Long-Term Impact of Intelligent Monitoring Technology on People with Cognitive Impairment: An Observational Study

Intelligent systems for people with cognitive impairment

Interactive smart home systems and intelligent interfaces have the potential to deliver AT at home which is particularly useful for elders, especially those with cognitive or physical impairments due to aging [20, 27]. In this context, assisted living systems combined with simple non-pharmacological interventions could help elders by promoting their QoL while decreasing cognitive-related problems. A recent systematic review has revealed six broad ICT disciplines [28]: 1) general ICT, 2) robotics, 3) telemedicine, 4) sensor technology (including “smart home” systems), 5) medication management, and 6) video games. In particular, smart home environments enable elders to live independently at home for a more extended period and reduce the caregivers’ burden by supporting their services [29, 30]. In our recent systematic review [31], we reached the conclusion that various actions have been undertaken so far in grounds of development, usability, cost-effectiveness, deployment, and ethics of assistive and health technologies across Europe. However, a fundamental problem hindering the adoption of ICT and smart home environments is that the elderly lack familiarization with the technological aspects of these solutions, thus reducing their usage to a minimum [32, 33].

Several studies have demonstrated the efficiency of assistive technologies to increase the independence, cognitive function, well-being, safety, and security of elders [17, 34–45]. For instance, in order to enhance the accuracy of medication administration and management, which is a common problem among elders, several applications have been designed, such as medication dispensers [21] and smartphone applications [46] to provide reminders and notifications. Many others have tried to address dementia-related problems, such as incontinence, by developing wetness-sensors [47, 48] and urine detection devices [48]. On the other hand, there are other systems, such as “ALARM-NET” and “CareWatch”, that recognize Activities of Daily Living (ADL) and improve dementia care provided by the caregivers [23, 50]. Other systems make use of detachable sensor arrays, such as wearable accelerometers [51, 52] and wireless heart rate monitors [53] or sensors for monitoring ADLs, such as hand washing [54]. Also, the “COACH” system [54] explored the efficacy of the proposed computerized device to assist people with dementia through ADL by using audio and/or audio-video prompts with six participants. Another study of the same size [55] examined the ability of a custom Body Sensor Network (BSN) to capture the presence of agitation against currently accepted subjective neuropsychological measures and to discriminate between agitation and cognitive decline. The study in [56] investigated the effectiveness of verbal instructions presented via technology in “helping persons with mild or moderate Alzheimer’s disease perform daily activities, such as morning bathroom routine and table setting, coffee preparation and dressing with four, two and one participant respectively.”

The aforementioned approaches are being used to capture daily vital signs of people, automatically detect ADLs which will enhance autonomy, recognize emergencies, and follow the disease progression of people with cognitive impairment [9, 57–59]. Other systems have been designed to offer support to elders with regards to time [60, 61] and space orientation problems [34, 62–65] or psychological support as a result of social isolation and lability [66, 67]. Moreover, there are several prompting devices, which assist people with memory disturbances on a daily basis [60, 68–70]. In general, various solutions have been developed to meet the elder’s needs, such as web-based information systems, video-calling, and electronic activity support systems [36, 72]. Furthermore, a more holistic approach, the so-called “smart homes”, has been proved to be useful for cognitive impairment related to AD [23, 27]. In detail, smart homes include several sensors (e.g., accelerometers, microphone arrays, pressure sensitive mats, gas sensors, etc.) and have been proposed as effective solutions to monitor users’ ability to cope with ADL [16, 72–75] or prevent significant incidents [61].

In a nutshell, several studies focus on the effectiveness of AT integrated into end-users’ domestic environments, to increase the quality of care, and reduce the burden of caregivers. However, state-of-the-art technology systems have some limitations. For instance, smart home platforms are often very comprehensive and require significant installation efforts, which make these systems difficult to be used and deployed in real settings [22, 77]. Also, most smart home systems use video-based technology or microphone recordings, which raise privacy issues (personal and data privacy) [78]. Other approaches require the user to wear “body-mounted” sensors or interact with complex systems for data acquisition [57, 79], an issue which requires user’s consent, which is not always feasible, especially for elders with cognitive impairment. Furthermore, intrusive systems are not well accepted both by people with dementia and related cognitive symptoms, as well as caregivers [80, 82]. Others have stressed that devices should be embedded in a person-centered model [68] in order to effectively apply adaptive interventions. However, it is a common opinion that “research into the use of ATs is widespread but in its infancy, consisting mainly of small-scale studies and few longitudinal studies” [78]. In a recent study [82], the authors have highlighted that it is of high importance to develop and integrate assistive technologies in a real-world setting, outside the laboratory, and perform longitudinal studies that will evaluate the interaction of the end users with technology [68]. Therefore, the main disadvantage of the current AT, is the lack of clinical evaluation trials, which will prove that such intelligent monitoring systems: 1) are useful and easy to be installed in participant’s home setting, 2) are sustainable, unobtrusive, and acceptable by the end-user, and 3) can provide clinicians with the analysis of multimodal, commentary information, intelligent coupling of clinical and user-related knowledge and high-level correlations in an adequate and objective way.

Aim of the present study

This paper presents new results on the usability and effectiveness of a previous sensor-based remote monitoring study for longer periods (up to a year). The design and goal of this longitudinal study in terms of data and not participants, aiming for data collection and acceptance from four months and up to a year has a two-fold aim: 1) to investigate whether the long-term use of sensor-based remote monitoring system at home can be accepted and sustained, and 2) to validate the beneficial impact of its long-term use, taking into account the tailored system-driven interventions, among different groups of people with mild cognitive impairment (MCI) and AD.

In previous studies, a multidisciplinary team with experience in ICT monitoring systems have designed, developed, tested, and evaluated the remote monitoring system for people with cognitive impairment [23, 83]. However, these studies have been limited to four users, for the duration of up to four months. The ultimate goal of our study is to pilot, maintain, and evaluate the long-term effects (up to a year) of a personalized sensor-based system to support non-pharmacological interventions for people with cognitive impairment, both in preclinical and more advanced stages. The sensor deployment enabled data collection from six users in their real homes, under real conditions, in the MCI and AD stages, from four months and up to a year, accumulating a substantial longitudinal dataset. It also facilitated ongoing co-design between system developers and participants and evaluation activities.

Before the installation of the system in the first participant’s home, there has been five-month extensive testing of the system in a laboratory home-simulated environment with over 90 participants [84, 85], where activity recognition (through a semantic fusion of sensor information) reached a recall and precision close to 82%. Based on the data analysis, the clinicians were able to introduce specific and personalized interventions to each of the six participants. Moreover, these interventions were adapted based on the information that the system provided to the clinician during the period of each home protocol. The present study aims to extend and further explore the longevity and acceptance of the technology for the long-term (up to a year) while maintaining acceptance, usability, and health benefits for the extended group of subjects. As an ultimate goal, it aims to leverage a combination of our proprietary technology and cognitive impairment insights with elegant design to empower older adults to intuitively interact with the technology and follow personalized care through their participation in non-pharmacological tailored interventions.

The proposed solution

For the system to be usable and acceptable, it integrates a variety of off-the-shelf Internet-of-Things (IoT) devices, interconnected and continuously feeding their measurements to the cloud. The market is now advancing and growing a lot with regards to wearable and ambient smart devices for smart home monitoring. Therefore, finding the optimal IoT device for most everyday use, at a given time, is a moving target. For this reason and to future-proof our solution, it integrates many popular devices in the market, with the additional benefit of choosing different devices for different users for further personalization. A middleware layer integrates not only multiple device application programming interfaces (APIs), but also a knowledge representation layer integrates data from different tools classifying the type of information it receives. Furthermore, semantic interpretation and fusion are enabled, to transform and combine various inputs into more complex behaviors and symptoms in daily life, presented the clinicians with an enriched application to monitor these important signals. The patients are also supported by being shown positive, motivational and inclusive indications from the monitoring system, effectively closing the loop between them and the clinicians.

More specifically, key features offered by the system are: 1) the clinician interface, which provides the clinicians objective and realistic information about the patient’s daily activity and health status (e.g., sleep patterns, physical activity, etc.); 2) focus on a series of domains of the target users’ life which might be supported by discreet and transparent assistance of a domestic assisted living system; 3) design and deployment of an intuitive user interface that will encourage users to monitor their daily functionality as measured by ADLs and interact with the system smoothly and efficiently, 4) use of context-aware technologies at home combined with tailored adaptive interventions guided by clinician, that will stimulate elderly people with MCI and AD diagnosis to increase their levels of physical activity; to enhance their cognitive state; to promote QoL and independence and to prevent future cognitive decline.

Compared to existing works [37, 86–88], the proposed system is deployed in real-home settings of people with cognitive impairment for hard-to-reach long-term periods of up to a whole year, monitoring multiple aspects of daily life, behaviors, and symptoms. Additionally, previous studies focused solely on capturing particular aspects and ADLs (e.g., agitation, sleep, handwashing, make phone calls) with a single type of sensor to explore or ameliorate disturbances common in the dementia spectrum (i.e., navigation) or used technology as means of communication of patients and relatives. Meanwhile, the proposed system uses multi-modal holistic monitoring to capture multiple aspects of life to aid in personalized interventions and monitor progress: sleep, movement, doing chores, watching TV, etc. Furthermore, the majority of assisted living systems are not exclusively designed for people with cognitive impairment related to AD. Instead, they target the general elderly population with a neurocognitive disability, while only a few intelligent systems have been designed exclusively for people with dementia and AD [89]. In light of those facts, we have developed, piloted, and evaluated a sensor-based system to assist clinicians in the care of people with cognitive impairment. The evaluation has proven that our solution is useful for clinicians, since it pinpoints specific areas of interest, and it can drive the way for the deployment of adaptive and personalized interventions.

From a technological perspective, the system integrates a universal middleware layer that connects to a multitude of off-the-shelf IoT devices in the market. AT algorithms perform sensor data analysis, interpret and aggregate raw numeric measurements according to users’ needs. The processed information is semantically stored and linked using advanced Semantic Web technologies. Intelligent reasoning and decision making further combines information and translates them into behaviors and symptoms and highlights problems that are meaningful to clinicians and patients.

The proposed solution builds upon and extends previous works, which use sensor-based systems for assessment trials and short-term home monitoring combined with clinical interventions. In detail, the underlying platform integrates various sensors from different manufacturers, various sensor modalities and processing algorithms for image, audio, and lifestyle information [90–92]. Small clinical laboratory trials have successfully employed this platform, subjectively measuring performance in functional tasks in a confined, closed world environment to assess the cognitive state of people with MCI and AD [83]. More than 200 individuals participated in the laboratory trials, yielding an 84% accuracy to cognitive state detection and participants’ classification [83, 85]. Meanwhile, various non-pharmacological intervention programs tackle cognitive, behavioral, and functional problems that commonly occur within the dementia spectrum [5, 93–96]. Combining these interventions with the short-term application of the system (up to four months) at the actual homes of individuals living alone, the remote monitoring has proved to enable decision making and, in turn, significantly improve and prevent cognitive decline [23]. To the best of our knowledge, this work presents a pioneering study of an intelligent sensor-based system that evaluates the positive impact of its long-term usage in several domains of cognitive and functional status, while, at the same time, proves its sustainability and acceptability based on actual user feedback.

System design and implementation

The proposed system follows a multidisciplinary approach that brings into effect the synergy of the latest advances in sensor technologies of multiple, complementary modalities, large-scale information fusion and mining, knowledge representation, and intelligent decision-making support. More specifically, the system integrates several logical layers, as depicted in Fig. 1, to serially process the information from raw, sensor data to higher-level behaviors and symptoms useful to clinicians. The end-goal is to integrate different sensing modalities, such as physical activity and sleep sensor measurements, combined with input from lifestyle sensors and higher-level image analytics with common semantic representation and interpretation. For more in-depth information on the technological aspects of the system, such as processing algorithms and knowledge models, the reader is referred to [97].

OPEN IN VIEWER

Fig.1

System architecture and logical layers: ambient and wearable sensors, sensor and image analytics, knowledge representation, semantic interpretation, and user interfaces.

The Ambient and Wearable Sensors layer is simply comprised of the hardware, or the devices in the system. The current selection of sensors consists of proprietary, low-cost devices, initially intended for lifestyle monitoring, repurposed to a medical context. Ambient sensors are installed at the end-users’ homes, embedded into furniture and daily-life objects, remaining unobtrusive, out of their sight and mind. The wearable sensors are comfortable, worn devices from bracelets to rubber-band watches, that the user can comfortably wear 24/7. The rise of wearables in the market allowed the selection of the least obtrusive ones, as studied in [51]. Regardless of manufacturer, the current ambient sensors are the Ambient Cameras, the Plug, Tag, Presence and Sleep sensors, complemented by a single wearable Wristwatch per end-user.

Ambient depth cameras (Xtion Pro, http://www.asus.com/Multimedia/Xtion_PRO/) are collecting both image and depth data to be later processed by suitable algorithms that infer gestures, activities, and locations in the area.

In the Sensor & Image Processing Layer, each device uses dedicated modules that wrap their respective API, retrieve data, and process them accordingly. The architecture is future-proof, as the development of new plug-ins can add sensors. Processing is performed to generate atomic events from sensor observations, e.g., through aggregation. For example, streams of power consumption from the Plugs are transformed into personalized events of using the appliance. The lack of such events means the device was not used at the time. Similarly, presence is recorded with IR events from Presence sensor and object usage from Tags. In the case of image data, computer vision techniques are employed to extract information about humans performing activities, such as opening the fridge, holding a cup, or drinking [98].

In the Knowledge Base layer, all atomic events and observations are mapped to a uniform semantic representation for interoperability and stored at the system’s Knowledge Base. This mapping is done according to ontologies, i.e., taxonomies of knowledge and their relationships. Both established and new ontologies are used to capture events, activities, behaviors and clinical notions.

The Semantic Interpretation & Problem Detection layer applies further semantic analysis, activity recognition and detection of problems, i.e., anomalies. This variety of derived information can later be used by domain-specific applications offering a tailored view to different types of users. The process of deriving this higher-level information involves aggregating the sensor and image data and applying logic-based algorithms on the information gathered so far in the knowledge base. The aim of the logic-based algorithms is to effectively combine atomic events to more meaningful, higher level notions. For example, events of presence in the kitchen, combined with cooking gestures, the use of a microwave oven and boiler deduce a “cooking” activity. A challenge here is to detect the exact activity duration between actually interleaving it for something else. Therefore, well-defined activities such as the development of new plug-ins can add sensors, or going to the bathroom are more accurately detected in time, as opposed to watching TV which is easily interleaved with chores or even sleep. The models for the activities are dictated by the clinicians after an interview of the end-user to pinpoint the areas, activities, and objects to focus.

The Clinician, End-user & Caregiver Interface layer contains the higher-level web applications which present that information visually to the users. At the core of this work lies the Clinician Interface (Fig. 2), intended to show the highest detail of information to the doctor or psychologist to support them to drive the intervention further. It pursues a holistic approach of integrating data capturing physiological, functional, and cognitive aspects of elders’ daily activity. It also focuses on the synthesis of multiple data sources to provide a holistic assessment to enhance clinical decision making and support a personal health record system. It utilizes existing hardware systems and focuses on data integration, processing, and visualization contributing to a better understanding of the interoperability of monitoring systems and maximizing knowledge generation using multiple data sources.

OPEN IN VIEWER

Fig.2

The Clinician interface. In the upper left side, a chart illustrates the comparison of two variables (physical activity-blue line and total time asleep-black line) as measured by two different sensors (wearable sensor and sleep sensor) for three different days. In the upper right side, a detail description of each modality as measured by several sensors indicating a specific activity and its duration is being presented. On the bottom side specific problems are being presented regarding sleep for each day after the setting of specific thresholds by the clinician.

Participants and settings

All participants were recruited from March 2015 to April 2017, via a memory and dementia outpatient clinic of the 3rd Department of Neurology of the Aristotle University of Thessaloniki, Greece (http://www.med.auth.gr/) and from the Day Centers of Greek Association of Alzheimer’s Disease and Related Disorders (GAADRD, http://www.alzheimer-hellas.gr/index.php/el/). We focused on stages 3-5 of the disease according to the Global Deterioration Scale (GDS) [99]. In detail, MCI participants were selected based on Petersen criteria [100] and deemed to be fit to take part in the study by a neurologist specialized in the field of dementia and a neuropsychologist. All AD participants fulfilled the Diagnostic and Statistical Manual of Mental Disorders (DSM-V) criteria for dementia of Alzheimer type [101] (APA, 1994) and the National Institute of Neurological and Communication Disorders and Stroke/Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria for probable AD [102]. Participants were diagnosed with MCI or AD according to their history, neuropsychological tests, structural magnetic resonance imaging, and other necessary laboratory examinations. The study was carried out following the Declaration of Helsinki and was approved by the GAADRD Scientific & Ethics Committee (17/26-02-2015).

Inclusion criteria consisted of participants who were aged 60 years or older (Table 1 presents the mean values of the age with the standard deviation for each group).

Exclusion criteria included: 1) any severe physical illness, 2) current psychiatric or neurological disorder, 3) history of drug or alcohol abuse and use of neuro-modifying drugs other than cholinesterase inhibitors or memantine in AD group, 4) having any somatic disorder that may have caused subjective or objective cognitive impairment such as stroke, other neurodegenerative diseases, traumatic brain injury, brain tumor, and alcohol abuse, 5) being under treatment less than 90 days before their participation in the study. During the neuropsychological assessment and clinical history interview, particular attention was devoted to rule out subjects referring to any symptom which may conflict with study procedures.

Twenty-one (n = 21) participants were initially selected to enroll in the study, with 18 being randomized to one of the three groups, as shown in Fig. 3. As a result, a total of 18 participants, six in each of the three groups took part in the study. In each group, there were four MCI participants and two with AD. The EG used the system for at least four months and received adapted interventions according to the system insights (mean±SD: age = 72.8±6.24). On the other hand, CG1 received only tailored interventions to ameliorate cognitive and behavioral-related disturbances based solely on clinical interviews and baseline neuropsychological assessment (mean±SD: age = 72.3±9.27); whereas the CG2, received neither the system nor any tailored intervention (mean±SD: age = 74.8±6.62). They all granted their written informed consent before their participation in the study.

OPEN IN VIEWER

Fig.3

Randomization procedure and group allocation of participants (N = 18).

More specifically, the EG consisted of six participants of which four out of six were diagnosed with amnestic MCI (User 1, User 3, User 5, and User 6) and two with AD (User 2 and User 4). Four out of the six users have used the system successfully for four months (User 1-User 4) [23] while two of them for twelve months (User 5 and User 6). The timelines for each user are presented in detail in Table 2. As it was a long-term study, attrition issues were anticipated. To address this, we made a great effort to first screen participants and to maintain contact, including weekly in-person visits and as needed phone calls to subjects to maintain adherence to the research protocol and to ensure the minimum of dropouts. We asked all our participants whether they wished to take part in an interview and gave them verbal and written information about the study. All accepted to participate and gave their consent, with anonymity and confidentiality by all researchers.

OPEN IN VIEWER

Table 2

Timetable of EG participants, who used the system (n = 6). User 5 and User 6 are the long-term users

Our study aims to minimize selection bias by using randomization and allocation concealment. After the baseline assessment, eligible and consenting people with MCI and AD were randomized to one of the three groups, the EG, CG1, or CG2. Until the completion of the last long-term follow-up, only the project leader and two researchers who were responsible for the interviews, system installation, and deployment had access to the information on group allocation. Due to the nature of the group therapy programs, blinding of the participants and instructor was not possible. However, all the independent evaluators were blinded with respect to group allocation, while the participants were not informed of primary outcome measure or the study hypothesis. To maintain group allocation confidential, participants were requested prior to each assessment phase to not reveal allocation or therapy content to the evaluators. Participants scheduled for qualitative studies were told that they must not talk to the evaluators about participation in interviews and neuropsychological assessment. In general, the interviews and neuropsychological assessment were performed in a way that does not reveal participants’ allocation.

This particular study entails a limited group of six participants ranging from MCI (n = 4) to AD stage (n = 2). This is because the aim of the study is to focus on the long-term 1) maintenance of remote monitoring to support interventions for continuous improvement and 2) acceptance of technology by people with dementia. The long duration of the study, combined with the large equipment installation, maintenance effort, and cost, directed the choice of six participants with full equipment. Despite their limited number, the continuous 24/7 collection of data from all aspects of daily life (physical activity, sleep, behavior, etc.) and zero drop-off rate resulted in a large dataset. Meanwhile, looking into literature (described above), similar studies in the line of ICT solutions for people with dementia [74] have been conducted with a limited number of participants for analogous reasons, namely with a sample size of six [54, 86], five [37, 87], four [56, 88], or three participants [103]. Based on their valid methodology, and aiming for larger data volume instead of participant numbers, the present study recruited eighteen individuals, six in the experimental group (EG) and twelve in two control groups (CG1, CG2). Notably, as this study establishes the long-term maintenance, continuous improvement and acceptance of the system, it serves as a stepping stone to larger trials.

Regarding privacy and data storage, all data stored by the system are anonymized by nature. In detail, numeric sensor measurements are processed and aggregated as described above, to be transformed to meaningful events such as Sleep duration, Interruptions, level of physical activity, and ADLs. Even data from depth cameras are processed to extract movement vectors, object recognition, and localization information to recognize ADLs. The original raw data, sensor measurements and images (depth and color) that can be used to identify people are not stored but rather processed and deleted. The qualitative data of sleep, physical activity, and ADLs cannot easily be traced back to the users’ identity, which is only known to the study collaborators, for which user consent has been given. Third parties entering the premises, were a very rare case, as participants of the EG were meticulously and carefully chosen for living alone (due to system limitations of identifying actors), while their close family and friends visiting were thoroughly informed. Even so, third parties could not easily affect and produce observations in the system, as sensors were either wearable or attached on personal items (drug box and cabinets) and cameras were installed in designated spaces (e.g., kitchen), still not producing recorded identifiable data.

Study design, procedure, and evaluation method

During the initial clinical visit, a psychologist from the team interviewed and applied specific neuropsychological tests to all participants (EG, CG1, and CG2) and their caregivers. After determining whether participants meet the eligibility criteria, the psychologist visited them for a second time so as to better understand their needs, to explain the study objectives and to provide informed consent for their participation in the study. All participants were very confident while handling technology, hinting the opportunity to get them involved more actively in their self-management. Participants also commented on the chance to socialize and interact with clinicians through the novel proposed solution. None of the participants had privacy concerns about the use of the technology in a research context or about the digital monitoring of their lives being since this was done in a confidential and anonymized manner.

The three groups included people with cognitive impairment in the AD continuum and each participant was randomized and allocated in one of the three groups (EG, CG1, and CG2):

The EG received both the intelligent system combined with non-pharmacological interventions. The interventions were based both on objective clinical observations through the system and participant’s feedback and subjective statements, as well as on the baseline neuropsychological assessment.

During the next visits in participants’ homes, the psychologist conducted a semi-structured interview with participants of the EG, so as to gain better knowledge and determine their lifestyles, habits, needs, and quality of life, and also to study the architectural structures and features of their houses. Necessities, expected assistance scenarios, functions to carry out, and technological solutions were recorded on a form. For each case, the clinicians of the team discussed a sum-up based on the needs and problems and assigned the tailored adapted non-pharmacological interventions (Table 3). Figure 4 depicts the entire study design. As a result, the clinicians could monitor participants’ vital signs and daily disturbances not only based on participants’ subjective statements, but also by taking advantage of the interface to determine objectively specific parameters (e.g., sleep, physical activity) (Fig. 5). After the observational period of each participant, the psychologist conducted a semi-structured interview in order to investigate the participants’ opinion about the system and administered the System Usability Scale (SUS) questionnaire (Supplementary Material). In Video 1, a whole presentation of the proposed intelligent monitoring system as a holistic approach, followed by non-pharmacological interventions, as deployed in the EG can be found in the Supplementary Material.

OPEN IN VIEWER

Fig.4

Study design concerning participant roles, evaluation methods, and goals.

OPEN IN VIEWER

Fig.5

Example of different sleep patterns as illustrated in the interface. In the left side the clinician can select specific patterns while on the right side a comparison chart presents visually the selected variables (y axis) and their distribution through time (x axis).

Neuropsychological assessment

Table 7 shows the neuropsychological assessment performance of all groups (EG, CG1, and CG2) at two different times: baseline (before the observational period) and post-trial (after the intervention). Asterisks (*) indicate a statistically significant difference between the two groups’ scores and bolded p-value shows the statistically significant difference between baseline and post-trial metrics of the same group. We used Bonferroni correction for multiple comparisons which yielded a p-value of (p = 1–³ = 0.05). Baseline performance was equal between the groups and for all tests with no statistically significant difference among the scores of the three independent groups, i.e., participant groups were cognitive-matched. Additionally, Fig. 6 depicts the neuropsychological scores of the three groups (EG, CG1, and CG2) both at baseline and post-trial examination, with mean and standard deviation error bars.

OPEN IN VIEWER

Fig.6

Graph illustration of baseline (top) and post-trial (bottom) neuropsychological assessment of the three groups (EG, CG1, and CG2).

Indeed, statistics show that the group utilizing the proposed system-supported interventions improved significantly compared to the other two groups. In the post-trial assessment, the EG showed improvement in the majority of neuropsychological tests (TEA elevator time test, TRAIL-B, RBMT-recall, BDI) or at least maintenance, i.e., stability (in the ROCFT-recall test). Improvement was statistically significant in RAVLT total [from M(SD) = 38.67(13.53) to M(SD) = 45.83(15.94), p = 0.03]. Not only in itself, but also in comparison, EG performed significantly better than CG1 (according to MMSE) and CG2 (according to MMSE, RAVLT-learning, and PSS). In detail, the Mann-Whitney U test shows that the difference of EG in MMSE [M(SD) = 28.33(1.86)] compared to CG1 [M(SD) = 25.33(1.51)], is significant p < 0.05 as well as compared to CG2 [M(SD) = 25.17(2.79)]. Also, in RAVLT-learning, the difference between EG [M(SD) = 9.00(4.05)] and CG2 [M(SD) = 4.00(1.90)] is significant p < 0.001, as well as in PSS, where EG [M(SD) = 3.83(8.2)] and CG2 [M(SD) = 15.33(3.50)] yield p < 0.001.

Meanwhile, the plain interventions group still performs better than the no-intervention group, which in fact deteriorates as was expected. CG1 in itself shows improvement or maintenance in some tests (ROCFT-recall and RAVLT-copy spelling), but a deterioration in others (TEA elevator and elevator time, TEA phone, TRAIL-B, ROCFT-copy, and FAS). Overall, its performance significantly improves in RAVLT total [from M(SD) = 35.50(13.28) to M(SD) = 42.00(16.22), p = 0.04]. Moreover, it improves compared to CG2 as well. CG1 showed better performance in PSS [M(SD) = 4.33(4.13)] in comparison to CG2 [M(SD) = 15.33 (3.50), p < 0.05]. The expected deterioration with no intervention is shown through statistically significant decline within CG2 in MMSE [from M(SD) = 27.00(3.16) to M(SD) = 25.17(2.79), p = 0.04], and RAVLT-recall [from M(SD) = –2.17(1.17) to M(SD) = –4.0(1.79), p = 0.04].

Clinical and system-generated observations for the EG

To investigate whether or not the proposed system and the adaptive clinical interventions can have positive long-term effects on participants’ physical and cognitive function, paired-sample t-test analysis was used to find improvement and changes observational period regarding participant’s sleep, physical activity, and daily functionality. The level of significance we set was α= 0.05.

This section presents observations via sensor monitoring and intelligent processing as the clinicians received them. They interpreted them while also making their observations via interviews, as in the other two groups. The usefulness and the correlation of system-generated and clinical observations for the EG group are discussed in the following subsections, each concerning different daily monitoring domains such as sleep patterns, physical activity, and ADLs. Notably, only the EG can be examined across these monitoring domains, since only this group had the advantage of utilizing the monitoring system for sleep, physical activity, and daily functionality.

To facilitate comparisons, the entire observational period was split into two periods, which from now on will be referred to as the 1st half and the 2nd half. For Users 1–4 with a total duration of 4 months, the two periods last 2 months each, and for Users 5–6 the extended periods last 6 months each. As interventions start to accumulate, changes are expected to be measured in the 2nd half. The following subsections examine observations in the domains of sleep, physical activity, and ADLs. Overall outcomes are reported for all Users and extensive observations for the prolonged users completing the study in this paper, User 5 and User 6. For the detailed observation of shorter, previous trials in sleep, physical activity, and ADL domains the reader is referred to [23].

A previous study [23] examined in detail Users 1–4 at an earlier stage for a shorter duration of four months. The main focus of the present paper is the longer duration, maintenance, and acceptance of the system and its outcomes. The present study expands with details of two new participants, for the much longer duration of one year, and accumulates results for the total of six participants with full installations (EG) and two control groups of equal size. Here, the examination of Users 1–4 provide sufficient background to show that findings for the two most recent participants, Users 5–6, monitored for the longer duration, are in agreement with the four previous ones. To show that, we included the detailed statistical analysis of User 5 and User 6 while comparing the neuropsychological tests of all six participants of the EG, compared to CG1 and CG2 (Tables 7 and 8). The detailed monitoring and intervention outcomes of the previous study in Users 1–4 are presented here briefly for completeness [23] concluded that increased physical activity had a significant impact on multiple sleep parameters. Also, improvement of daily functionality was linked also with a better cognitive state, as measured by neuropsychological assessment, after the intervention period, i.e., the 2nd half of the study. More specifically, most of the values in daily physical activity of the four participants (Users 1–4) were low during the 1st half. On the contrary, after the intervention program, higher values were yielded. The statistical analysis has confirmed this significant improvement over time regarding physical activity and ADLs for all participants. In particular, regarding User 1 statistically significant better performance was found in the duration of kitchen presence, which drives the way to suggest that the participant was spending more time for meal preparation and cleaning during the 2nd half, which can be explained by increased interest and involvement in daily tasks and house chores. Also, User 2 and User 4 were found to have statistically improved in bathroom presence and taking care of their personal hygiene, since more frequent visits were observed during the 2nd half. In this vein, User 3 was found to have been significantly improved regarding moving intensity, while all users (User 1–4) were found to use TV less during the 2nd half compared to the 1st half.

Sleep patterns

The sleep patterns’ data, from the Sleep Sensor, were also used to obtain estimates of sleep parameters, including the total time of sleep, light (shallow) and deep sleep duration, the number of interruptions, REM sleep activity, total time awake in bed, and sleep-onset latency. To investigate our hypothesis that the proposed system followed by adaptive clinical interventions can have positive effects on participants’ physical and cognitive function, paired-sample t-test analysis was used to discover significant changes in observations between the 1st half and 2nd half of a participant’s sleep, physical activity, and ADLs.

All users had significant improvement in the majority of sleep parameters (increased duration of total time asleep, decreased the duration of sleep latency, decreased number of interruptions and sleep latency, increase the duration of deep and shallow sleep during nighttime activity). Regarding sleep, observations for both User 1 and User 2 showed that there was a significant increase in the duration of total time asleep and deep sleep duration during the 2nd half compared to the 1st half. In particular, User 1 found to have a significant decrease in sleep latency and shallow sleep during the 2nd half. Regarding User 3 and User 4, there was a significant decrease in the number of interruptions and sleep latency. Also, User 3 compared to the observation in the 1st half, found to have a significant increase in deep sleep duration during the 2nd half. While the study in [23] confirms the abovementioned improvements for Users 1–4, this study shows that the results extend and are coherent with long-term users (User 5 and User 6), who used the system for a much longer period, also demonstrated a significant improvement in the majority of the sleep parameters. More specifically:

Regarding User 5, statistical analysis revealed a significant increase in the duration of total time asleep in the 2nd half (M = 7.10 h, SD = 3.41) compared to the 1st half (M = 5.98 h, SD = 4.18); t (125) = –2.25, p = 0.02, decrease of number of interruptions in the 2nd half (M = 2.4, SD = 2.50) with respect to the 1st half (M = 3.72, SD = 2.91); t (176) = 4.54, p < 0.0001 and increase of shallow sleep duration in the 2nd half (M = 3.35 h, SD = 1.98) compared to the 1st half (M = 3 h, SD = 1.97); t (186) = 3.67, p < 0.0001. Although, the time spent in bed awake as well as sleep latency decreased in the 2nd half (M = 23.58 min, SD = 22.7) and (M = 5.85 min, SD = 5.1) with respect to the 1st half (M = 26.15 min, SD = 25.18); t(182) = 1.040, p = 0.30 and (M = 23.58 min, SD = 7.18); t(168) = 0.54, p = 0.59 respectively, there was no statistically significant difference between metrics of the 1st and the 2nd half. In this common line, deep sleep duration also increased during the 2nd half (M = 1.42 h, SD = 0.75) compared to the 1st half (M = 1.35 h, SD = 0.98); t(139) = –0.62, p = 0.54, without any statistical significant difference.

Regarding User 6, we observed statistically significant improvement in all sleep parameters. More specifically, there was a significant decrease of sleep latency during the 2nd half (M = 8.4 min, SD = 13.15) than in the 1st half (M = 8.8 min, SD = 10.06); t (239) = 2.3, p = 0.02, decrease of total time in bed awake in the 2nd half (M = 66.9 min, SD = 59.9) than in the 1st half (M = 97.3 min, SD = 62.01); t (172) = –4.95, p < 0.0001, increase of deep sleep duration during the 2nd half (M = 92.9 min, SD = 53.2) than in the 1st half (M = 68 min, SD = 43.4); t (163) = –4.49, p < 0.0001 and increase of shallow sleep duration in the 2nd half (M = 3.1 hours min, SD = 1.82) than in the 1st half (M = 2.67 h, SD = 1.54); t (161) = –2.22, p < 0.02. Also, there was a significant increase in the duration of total time asleep in the 2nd half (M = 7.58 h, SD = 4.20) than in the 1st half (M = 6.24 h, SD = 3.24); t (158) = –3.00, p = 0.003. Finally, there was a significant decrease of the number of interruptions in the 2nd half (M = 4.75, SD = 3.65) than in the 1st half (M = 3.67, SD = 3.77); t (251) = –3.33, p = 0.001.

The improvement in the above-mentioned sleep parameters is confirmed by the day-to-day monitoring system observations on the Clinician Interface. Based on the system’s visualizations of a day’s observation, as shown in Summary, Per Minute resolution, both participants, User 5 and User 6, had sleep problems as they also stated to the psychologist during the first and the subsequent visits of the 1st half period. Our participants, in the beginning, were for more than an hour in bed awake; their total shallow sleep duration was less than 3 h, and their Deep Sleep duration was limited; almost absent. Furthermore, more than four interruptions during nighttime sleep activity were detected, a situation which is commonly observed in people with cognitive impairment [118, 119]. On the contrary, during the 2nd half, fewer problems were noted in almost any given daily view of Summary, Per Minute, such as the one in Fig. 7. As shown in Fig. 9, User 6 showed great improvement in the majority of sleep parameters during the 2nd half of the observation. In the 2nd half, the clinician observed through the system, decrease in the duration of total time awake in bed and increase in the duration of the total time of deep and shallow sleep as well as total time asleep (Fig. 8).

OPEN IN VIEWER

Fig.7

Detected sleep qualities (number of interruptions and sleep latency) for the whole observation period (1st - upper image and 2nd half - lower image) of the User 6, as illustrated in the Clinician interface. The x axis presents the months, while the y axis the number of interruptions and sleep latency, respectively.

OPEN IN VIEWER

Fig.8

Detected sleep qualities (total time in bed awake, total time light and sleep, total time asleep, sleep latency (seconds), and number of interruptions) from two indicative days of the 1st half (A) and the 2nd half (B) of the User 5 participant, as shown on the Clinician interface, One Day Summary section.

OPEN IN VIEWER

Fig.9

Specific sleep patterns (total time in bed awake, total time deep sleep, sleep latency, and number of interruptions) of the User 6 for the whole observational period (1st and 2nd half) as shown on the Clinician interface in per Month Comparison section. The x axis presents the months, while the y axis presents the number of interruptions, sleep latency in seconds, total time deep sleep, and total time in bed awake duration in hours and minutes.

Furthermore, by using the interface of the system, the clinician was able to define specific sleep parameters and set thresholds in the Dashboard section for each activity (e.g., REM). For instance, sleep duration: 6 h, number of interruptions: (more than) 3, sleep latency: (more than) 30 min, days of reoccurring problems: 3. Thus, regarding User 5, it was revealed through the interface that there was a substantial gradual decrease of the number of interruptions, short sleep duration and total time awake in bed during the whole observational period (January-March 104 problems detected, April-June 72 problems identified and July - Oct 66 problems) (Fig. 10a). With regards to the User 6, a noticeable decrease of number of interruptions and total time awake in bed was detected, with an increase of light and deep sleep duration at the same time (April-May 181 problems detected, July-September 180 problems, October-December 131, January-March 95, April 85 and May-June 44 (Fig. 10b). In comparison with our Users 1–4 [23], Users 5 and 6 had fewer problems after the observational period of one year.

OPEN IN VIEWER

Fig.10

Detected sleep problems (in bed but awake, short sleep duration, and large number of interruptions problems) for the whole observational period (1st and 2nd half) of the User 5 (A) and User 6 (B), as shown on the Clinician interface, per day summary section.

With sleep parameters being of special interest and in order to identify which particular parameters are correlated with specific neuropsychological tests, we conducted Pearson correlation between neuropsychological tests and mean values from several sleep parameters. The most significant and important correlations between sleep variables and post-trial neuropsychological assessment from EG are presented in Table 8. It is clear that after the 2nd half, the post-trial assessment of CDR was negatively correlated with total time asleep, which indicates that EG participants who were spending more hours at nighttime sleep activity, had better CDR scores (lower values) and consequently better overall cognitive function. Moreover, Total Time Asleep duration found to be highly correlated with the RAVLT learning score, a finding which highlights also the importance and the impact of more sleeping hours at nighttime in the learning process. Correlation between ROCFT copy and total time shallow sleep provided a rich set of quality information about our EG participants, where results indicate positive correlations regarding these two metrics.

Physical activity and activities of daily living

In general, the main difficulties in performing ADLs at home observed in our participants were found to include item selection and passivity in initiating actions autonomously. Our participants had very decreased physical activity (measured by the system as moving intensity) in the 1st half, and they did not get involved in housework such as cooking or cleaning. Instead of active participation in multiple activities, they preferred to stay at home and watch TV for many hours. The majority of our participants had problems to complete household chores. More specifically in the initial visit, they indicated worries about physical health and trivial matters (e.g., doing household chores correctly), as well as physical symptoms, including feeling restless and jittery. Therefore, we decided that the visualization of daily physical activity and monitoring of daily activities could provide clinicians with relevant and useful information regarding the progress of the participants in time. Also, in Comparison per Day or per Week section, correlations between several sleep patterns (e.g., number of interruptions and total time asleep) or/and comparisons between electrical appliances (smart plugs and tag sensors) could also provide clinicians with useful information regarding daily activities and if a prolonged activity (e.g., TV use) negatively or positively affects another activity (e.g., sleep total duration) (Fig. 11).

OPEN IN VIEWER

Fig.11

Correlation between TV use and sleep patterns (Sleep Latency and Number of Interruptions) for the User 5. The x axis presents the specific duration while the y axis presents the number of interruptions (blue line), the sleep latency (black line) in seconds, and the TV use (green line) in hours and minutes. The arrow highlights an interesting observation of increased TV usage linked to sleep latency and interruptions.

Graphs for a specific period highlight specific problems regarding physical activity during the 1st half and significant improvement after adaptive clinical interventions (e.g., exercise programs, dance, daily exercise, etc.). Most of the values in Daily graphs of physical activity for our participants were very low in the beginning while after the intervention program and installation procedure, they were significantly increased. The statistical analysis also confirmed the significant improvement over time regarding physical activity and activities of daily living for all participants.

Regarding User 5, even though there was an improvement in the majority of the domains above significant difference was found only between 1st and 2nd half scores of kitchen presence. The participant spent many hours in kitchen for meal preparation and cleaning in the 2nd half (M = 2.40, SD = 1.41) than in the 1st half (M = 1.54, SD = 0.62); t (63) = 4.63, p < 0.0001. Moreover, statistically significant improvement was found regarding physical activity (moving intensity) during the 2nd half (M = 43.26, SD = 10.29) compared to in the 1st half (M = 13.09, SD = 12.19); t (51) = –2.08, p = 0.04. However, no statistically significant differences were found between 1st and 2nd half regarding washing machine, TV use, and iron use (Fig. 12).

OPEN IN VIEWER

Fig.12

Comparison chart of physical activity (moving intensity–orange line) and sleep parameters for the User 6. The x axis presents the duration of the observational period while y axis presents the total time asleep, the total deep sleep duration, and total shallow sleep in hours and minutes.

Regarding User 6, there was an improvement in all domains of ADLs. Observations in the 1st half indicated that User 6 could not manage to perform many of the household chores independently, but was in need of support from others. If she started to do them, she spent less time and did not complete the majority of them successfully, especially complicated tasks (e.g., cooking, ironing). During the observational period, it was obvious that User 6 had a statistically significant improvement concerning ADLs. The clinician noted positive changes, through the Clinician Interface, in physical activity, laundry, TV usage, and ironing. Statistically significant difference was found with regards to 1st and 2nd half of the kitchen presence duration, since the participant spent many hours in kitchen for meal preparation and cleaning in the 2nd half (M = 2.40, SD = 1.41) than in the 1st half (M = 1.54, SD = 0.62); t (63) = 4.63, p < 0.0001.

Combination of multiple monitoring domains

One of the most important correlations between metrics or between different sleep patterns of our participants is presented in Figs. 13 and 14, based on the visualizations from the Clinician Interface. These figures show the long-term impact of non-pharmacological interventions in specific patterns. Figure 13 shows how the system presents and measures a poor night for User 6, with more than 7 interruptions and an hour of sleep latency, and a healthy one with less than 5 interruptions and sleep latency less than 8 minutes. Moreover, correlation between deep sleep duration and total time asleep provided a rich set of quality information about the User 5 with cognitive impairment during the whole observational period, which supports a trend of correlation between these two parameters (Fig. 14), similarly to the User 6, where expected results observed regarding long-term improvement in different sleep patterns (Fig. 15).

OPEN IN VIEWER

Fig.13

Comparison per day chart between sleep patterns for User 6. The graph shows the correlation between Number of Interruptions (blue line), Sleep Latency (black line) in seconds, Total Time Asleep (green line), and Total Time in Bed but Awake (orange line) in hours and minutes (y axis) per day (x axis). The arrows show how the system presents a poor (red) night for User 6 with high latency, low total sleep and high interruptions and a healthy (black) night sleep with low latency and interruptions.

OPEN IN VIEWER

Fig.14

Correlation between sleep metrics (Total Time Asleep and Deep Sleep) for User 5. The x axis presents the total duration of sleep, while the y axis presents the total time of deep sleep duration, both in hours and minutes.

OPEN IN VIEWER

Fig.15

Comparison per day chart between sleep patterns for User 5. The graph shows the relationship between multiple variables (Number of Interruptions, Sleep Latency, Total Time Asleep and Shallow and Deep Sleep Duration, Total time in bed but awake) in a specific time period where all sleep duration metrics increase.

User statements, SUS, and feedback for the long-term use of the proposed system

The perceived usability and user satisfaction of the device were measured using SUS [120]. The SUS provides a “quick and dirty”, reliable tool for measuring the usability of a system or a device. It consists of a 10-item questionnaire with five response options for respondents; from strongly agree to strongly disagree. The scale was administered in EG at the end of the observational period. The SUS scores were calculated according to the standard way of calculation of this questionnaire (https://www.usability.gov/how-to-and-tools/methods/system-usability-scale.html), namely by assigning a relative score to each item and performing a calculation with their sum.

The SUS mean score for the EG was above average (EG = 81.6, M = 70.0) suggesting users accept the system and find it useful. Table 10 presents the SUS total score as well as the responses to each item from the EG. More specifically, the majority of them reported that they would like to use the system frequently (Item 1) and they found quite easy to use it (Item 3). Only User 2 reported that he would need the support of a technical person to be able to use this system.

Moreover, after the observational period, a semi-structured interview took place between the psychologist and participants in order to find out any pros and cons of the long-term use of the system by the participants (EG). Despite the fact that the literature supports the argument that, for some older adults, technology monitoring is not acceptable in any format [121], in our study all users recognized our system as non-obtrusive and easy for long-term use. In Table 11, some testimonials are presented.

Limitations

Nevertheless, there are a few limitations in this study that should be considered for future research. To begin with, our limited number of participants is for now compensated by the huge amount of data generated, analyzed, and used for each one of the participants. On the other hand, based on the specific characteristics, requirements and the long (one year) duration of the study, meaningful and objective clinical observations of improvement have been made and these interesting outcomes of our study will shed light to the field of Dementia and Assistive Technology. Monitoring was complex and continuous; it involved a large number of heterogeneous sensors, data collection around the clock (24/7) for several months and up to a year for all participants. Similar published studies in the line of ICT solutions for people with dementia have been applied to six [86] and even fewer participants [74, 103]. Therefore, the number of six participants in the full-installation and intervention group was chosen as the minimum stepping-stone to larger studies. Towards expanding to more participants, we plan to include more people, in collaboration with patient organizations such as Alzheimer Hellas and include people with other types of dementia (i.e., Lewy body dementia, frontotemporal dementia, etc.) and cognitively impaired patients at various stages of disease (i.e., subjective cognitive impairment, severe dementia, etc.). Therefore, more research is needed to determine the long-term effects of combined intelligent monitoring technologies and non-pharmacological interventions in general elderly populations and large-scale pilots with people with cognitive impairment within AD. Given the reflections of intelligent monitoring effects combined with non-pharmacological interventions on cognition and behavior, it would be of great importance to also investigate changes of other well-established biomarkers (e.g., brain volumetric changes) so as to gain a holistic view of people with cognitive impairment in all aspects. Moreover, future studies could consider the fact to test the combination of the intelligent monitoring system with non-pharmacological interventions to other types of cognitively impaired people (e.g., Central Nervous System disorders) to gain knowledge about its potential benefits to such population. An increased number of participants could help investigate algorithmic models to automatically suggest interventions intelligently according to context.

Conclusion

There are many challenges highlighted in our study such as long-term adherence, engagement with technology, deployment of non-pharmacological interventions based on system’s visualizations, and making new activities (e.g., use of UI) part of participant’s daily routine for a long period. Therefore, our study presents the integration of an ICT home-based system that incorporates technology, so as to maintain and even enhance daily functionality, cognitive function, well-being and QoL of people with cognitive impairment related to AD. Our data show that with appropriate design, participant training, and adequate monitoring, a smart home system can capture regular data on sleep, daily functionality, and physical activity in large multinational studies.

The “smart home” system implemented aimed to enable nonobtrusive monitoring of its residents for a long-term period. Therefore, it involved different levels of technological sophistication, ranging from wearable devices to smart environments that continuously monitor residents’ activities and physical status and adapt to residents’ needs. The system proved to be capable of providing information on current patient behavior that can help quantify the patient’s cognitive impairment, estimate the patient’s self-dependency, and facilitate formal and informal care of cognitively impaired patients living alone. In detail, the system enabled individualized interventions which improved several cognitive functions compared to other groups. As a result, the proposed system is scalable, sustainable, and leads to reduced consumption of clinical time and hospital resources, and increased home residence, self-care, and self-management. We hope that the outcome of our study, together with further long-term investigation and research, will justify additional interventional studies with a focus on user-acceptance.