This study and related documents were approved by the local Ethics Committee (Comitato Etico IRCCS [Istituto di Ricovero e Cura a Carattere Scientifico] Regione Lombardia, Sezione IRCCS Fondazione Don Carlo Gnocchi, ID 07_07/04/2021, April 7, 2021, amended on May 26, 2023). Due to severe dementia, informed consents were obtained from legal tutors of the participants. Any data or information was recorded in an anonymized form. No compensation was provided to participants.

Participants were recruited from the Alzheimer Unit of the Palazzolo Institute, Don Carlo Gnocchi Foundation, Milan. As it is a pilot study, no minimum sample size has been identified, but all resident guests of the experimental site were contacted for possible enrollment.

The inclusion criteria were an autonomous locomotor capacity (even with the help of walking aids such as, for example, a cane or a crutch) and the frequentation of common areas of the residential unit. The exclusion criteria were the presence of behavioral disorders incompatible with the research objectives (delirium, psychomotor agitation requiring psychoactive pharmacological interventions, acute clinical events limiting autonomous movement, and nonneurological pathologies strongly limiting motor activities), the need for assistance for walking by a caregiver, any dermatological problems preventing contact with wearable sensors, any confined conditions or quarantine for COVID-19 risk management.

Reference data from age-matched healthy participants were extracted from previously published datasets [11,30] in order to build a control group with a comparable average age.

The chosen assessment battery included Hendrich II Fall Risk Model [31], Tinetti Scale [32] when applicable, Barthel Index [33], Mini-Mental State Examination [34], Clinical Dementia Rating [35], and Scheda di Osservazione Intermedia di Assistenza [36]. Moreover, an Event Monitoring Diary (clinical events, changes in pharmacological and nonpharmacological interventions, falls, discomfort, and/or incongruous interaction with sensors, and temporarily leaving the ward) was compiled during the experiment.

Participants wore, at their dominant wrist, a wristwatch-like waterproof device including a triaxial accelerometer for 1 week. Accelerometric raw data (measurement unit is one thousandth of earth gravity constant g, termed in the following as milli-g), sampled at 50 Hz, were stored in a resident memory (Geneactiv, Activinsights). The device was not to be removed unless the participant was willing to. The 3 acceleration components allowed for the computation of the vector norm, from which the motor activity (MA) index was derived as the SD of the vector norm in 1-minute epochs [37], thus resulting in a 7-day MA profile composed of 1440 values per day (Figure 1).

A sorting operation on the MA profile produced a distribution plot (Figure 2) which, through proper thresholds (10 milli-g for transition from no to light physical activity [LPA], 100 milli-g for light to moderate, and 350 milli-g for moderate to vigorous, for a sensor located on the nondominant wrist) [38], eventually corrected for dominant side (previous threshold values multiplied by Coeff_Dominant=1.16926), allowed computation of physical activity indices: the percentage duration of the phases with no physical activity (NPA), LPA, moderate physical activity (MPA), and vigorous physical activity (VPA).

An additional analysis was performed on the raw acceleration profile to detect the occurrence of regular patterns that are likely to be associated with a pseudoperiodic function such as locomotion. The raw signal was windowed in 4-second epochs, and an autocorrelation analysis was applied to the signal window. The possible occurrence of lagged peaks of the autocorrelation function was searched, adopting a threshold of 0.7 (a perfect periodic phenomenon has a unitary peak value, while white noise is denoted by null values of the autocorrelation function), which had been set, based on trial-and-error attempts [39]. The duration of such pseudoperiodic behavior (PPB) phases was quantified by their percentage duration (PPB).

Two more indices were computed to quantify the quality of nighttime sleep and daytime wake time: nighttime sleep efficiency (SEff), the percentage duration of NPA phases between 11 PM and 7 AM, the a priori “night” time interval, which is related to SEff [40]; analogously, daytime wake efficiency (WEff), the percentage duration of LPA, MPA, or VPA (ie, the complementary duration of no activity phases) between 9 AM and 7 PM, the a priori “day” time interval, which quantifies WEff. The choice of the strict timing for “day” and “night” time intervals was based on the daily routines established within the long-term care unit: all guests are expected to be in bed during the “night” interval and awake during the “day” interval, while this assumption is not true for transition intervals (7 AM-9 AM and 7 PM-11 PM).

The chosen location tracking system is composed of individual wearable devices and a network of environment fixed-position antennas (TapMyLife Inc). Monitored participants wore a miniaturized Bluetooth beacon (dimensions 32 mm × 23 mm × 3.3 mm, weight 4.0 grams) at one of their ankles by means of a biocompatible plastic bracelet. Such a device continuously emits a unique ID code, possibly received by a Bluetooth receiving antenna positioned nearby (according to the Bluetooth technology, approximately up to 10 meters in open space). All environments where participants resided (private bedrooms and any common spaces they spend time in, such as dining rooms, living rooms, corridors, etc) were equipped with Bluetooth receiving antennas: 1 antenna for each private room (PRI; 13 double rooms and 4 single rooms), 2 antennas for each large public room (3 PUBs), and 7 antennas along the main corridor, as shown in Figure 3. The frequency of the Bluetooth emitter is partly determined by the asynchronous Bluetooth protocol, generally in the magnitude order of 1 Hz. As a single Bluetooth emission can be received by several antennas, a proprietary signal algorithm identified the closest antenna and updated the beacon location (ie, the identified receiving antenna position) only when the beacon switched to a new antenna. Therefore, the source raw data provided by the location tracking system was natively asynchronous, and the current location of a beacon was determined by the position of the last antenna that detected the beacon itself. In order to summarize the location data, the percentage values of time spent in PRIs, common living rooms (PUBs), and corridors were computed.

The single monitoring experiment lasted for 7 days. Three experimental sessions were performed in July (Session A), September (Session B), and October (Session C) 2023. For short-term repeatability assessment (3-week time interval), Sessions B and C were compared. For long-term repeatability assessment (3-month time interval), the Sessions A and C were considered.

The variables computed starting from the actigraphic recordings were:

Indices were computed on the 7-day-long recording. When the full-length recordings included a portion of scratch recording for any reason, for example, because of participants taking off the sensor, the scratch portion was removed from the recording, and the remaining time duration was further reduced to the highest possible multiple of 24 hours. The previous algorithm allowed for the computation of the indices on recordings whose length was a multiple of the basic 24-hour circadian cycle.

The difference between outcome variables as obtained from control data and from pooled experimental data was quantified as percentage variation of median values and presented as boxplots. The difference was assessed by the Wilcoxon rank sum test (as implemented by “ranksum” function in MATLAB software, version R2016a, The MathWorks Inc), a nonparametric test for 2 independent populations.

Test-retest reliability was presented as scatterplots and assessed through the intraclass correlation coefficient (ICC). ICC was computed with a 95% CI, using a 2-way mixed-effect, absolute agreement, single measurement model. Significance level was set at 0.05. The significance level was corrected for multiple comparisons according to Bonferroni-Holm.

The adopted guidelines for interpreting ICC [41] meant “poor” for values less than 0.40, “fair” for values between 0.40 and 0.59, “good” for values between 0.60 and 0.74, and “excellent” for values between 0.75 and 1.00.

Instrumented methods for physical activity detection and tracking are valuable tools for studying individuals with dementia [42], complementing self-report measures such as questionnaires and surveys.

The presented method has provided a quantification of various aspects concerning physical activities during wake time, sleep, and social interactions among the components of the participants’ group. Besides the characterization of average behaviors as quantified by the various indices, it is worth noting that distinctive individual behaviors, different from the average patterns, were consistently identified as well, as shown also by the high test-retest reliability of the computed indices. In fact, this sensitivity aspect supports the use of such methods not only in group studies but also in clinical settings where accurate individual characterization and management are essential.

The indices considered in this study were intended to depict a multidimensional figure of each measured participant. Along with indices related to levels of physical activity (NPA, LPA, MPA, and VPA) [38], indices concerning sustained walking (PPB), social interaction (CC) [11], quality of nighttime rest (SEff), and daytime vitality (WEff) were included.

The indices’ definitions are generally self-explanatory. Only the rigid identification of wake-time as daytime and sleep-time as nighttime adopted in the wake and SEff indices may be regarded as a critical choice. An alternative choice would be a functional identification of wake epochs and sleep epochs [43]. In fact, the relatively simple original definition of SEff in sleep studies (percentage of time asleep on the total time spent lying in bed, with these 2 measures obtained, respectively, from EEG (electroencephalography) polysomnography and observation log) required a redefinition of the indices when wrist accelerometry was introduced [44]. Interestingly, a literature review [45] has highlighted that the use of wrist accelerometry in sleep studies offers several advantages, such as high participant compliance, but also drawbacks, including the potential under- or overestimation of sleep indices. In this study, a rigid timeline approach was adopted to meet the requirements of clinical partners, who were specifically interested in monitoring residents’ compliance with the daily routines established within the long-term care unit. According to these routines, all guests are expected to be in bed during the “night” interval (11 PM-7 AM) and awake during the “day” interval (9 AM-7 PM). The “day” and “night” time-intervals are separated by 2 transition time intervals, during which “go-to-bed” (7 PM-11 PM) and “get-out-of-bed” (7 AM-9 AM) events may occur. Interestingly, these events do not occur simultaneously across the guests. Consequently, the inclusion of the transition intervals within either the “day” or “night” intervals would have implied the risk of false positive or false negative classifications. Importantly, omitting the transition intervals does not compromise the reliability of the efficiency indices, which are computed over time windows of 8 hours (“night”) and 10 hours (“d”). These durations are consistent with the minimum actigraphic recording time of 8 hours required to obtain reliable indices [46]. Therefore, we consider the rigid timeline approach reasonable and justify the inclusion of the transition intervals as a strategy to improve the reliability of the computed indices.

The indices obtained in the experimental group were compared with proper age-matched normative data extracted from already available datasets [11,30]. The only exceptions are for the CC index whose normative data were from a younger group (mean age 72, SD 7.6, range 58-87 years) [11], and for the location tracking indices which were based on a specific local network of BLE antennas and therefore did not allow neither the recording of normative data in the long-term care residential unit, nor the option to set up a duplicate of the residential unit for healthy older adults, was considered.

The installation of the system of BLE antennas was performed in the weeks preceding the experiments without major interference with the daily flow of activities in the care unit. The accuracy of the system (ie, the correct classification of the room where the specific BLE beacon was in) was obtained as a result of a calibration procedure following the installation. The experimental data collection was carried out smoothly without any potentially confounding events, particularly, as described by the clinical logs, the participants constantly resided in the long-term care unit without leaving it. The participants occasionally received nursing assistance in their activities, particularly hygiene-related ones, in which partly passive movements could be interpreted as LPAs. However, these nursing activities were limited in time (few minutes), therefore were not considered to alter the experimental outcome significantly. All accelerometric recordings were a posteriori checked for identifying sensor removal (a constant baseline clearly distinguishes these from actual measurement, even from a still sleeping participant) and eventually tagged to exclude these time intervals from further analyses.

The indices related to physical activity (NPA, LPA, MPA, and VPA) evidenced significant differences between the group of persons with dementia and age-matched healthy participants: a slightly larger, though not significantly different after P value correction, duration of phases of NPA, a comparable LPA, a significantly reduced duration of MPA, and a substantially absent VPA. This figure is compatible with a commonly expected behavior in a nursing ward: increased sleep or rest, even in a seated position, particularly during the day [47], and drastically reduced activities requiring a relevant physical involvement [48]. A perusal of the individual data allowed us to identify individuals who showed activity profiles that differ from average behavior: particularly, 2 cases evidenced a very low value of NPA, which was compensated in 1 case (S18) by large values of LPA, while the other case (S23) evidenced the largest values of MPA. Interestingly, S18 had the lowest Barthel score among participants, while S23 was the youngest and had the lowest Tinetti score.

The index PPB, related to a pseudoperiodic motor behavior, is to be interpreted as an index of steady walking. Indeed, while the PPB index was, on average, not significantly different between controls and this study’s group, a small subgroup of participants showed higher than normal PPB (S25, S23, and S07 above all controls in all 3 sessions, S13, S06, and S09 above at least 9 of 10 controls in all 3 sessions). The motor behavior of these participants resembles the so-called “wandering” behavior [49,50], a continuous, uninterrupted walking along the allowed paths, such as up and down a main corridor. Specifically, in this study’s group, 6 individuals had this wandering behavior, clearly detectable in PPB scatterplots. Interestingly, this behavior is very consistent and can be observed also in repeated sessions, weeks, or months apart.

The CC index, quantifying the accordance between MA profiles of any possible couple of participants, is the key variable in the so-called social actigraphy approach [11] that aims at identifying congruent synchronized behaviors induced by the social proximity (in the original paper due to the marriage or cohabitation, in this study due to the common belonging to the community living in the long-term residential unit). In the presented data, the CC index showed significantly lower values in the experimental group, compared to reference data published in literature [11]. This result was initially unexpected, since a simple assumption implies that a tight social community, such as the one composed of guests of a nursing home, should be characterized by higher between-subjects CC values. The presented data evidenced, on the contrary, that participants with dementia show larger between-subjects differences in the MA profiles (corresponding to lower CC values) than observed in healthy participants. This outcome is surprising when the attention is all upon the social context, with common activities such as having meals together or participating in socialization events, since it would imply a certain “temporal synchronization” of motor behaviors. On the contrary, this result could be explained by the fact that people with dementia are known to often show apathy and withdraw from social interactions [51], and sometimes individuals with dementia may even show sociopathic traits [52].

The index SEff showed, on average, comparable values between healthy participants and persons with dementia. Interestingly, a large number of this study’s group was characterized by lower-than-normal values, thus evidencing a troubled management of the hours eligible for sleep or at least rest. The repeatability of these lower-than-normal values lets us identify a subgroup of the guests of the residential ward that need special attention concerning the management of the nighttime. Without entering into specific details, several actions can be performed, including a revision of pharmacological or nonpharmacological interventions [53-55]. Nonetheless, the present method lets us consider also more practical choices such as a change of the bedroom and related roommate. The index WEff evidenced a tendency, not statistically significant after P value correction, of lower-than-normal values for people with dementia, implying reduced percentage of activity time and increased resting, possibly daytime napping [56]. Symmetrically with previous comments on SEff, it is possible to identify a small subgroup that evidenced a higher-than-normal WEff, thus presenting themselves as the guests who never sleep during the day and are always active.

The indices related to location tracking (PRI, PUB, and corridor), that is, the percentages of total time spent in specific indoor locations, have no reference normative data. The interpretation of the observed values relies on comparison with other indices. The PRI values, representing the time spent in PRIs, were found to be slightly larger than NPA, the time characterized by no activity. This result is consistent with the fact that PRIs were mostly used for resting and sleeping. Conversely, the sum of PUB and corridor durations (ie, the time spent in any public space) was found to be slightly lower than the sum of LPA, MPA, and VPA durations (ie, the time spent at any level of activity), thus implying that MA might occur even in PRIs.

The ICC values of the physical activity indices (NPA, LPA, and MPA) were excellent either in the short-term or in the long-term, with just a slight decrease of the values in the latter case. No ICC was computed for the VPA because vigorous activity was substantially absent in all cases. When considering PPB as a proxy for the amount of performed steady gait, the ICC was excellent in both short-term and long-term, and, to our surprise, it scored a higher value in the long-term. We are not willing to interpret this latter point as significant of some underlying phenomenon; rather, we consider it the result of a stochastic effect. Lower, but still good, ICC values were observed for the social actigraphy index CC. In fact, the CC values were substantially unchanged in all sessions, both after 3 weeks and after 3 months. As the CC values depicted a figure of a substantial absence of relations among the MA profiles of participants, even compared with control data, our interpretation is that people with dementia living in long-term care residences show signs of social withdrawal in their lifestyles as well as peculiar behaviors that may occur, such as the wandering one that largely differs from the inactive one which characterizes the majority of guests. Excellent ICC values were observed for SEff in the short-term. Indeed, participants tended to keep their night habits, which included either high SEff behaviors or persons who spent a large part of the night asleep. It is to be noted that the clinical management rules of the experimental residential unit excluded a general use of antipsychotics [57], thus increasing the possibilities for a person to be free to stay awake and even wander, as it was seldom observed, during the night. The test-retest reliability of WEff was between good and excellent. As for night habits, even during the daytime, participants had different behaviors (some stayed awake the whole day and others had frequent day naps). Those behaviors were individually consistent and repeatable. The location tracking variables, that is, the percentage time spent in different locations inside the residence unit, showed mainly fair values of reliability. The relatively lower values of the location tracking indices’ ICC disagreed with the higher ICC values of the accelerometer-based indices. It is an open question how much of this low reliability of the location tracking indices is due to the actual, and different from the acceleration, nature of measured variables, and how much is due to a different behavioral performance. As, occasionally, we observed that the system delayed the detection of a displacement of a participant, that is, of the Bluetooth tag he wore, from one space to another one, we believe that these technical occurrences might have decreased the overall reliability of the Bluetooth measure, leading therefore to a decrease in the reliability of the variables which were derived from that system.

The obtained ICC values should be compared to ICC values found in the literature. When considering indices obtained by a questionnaire, a published study reported ICC ranging from 0.53 to 0.68 for total sedentary time, while for more focused domain-specific sedentary times, ICC ranged from 0.36 to 0.76 [24]. Another study concerning a self-administered questionnaire quantifying physical activity indices reported ICC values ranging from 0.56 to 0.91 for physical activity indices [25]. A questionnaire administered to caregivers of persons with dementia scored an ICC ranging from 0.87 to 0.95 [26]. When considering sensor-based assessments, a study on patients with knee osteoarthritis reported ICC ranging from 0.70 to 0.93 for physical activity indices [29]. In a gait monitoring approach in older adult participants, it was observed that higher repeatability was observed among the prefrail or frail group (ICC>0.78) compared to nonfrail individuals (0.39<ICC<0.55) [28]. In a study, in which the accelerometer data were processed with different algorithms, the test-retest reliability of the physical activity indices was generally very good to excellent (ICC=0.70‐0.90) [27]. All considered, the test-retest reliability found in the present study is comparable with the already published literature.

The presented experiment has some limitations. First, the location tracking method required the installation of a hardware infrastructure embedded in the building, which is not movable to another setting (this limitation obviously does not apply to the wrist accelerometer setup). As a consequence, control data obviously cannot be obtained from healthy age-matched individuals living in the same environment. This limitation might be overcome in the future, in the light of a rapidly evolving technological advancement, if detection antennas for an ultrawide band become ubiquitous indoors [58] or with AI-enabled surveillance cameras, provided that privacy concerns are solved [59]. Second, the activity tracking method had a compliance which was not 100% because some individuals did not agree to wear the wrist sensors or removed them during the experiment, resulting in a generic discomfort, though the clinical partners had preliminarily warned about these refusals as a consequence of the dementia affecting the unit residents. In the latter cases, when the wrist sensors measured for more than 24 hours, the data were retained for multiples of 24-hour durations, and outcome variables were computed on the available full days. Nonetheless, such occurrences were seldom observed. Despite the authorization given by the legal tutors, our participants, persons with dementia, were still able to refuse to participate or at least to reject the wrist or ankle sensors because they felt disturbed by their presence. The dimension and location of the sensors might play a role in their refusal, so it may be expected that future technological developments, letting smaller sensors be embedded in the garments or fixed into adhesive patches, will reduce this noncompliance. Interestingly, compliance problems arose mostly for the wrist sensor, as the participant could perceive and grasp it. Third, the control group for the accelerometric data was built from an available dataset of healthy older adults living in the community [11,30] to be comparable as to average age, but the age range was narrower (80‐87 years), 9 of 25 participants were above the maximum control age and 7 of 25 participants were below the minimum control age. This is due to the difficulty in recruiting healthy older adults in their nineties living in the community [60] while younger healthy older adults were excluded in order to have a matched average age. Fourth, the choice of monitoring activities with wrist accelerometry (adopted in 22% of the published actigraphic studies) was critically compared with the option of performing waist accelerometry (in 48% of the studies) [61], as an accelerometer on the waist is closer to the body center of mass and, therefore, considered more informative about physically engaging activities. The wrist choice was justified by the relevance of upper limb movements in older adult residents in a long-term care unit, as locomotor activities might be quite reduced, and by feasibility reasons, as an accelerometer positioned on the lower back with a belt would greatly interfere with autonomous bathroom activities of older adults, would not have had a stable position, and would be disturbing when lying in the bed. Moreover, the integration of a gyroscope sensor into the monitoring watch was not useful, and therefore dropped, as gyroscopes are mainly involved in activity classification, which was not an objective of this study.

In conclusion, the method involving activity tracking by a wearable accelerometer and location tracking through the detection of a wearable Bluetooth beacon by a network of Bluetooth antennas was proven to be feasible. The computed indices quantified the activities and social behaviors of the long-term care unit guests, providing a novel multidimensional characterization of the individual guest behavior, which considers the physical activities, the WEff or SEff, as well as the social interactions in terms of agreement between the activity profiles among participants. Though the results about activity levels in persons with dementia are in line with intuitive knowledge (ie, people with dementia move less than age-matched healthy participants), the assessment provided novel evidence against intuitive knowledge when quantifying the strict and continuous social interactions involved in a residential long-term unit. Such interactions, quantified by the CC index, resulted in lower values than those observed among healthy older adult couples not even living together. Apart from behavioral evidence, the fundamental novel information provided by this study consists of the quantification of the test-retest reliability of the computed indices, which resulted in good to excellent values, thus possibly supporting future power analyses in designing longitudinal studies on the activity tracking and localization of persons affected by health problems or frailty, and particularly by dementia or cognitive decline. These test-retest features complement the ones already available in the literature concerning studies on healthy adults [27], on walking bouts in frail older adults [28], and in people with orthopedic problems [29]. The results presented here will possibly support future studies able to characterize the effects of aging and changing social situations with a longitudinal design, or the effect of a possible therapy in a randomized controlled trial design. Interestingly, the methodology proved to be able to clearly identify behaviors, such as, for example, the wandering behavior, expressed by a minority of individuals inside this study’s group of persons with dementia, thus supporting a possible use of the methodology in clinical settings, either as a way to characterize the individual or a means to detect changes in their behavior. Moreover, the computed indices might support a clinical team in improving the guests’ experience [62] and in tailoring a personalized and coordinated across-subject rehabilitation program.