89
] in relation with other activities such as walking, drinking water, writing and texting. The applied classification algorithm is RF from Apache Spark. Even though the data were misclassified sometimes using that approach, the application of the near-field communication (NFC) sensors may improve the results. Medication intake behavior in relation with the eating and the drinking behavior is researched in [90
]. The experiments were conducted using data collected from gyroscope and accelerometer sensors and the performance of various classifiers such as K-NN, NB, DT, multilayer perceptron (MPC), RF and HMM was evaluated. Moreover, two out of the five analyzed different users were seniors. The results presented in that article are very promising, and the system should provide an adequate basis for smart reminder triggering in the case of the autonomously living seniors. However, that approach was tested only on five seniors the method does not consider special characteristics of the older adults that may have dementia such as memory related problems or unpredictable behavior. The medication intake activities were considered in [91
] using a solution based on data collected from a smartwatch in relation with other activities such as texting, writing, walking and bottled water intake. In terms of F1Score, the Gradient Boosted Trees (GBT)-based approach returned better results than the approaches based on LR, SVM and RF. One advantage compared to other approaches from literature that consider the medication intake activity is the fact that it considers at least two types of medication intake. In [92
] the medication adherence is considered using a ML approach that is based on body worn sensors. The medication intake activity is analyzed considering correlated activities such as drinking, taking chocolate and eating and the applied machine learning classifier is the DT classifier.
2.3. Social Robots Driven Intervention
Social robots are nowadays seen as key technology for supporting older adults care at home or in care institutions. Even from the breakthrough of the assistive robot’s technology two main advantages were identified: the functional capabilities and the affective aspects [93
]. They can offer different care functions (physical activity, affective therapy, cognitive training, physiological therapy, etc.) while they can help to increase the quality of life of older adults through companionship and social interaction. There are basically two classes of social robots: physically assistive robots that are focused to perform physical tasks and socially assistive robots that can be used for the social and psychological needs of older adults [94
Few literature studies offer a clear view over the current status and trends for social robots being focused most on old robot models and their potential application in the older adult care domain [93
]. These studies highlight PARO as one of the first social robots with immediate applicability for older adult care. PARO is an interactive robot developed by Japanese AIST research institute that uses animal therapy as a care method. It has reached its 8th generation with a price around of €5000 and is successfully used in hospitals and care homes across the world, but it can be easily used in the older adult’s home. PARO uses tactile, light, audition, temperature and posture sensors to learn the surrounding and interact with older adults and can dynamically adapt to the user’s actions and preferences. As benefits, PARO can reduce patient stress and indirectly of their caregivers, stimulate patients and contribute to the socialization part of their life. A recent study has shown that PARO can be very useful for improving quality of life of older adults with dementia and Alzheimer, affection and social interaction, reducing depression and anxiety—even for reducing pain medication usage [96
]. Another comprehensive evaluation study highlights the potential usage of PARO as a pet therapy for older people with dementia while identifying barriers for its wide usage such as the fact that users’ needs and experiences are not properly taken into consideration and that it fails to meet the actual clinical needs [94
]. Besides these, compared with nowadays social robots, PARO has no functionalities for requirements such as stimulating older adults’ memory or cognitive functions, increasing physical activities or aiding the caregiver with care information.
The Pepper robot developed by SoftBank is a humanoid robot featuring multi-modal communication able to recognize faces and basic human emotions, is capable of human interaction directly through conversation [97
]. It can also exhibit body language, perceiving and moving around and is currently used in schools for child education or different businesses mainly as information point for users. Pepper is priced around €15,000 (acquisition price and subscription fee for three years) and has spawned multiple projects for developing new instances of its basic features such as ASIMO, COMAN or Enon and EU research projects such as CARESSES, CROWD-BOT or ANIMATAS. As for its underlying technology, Pepper uses a six-axis inertial measurement unit (IMU) sensor, microphones, cameras and 3-D sensor, laser sensing modules, loudspeakers, sonar sensors, infrared sensors, tactile sensors, bumper sensors and features an attached tablet for direct interaction and configuration. For the specific use–case of older adult care, Pepper was successfully used in healthcare and older adult-care facilities mainly as narrative-memory-based human–robot companionship [98
] and medicine taking reminding, encouraging older adults to keep active and helping them keep in touch with family and friends [99
]. Recent research approaches have used Pepper as older adults’ companion for suggesting personalized physical activities in the context of active aging [100
]. The proposed solution uses deep learning methods on Pepper recorded information to classify the exercises and to schedule personalized physical activities. The robot is integrated in the context of a robot system named PHAROS which in addition to the robot component it contains a component for human exercises recognition which applies deep learning models on the data recorded by Pepper and a component which recommends physical exercises periodically considering the data from the agenda of the users. In [101
] Pepper robot is used in the context of a system called Crowd of Oz (CoZ), an open-source system that allows conversational tasks. The objectives of CoZ are to enhance both the contextual and the social awareness of the workers, to manage the asynchronous nature of the workers during the conversational task and to support the task performance of the workers. The Pepper robot has also been used for the case of people that have special conditions such as schizophrenia or dementia during recreational or rehabilitation sessions [102
]. Even though the humanoid robot can elicit simple instructions for simple activities such as physical exercises, numerous improvements are required in order to deploy the humanoid robots in the long-term care. Moreover, other aspects should be considered such as the motivation of the older adults or a sense of calmness from the clients’ side. Some benefits of the application of Pepper in hospitals [103
] are the lowering of the stress levels and the contacting of the family members when the older adults are unable to contact anyone especially in critical situations when something happens to them. However, there are still further research directions that should be considered such as the recognition of a wider range of emotions.
The Nao robot is reported in several approaches as useful coaching assistant [104
]. In [104
] a particular use case for Nao is presented as an illustrative scenario for social robot driven intervention, an autonomous exercise tutor. The robot is capable to learn from a human, to generate feedback situations such as speed and amplitude adjustment, mirroring detection and no motion. However, the older adult participants did not prefer the intervention of the robots due to the poor social skills and because they perceived the sessions with the robots as one-to-one instead as a social event. In [105
] a Nao socially-assistive humanoid robot is tested in the context of a smart home environment. Even though the results presented there suggest that the interaction between the robot and the older adults is not characterized by anxiety, the maintenance of high levels of enjoyment for the older adults for prolonged time periods is still a challenge. Nao was also used to detect the behavioral disturbances of the people with dementia [106
]. The authors aimed to evolve the role of the humanoid social robot to a technological support tool that functions autonomously, to drive to the resident and alert the staff or distract the resident temporarily when a behavioral disturbance is detected.
There are several other social robots that are reported in the literature as potential solution for older adult care intervention. Zora is one of these and can be used in interventions for the care personnel and for the older adult-care institutions [107
]. The results show various types of impacts, ranging from negative to positive. A part of the participants suffered from memory disorders and others required round-the-clock care. The impact on the care personnel differs very much from the impact on the older adults. In the case of the care personnel the impact was on dimensions such as the working atmosphere, the professional development and the competences, while in the case of the older adults the impact was on the physical activities, the older adults’ interaction and the sensory experiences. Authors of [108
] consider a mini robot as support for a motivational decision-making system (DMS). The stimulating and the improvement of the interaction of the robot with the users is considered from various perspectives such as the performance of cognitive exercises and the performance of educational games. In the scenario described in that article the mini robot has two motivations, namely a social motivation and a relaxing motivation, and those two motivations must be in equilibrium while the robot interacts with a user. However, one drawback of that approach is that the robot is modeled using only two motivations and many other complex aspects should be considered when the users have specific health conditions. Another approach [109
] proposes the learning of the social gestures through imitation through a humanoid robot called Tangy that is programmed to avoid the self-collisions and to generate arm trajectories that are collision-free. Even though the human demonstrators considered in that article do not have health conditions, the research considered in that article can be further adapted to stimulate the interactions of the older adults with the social robots. Another study that considers the application of social robots for older adults’ intervention is presented in [110
] where the authors propose Stevie robot that is tested and validated on different categories of users such as the residents of a retirement community and the healthcare personal of that community. However, the capabilities of the robot should be extended in order to apply that robot in the case of the people with dementia or with different types of physical disabilities. A combination of more than one robot for developing behavioral intervention systems but focused on the children with autism spectrum disorders is proposed in [111
]. The approach uses iRobiQ and CARO robots for providing training and support.
summarizes the main approaches analyzed in the context of older adults’ robot driven coaching and intervention organized by the type of social robot intervention used.
2.4. Technology Limitation and User Acceptance
Older adults are keen to continue living in their own homes rather than move into residential institutions and the assistive technologies reviewed in the above sections have the potential of providing the need support in managing various problems of their daily life. However, there are specific challenges that need to be addressed for both technological development perspective and integration with the care models as well as regarding its acceptance by the older adults.
From the technological perspective, the nowadays older adults assistive services are shifting towards the use of objective monitoring using IoT sensors and sensors networks, but in general, they lack in personalization when it comes to addressing the older adults wishes and needs and are relying on dedicated healthcare resources for assuring the intervention and continuity of care. The assistive service models in the area of managing and carrying of older adults are focused on providing general information and awareness of the specific disease. Being overwhelmed due to daily life activities and carrying duties the end-users are no longer interested in general-purpose information, but they want personalized target support. In addition, most of today’s care process assessment is still relying on self-reporting of “perceived behaviors”, but this kind of models are proven not to be viable in the case of elders who mostly deals with problems such as forgetfulness and confusion and chaining mood and behavior. Recent IoT advancements and the development of miniaturized sensors have the potential of changing this situation, by enabling the remote and daily monitoring of important care aspects such as adherence to recommended therapy and lifestyle changes delivering more coordinated innervation through the means of social robots. In addition, this has the potential of improving the older adults’ engagement and adherence and timely communication among all parties involved in the care process. Moreover, there is a need of integrating within the nowadays care models novel technology solutions supporting the non-face-to-face interaction and follow-up of older adults, such as advanced robot-supported verbal communication tools with caregivers and patient and/or family/caretaker support for self-management, independent living and activities of daily living. All these limitations need to be systematically addressed to support the perspective shift for next-generation of coordinated care service models for older adults which are relying on ICT-based pervasive and objective monitoring of daily live functioning using advanced IoT sensors, quantifiable metrics of assessing the elders’ deviations signaling conditions decline and timely interventions supported by robots/tablets/avatars.
Acceptance of these novel technologies well as their daily usage may be challenging for older adults. As shown in the literature, several technology acceptances factors can be correlated with aging ranging from technology costs, privacy implications and usability up to social aspects such as the implications for family or friends [112
]. Several acceptance models can be used to support the process as technology acceptance model (TAM), unified theory of acceptance and use of technology (UTAUT) or senior technology acceptance model (STAM) [113
] by assessing indicators such as technology perceived usefulness, ease of use, experience or social implications. However, even these models cannot capture aspects such as user acceptance in time and miss key aspects in the older adult life such as cognitive decline and social isolation [114
Monitoring older adults’ activities is the most common identified barrier for older adults’ acceptance since it involves breaking their privacy by installing smart devices such as wearables or sensor networks in their homes and collecting sensitive data [116
]. Even that at beginning the older adults may not trust in these technologies, after using them they start to perceive them as important. A key role for accepting in home remote monitoring is attributed to family, friend and caregivers that can assist, train and support the older adult in embracing the change. Passive monitoring-based on IoT devices is the most suitable approach for smart home data collection while wearables technology seems to have a good acceptance ratio especially to well-educated older adults that appreciate the possibility of self-monitoring their health status [117
]. In contrast, active monitoring such as monitoring older adults with cameras in their environment is raising major privacy and ethical concerns and consequently it encounters their reluctance of using such systems even though they are useful for users that have health problems such as dementia [118
]. In domestic environments, sensor-based monitoring infrastructures usually require combining heterogenous sensors and devices for capturing information regarding the older adult behavior leading to another major issue for developing smart AAL systems, namely data heterogeneity. However, this issue was thoroughly researched leading to different solutions ranging from involving web semantics and sharing a common context model for building smart objects in heterogenous IoT networks [120
] to defining and using ML and big data technologies to process time-series-based heterogeneous and distributed streams of data in a unitary approach [62
The benefits of assistive technologies come from their core Artificial Intelligence (AI) components that use advanced ML techniques for taking decisions, alerting or providing support for doctors and caregivers. The user acceptance for these technologies relates to their understanding of the brought benefits and potential improvement for their quality of life. In general, older adults, doctors and caregivers agree and seem to be comfortable to us AI/ML technologies for assisting and improving the care process [123
]. They are usually adapted to particularities of older adult’s health status such as mild dementia [83
], medication and polypharmacy [88
], cardiovascular problems [15
] or physical activity [76
]. However, the acceptance of such technology is usually influenced by factors such as age and education [125
], while the adoption of intelligent technologies is corelated with the older adult loneliness and with the support provided by caregivers [114
The use of social robots for older people care at home or in initialized care is a new technological trend for supporting the intervention in assistive care models. Existing studies for the social robots’ acceptance are relatively new and conclude that further research is required for assessing acceptance rate, but in the same time clearly identify the care areas where the robots can be successfully used as physical assistance, safety/monitoring and social companionship for an older adult [124
]. Other surveys show that older adults have positive thoughts regarding the usefulness, utility, safety and trust of a social robot, while doctors and caregivers consider that the robot is a useful tool for rehabilitation [127
]. In general, more positive acceptance is found after the users interact with the robot after a period, while the acceptance level is directly influenced by the robot’s social capabilities [128
]. Other authors pinpoint that usefulness, adaptability, enjoyment, sociability, companionship and perceived behavioral control are important for the high acceptance rate of social robots [129
]. One specific factor for accepting the social robot as a care companion is loneliness, the social robots may offer support and companionship especially at home [130