The global prevalence of dementia is expected to reach 139 million by 2050 [1], and in countries such as Norway, more than two-thirds of nursing home residents live with dementia [2]. Monitoring neuropsychiatric symptoms (NPSs) and changes in physical activity typically relies on proxy-rated assessments, which are limited by subjectivity, retrospective bias, and low sensitivity to subtle or short-term change [3]. As health care systems face increasing workforce shortages, there is a growing need for scalable, objective approaches that can support continuous monitoring and individualized care [4]. Recent advances in sensing technologies, including wrist-worn wearables and contactless radar systems, offer minimally obtrusive, real-time data collection in naturalistic environments and may help address these challenges.

Physical inactivity is widespread in nursing home environments [5,6], with daily routines dominated by sedentary behaviors such as sitting, resting, and sleeping [7]. Despite the importance of physical activity for maintaining function, there are no standardized recommendations for activity dosage in people with advanced dementia [8]. Although the relationship between daytime activity and sleep quality remains insufficiently understood, several studies suggest that higher daytime activity may be associated with improved sleep [9,10]. Functional ability is typically assessed using proxy-rated activities of daily living (ADL) questionnaires, but these tools are retrospective and infrequently administered, reducing their use for detecting early or subtle decline, which is an important limitation given the strong association between functional loss and increased morbidity and mortality [11].

Accelerometry provides several methods for quantifying physical activity, including step counts, intensity, frequency, and vector magnitude [12], and common metrics derived from raw acceleration data include Euclidean norm minus one (ENMO), mean amplitude deviation (MAD), Monitor-Independent Movement Summary (MIMS), activity index, and Rate of Change Acceleration Movement (ROCAM) [12,13]. A recent feasibility study in nursing home residents with dementia demonstrated that commercial wearables could accurately classify physical activity over a 24-hour period [14], supporting the promising potential of sensing technologies in dementia care [15,16]. However, research validating activity classification and intensity thresholds in older adults with dementia remains limited, as demonstrated in a systematic review by Gorman et al [13] exploring physical activity and sedentary behaviors in older adults, finding that there are no standardized accelerometry-based thresholds for moderate-to-vigorous or sedentary activity, with most cut-offs validated in healthy adults rather than cognitively impaired populations. Research using wrist-worn accelerometers and actigraphy in a cohort of 415 nursing home residents in Australia demonstrated that these devices could capture patterns of inactivity and sleep duration in this population. The study also highlighted practical challenges, particularly around device accuracy and participant adherence, which was particularly low (<50%), concluding that additional work is needed to refine sensor placement, device characteristics, and methodological approaches for use in a nursing home setting [7].

Sleep disturbances are prevalent in nursing home residents and affect up to 90% of those with dementia [17], with common issues including insomnia, sundowning, restless legs, and fragmented sleep. Sleep quality is particularly compromised in people with dementia residing in nursing homes [18,19], and although nonpharmacological interventions are recommended [20], psychotropic medications remain widely used and may worsen sleep or increase the risk of injury [21], contributing to further NPSs such as agitation, anxiety, and nighttime restlessness [18]. Key sleep metrics, including efficiency, latency, wakefulness after sleep onset, fragmentation, regulation, and total sleep, decline with age and are further impaired for people with dementia [22]. Autonomic regulation of sleep also changes with age and neurodegeneration, particularly regarding parasympathetic nervous system involvement in the dampening of the sympathetic nervous system during deep sleep [23]. Sleep durations of 6 to 9 hours have been associated with improved parasympathetic function in older adults with neurodegenerative disease [24].

Polysomnography (PSG) remains the gold standard for sleep assessment but is often impractical and burdensome for people with dementia in nursing homes [25]. Consumer wearables provide sleep estimates using accelerometry and photoplethysmography, but they frequently overestimate sleep in individuals with sleep dysfunction [26] and vary in accuracy [27]. Newer commercial wearable models offer sleep staging, and although sleep sensitivity tends to be adequate, the poor detection of awake periods reduces accuracy, making these devices not yet adequate for research or clinical standards and not validated in an older adult population [28]. Radar-based systems [29] offer a contact-free alternative capable of monitoring sleep in free-living environments. These systems produce digital biomarkers comparable to PSG and show promising accuracy [30], though further research is needed to confirm their reliability and clinical use for longitudinal monitoring in real-world dementia care settings.

This study aims to explore the long-term capabilities of multimodal sensing technologies for assessing physical activity levels and sleep quality using selected digital biomarkers for nursing home residents with dementia. Objectives were for observation to be aligned with real-world conditions in which such sensing technologies would be applied within a nursing home environment, and to assess whether distinct differences in selected digital biomarkers can be observed accurately and reliably longitudinally.

This study is related to an observational pilot study: digital phenotyping for changes in activity at the end of life in people with dementia (DIPH.DEM): an observational trial based on sensing technology. DIPH.DEM is a cornerstone study for a larger observational study titled “Decoding Death and Dying in People With Dementia by Digital Thanotyping” (5-D; REK 657596, NEM 2023/166). DIPH.DEM was designed to improve the overall quality, recruitment strategies, and efficacy of the larger 5-D study. This paper is an extension of a previous article [31], which used the DIPH.DEM baseline data (description in Table 1) to investigate the association between daytime activity and sleep quality in people with dementia residing in a nursing home. The DIPH.DEM cohort and study design have previously been described in detail in 2 previous articles [31,32].

Participants (n=11) 65 years or older were recruited from 2 dementia wards in a Norwegian nursing home. Eligibility for the study included that the participants had lived in the nursing home for more than 6 weeks and had a previous diagnosis of moderate-to-severe cognitive impairment or dementia, confirmed within the medical record and/or using the Clinical Dementia Rating (CDR) scale administered at baseline. Exclusion criteria included the presence of delirium at the baseline assessment and a current life expectancy of less than six weeks as suggested by the multidisciplinary nursing home team. Data were collected from February 2024 to May 2025. The proxy-rated questionnaires and continuous digital monitoring using sensing technologies, Garmin Vivoactive 5 or Garmin Venu 3 and Vital Things Somnofy, were conducted at 3 distinct time points: baseline, 6 months, and 1 year.

Verbal and written informed consent were obtained in direct conversation with the residents who demonstrated sufficient capacity to consent. For those lacking this ability, we obtained consent from the resident’s legal guardian, usually a family member, after explaining the aims and protocol of the study. The DIPH.DEM study was approved by the Regional Committee for Medical and Health Research Ethics (REK) in Norway in October 2023 (approval number 634938).

Traditional outcome measures used in this study were the Neuropsychiatric Inventory–Nursing Home version (NPI-NH) [33], assessing symptoms of sleep disturbance (Nighttime Behaviours section K), and the Physical Self Maintenance Scale (PSMS) [34], assessing functional activity levels and adverse effects of inactivity. The NPI-NH (section K) [35] includes questions about difficulty falling asleep, frequent nightly awakenings, wandering, dressing in the middle of the night, and waking earlier than normal to begin the day. The PSMS [36] includes functional activities of ambulation, bathing, grooming, dressing, toileting, and feeding. To define the severity and level of disturbances within this study, we chose to use a global scoring system from 6 to 30 (6 questions and 5 possible answers), with lower scores indicating lower dependency and higher scores indicating higher dependency.

Additional baseline measurements reported in this study were the General Medical Health Rating (GMHR) [37] for the assessment of general health and comorbidities, 4AT (Assessment Test for delirium) [38], and the CDR [39]. The CDR questionnaire (range 0‐3) is a widely accepted global scale developed to clinically identify the presence of dementia and stage its severity (score of >0.5: cognitive impairment/likely dementia; 1: mild; 2: moderate; 3: severe dementia). Within our study, participants were included in the study with a score equal to or greater than 1 on the CDR. The GMHR (range 1‐4) assesses the number and impact of comorbidities (1=very good; 2=good; 3=moderate, and 4=bad health). The 4AT Test (range 0‐12) is a screening tool for delirium and consists of 4 questions evaluating alertness, cognitive function, and acute change, with a score of 4 or above indicating potential delirium. All traditional questionnaires used were previously validated and provided in the Norwegian language.

The sensing technologies used in the study were Somnofy (Vital Things AS, Trondheim, Norway) [40] version 0.7 with sleep algorithm 1.0 and Garmin (Garmin Ltd., Switzerland) Vivoactive5 smartwatch [26] firmware version 11.14 or Garmin Venu 3 smartwatch (X) firmware version 12.11. Digital biomarkers used in this study were acceleration-based physical activity estimates using ENMO, sleep efficiency (SE), wake after sleep onset (WASO), sleep regulatory index (SRI), total sleep time (TST), sleep fragmentation index (SFI), and no presence (time out of bed) (Table 1). Garmin Vivoactive5 was used for data collection at baseline and 6 months and upgraded to Garmin Venu3 at the 1-year time point for improved battery capacity and data storage capabilities. Both smartwatch models contained accelerometers used to capture raw acceleration signals. The Venu3 includes additional sensors, a barometric altimeter, gyroscope, and the updated Elevate V5 heart-rate sensor, while the earlier model uses the V4 sensor. Only accelerometer-derived raw data were analyzed in this study. Participants wore the device continuously for 7 days at each data collection point (baseline, 6 months, 1 year), including during sleep and bathing.

Garmin wearables, including Vivoactive and Venu models, demonstrate acceptable validity and strong interdevice reliability for step counting and movement assessment in laboratory and free-living settings [41,42]. Prior studies report accurate step detection across walking speeds and age groups, including older adults with slower gait patterns [43,44]. Garmin devices have also shown reliable estimation of physical activity using raw acceleration data [45], and ENMO-based thresholds applied to wrist-worn Garmin sensors provide high sensitivity and specificity for distinguishing sedentary and light-intensity behaviors [46].

Descriptive statistics and narrative synthesis are provided including mean, median, minimum, maximum, and SD for all relative variables. The data were tested for normality and analyzed using MATLAB R2023a and STATA SE18.5; outliers were detected using box plots. The data were dependent, nonparametric, and not normally distributed resulting in the use of the Friedman test to assess differences between the 3 data collection time points. The Friedman test is used for one-way repeated measures ANOVA by ranks. A post hoc evaluation of the specific differences between the time points in which the null hypothesis was rejected was performed using the Wilcoxon signed-rank test. Kendall W interpretation of effect is as follows: 0.1‐0.29 small effect size, 0.3‐0.5 moderate effect size, and >0.5 large effect size. A sensitivity analysis was performed as a subanalysis of the significant results by removing influential extreme outliers, which were defined as totals beyond 1.5 times the IQR. The results were considered statistically significant at a P<.05, and at a χ² (2 df) critical value of ≥5.991. Reliability as temporal stability is reported using an intraclass correlation coefficient (ICC 1,1 and ICC 3,1) defined with scores ranging from 0 to 1 and score interpretation: <0.5 as poor agreement, 0.5 to <0.75 as moderate agreement, 0.75 to <0.90 as good agreement, and 0.9 to 1.0 as excellent agreement.

This study aimed to explore the use of sensing technologies in the assessment of longitudinal changes in physical activity levels and sleep quality in people with dementia living in a nursing home. We found that nighttime physical activity levels (ENMO) and 4 distinct digital sleep biomarkers (TST, SE, WASO, SRI) showed discrete changes across the 1-year data collection period, suggesting that these selected digital biomarkers could detect modest variations and declines in activity and sleep quality over time. Temporal stability of the sleep metrics was moderate to excellent; however, further inspection of the reliability of all longitudinal ENMO measured physical activity level data was poor. This suggests that the acceleration metric was unreliable as a longitudinal assessment tool and would be best applied to a single time-point, where reliability remained moderate to good or that for real-life use, additional sensor modalities and positions are required for better quality data. Adherence and acceptability of the selected technologies remained high across the 1-year period, with no adverse events, suggesting that the chosen devices and level of unobtrusiveness were a good match for the cohort. This study’s findings are important in the illustration of the capabilities and limitations of sensing technologies and their corresponding digital biomarkers as clinical decision-making tools for people with dementia. It also demonstrates the trade-off between the level of unobtrusiveness and quality of data when choosing sensing technologies for use within more vulnerable groups.

The standardization and clarification of measurement context for digital biomarkers is needed, as emphasized in a recent meta-analysis [48] stating that accelerometry is increasingly being used to quantify physical activity, yet daily activity metrics produced using acceleration data remain heterogeneous and primarily reflect overall movement rather than specific health-related exertion. In our study, ENMO-derived activity levels distinguished long-term group differences; however, the longitudinal reliability of these measures was poor. This is consistent with previous work [49,50] stating that the most accurate results and models occur with thigh or low back placement of sensors for the monitoring of physical activity in older adults. Therefore, device placement and user burden are central considerations; in this study’s nursing home population, a wrist-worn device was selected to minimize obtrusiveness and encourage adherence for older, potentially frail residents. While wrist-based wearables may be appropriate for single time point assessments, long-term monitoring may require alternative placements, such as the hip or low back, or integration with additional sensor modalities, to improve long-term reliability.

The World Health Organization (WHO) published the Global Action Plan on Physical Activity (GAPPA) [51,52], providing policy guidance for increased activity levels, reduced sedentary behavior, and tailored interventions across age groups. The report highlights limitations of self- and proxy-reported measures and advocates integrating wearable technologies into global physical activity surveillance. It also calls for standardized protocols to examine dose-response relationships between activity, sedentary behavior, and health outcomes, and for future research that enables data triangulation, refined estimates of total activity volume, and the identification of age- and disease-specific associations [53]. This is a promising step toward future global clinical use and recommendations for standardized metrics of accelerometry-based data; however, focused research on more vulnerable groups, such as people with dementia, is still lacking.

The landscape of sleep-monitoring technologies for individuals with dementia has evolved rapidly, integrating passive and wearable systems tailored for long-term care and home use. Traditional actigraphy remains widely used to assess sleep-wake patterns and circadian disruptions in people with dementia [48]; recently, more advanced noncontact sensors, such as under-mattress sleep mats leveraging ballistocardiograph and respiration monitoring, and radar technology, are being deployed to track heart rate, breathing, movement, and fragmentation of sleep, reducing the need for wearable devices [27,30]. Due to the rapid development of new sensing technologies that rival gold standard PSG, a recent study emphasized the importance of the evaluation of novel sleep technologies conducted within longitudinal, real-world settings including people with advanced dementia, as a pathway for future device validation [30]. These recommendations reflect a shift toward more nuanced, long-term, and minimally burdensome monitoring modalities that are critical for capturing sleep changes in nursing-home and real-world settings, while addressing the unique compliance challenges of dementia care.

The fundamental differences between traditional proxy-rated questionnaires and digital biomarkers derived from sensing technologies may capture distinct phenomena and often yield divergent conclusions [54]. The NPI-NH and PSMS are established proxy-rated instruments commonly used in nursing home research to assess behavioral symptoms, functional ability, and sleep-related disturbances. However, these questionnaires do not capture the same dimensions of sleep quality represented by the digital measures used in this study. Differences in data capture frequency also limit comparability and validation between digital and traditional tools, as traditional questionnaires are typically completed every 30 to 90 days, while sensing technologies generate continuous, real-time data. Because proxy-rated tools and sensor-based measures assess different constructs and operate across different temporal scales, they are not directly interchangeable. Future research should incorporate assessment instruments that more closely correspond to digital sleep and activity metrics to strengthen construct validity and improve clinical interpretation. We cannot, therefore, conclude that sensing technologies provide superior assessments; rather, they can offer a unique definition of physical activity levels and sleep patterns through digital biomarkers, which can strengthen and fill in the blanks between periodic traditional proxy-rated assessments.

Due to rapid development and consequential obsolescence, the future validation of sensing technologies should involve repeated measures within the same modality to accelerate validation and contextualize digital biomarker results within target populations, such as people with dementia. Global consensus on the use of a standardized metric for accelerometry data and clinical translation of the corresponding thresholds should be prioritized to improve the generalization and applicability of results in real-world settings. Establishing population and context-specific thresholds, which account for age, health status, cultural, and environmental factors, will be essential to improving generalizability, applicability, and interpretability in future studies. Until these challenges are appropriately addressed, integrating proxy-rated assessments, inclusion of human-in-the-loop interpretation for data, and cautious interpretation of digital biomarkers is most appropriate for vulnerable populations such as people with dementia residing in nursing homes.

Although this study demonstrates the use of sensing technologies and digital biomarkers, it is not without limitations. The sample size, though appropriate for a pilot study design, is small, and the results should therefore be interpreted with caution. A sensitivity analysis was conducted to account for the small sample size, the influence of extreme values, and to clarify which effects were most robust. After removing outliers and aggregating repeated measurements, the overall pattern of change remained consistent with the original results, but only WASO and TST retained statistical significance, indicating that these domains showed the strongest and most reliable longitudinal shifts. Clinically, this suggests that while participants exhibited meaningful variability across physical activity metrics and multiple sleep parameters, the most stable and reproducible changes over time occurred in wakefulness and sleep duration, whereas other measures may require larger samples to detect effects with greater confidence. Regardless, further investigation is warranted based on these preliminary results. Further, this study uses the universal metric ENMO to interpret the acceleration data into physical activity levels; however, there is currently no standardized metric established for clinical translation of physical activity digital biomarkers, and another method could produce varied results. This study’s objective was not to validate specific sensing technologies, and we would like to note that the nature of rapidly developing technologies creates a challenge for the timely validation of these novel devices and a gap between the use of these devices and digital biomarkers in research versus real-world applications.

Sensing technologies are promising for supporting more objective, data-driven care models in nursing homes. Acceleration-based metrics from wrist-worn devices appear most robust for cross-sectional assessment, whereas contactless radar-derived sleep biomarkers demonstrate stronger suitability and accuracy for both snapshot and longitudinal monitoring. Our findings underscore the need for cautious, context-aware interpretation of digital biomarkers. Further work is required to establish the accuracy and reliability of these sensing methods before they are incorporated into clinical decision-making frameworks.