From June 2016 to November 2017, patients were recruited after admission to a subacute rehabilitation center in Los Angeles. A longitudinal study of the physical therapy, occupational therapy, and sensor-based data assessments was performed. The study cohort contains patients admitted to a skilled nursing facility for an intended rehabilitation course of no more than 21 days. After this period, patients were either re-admitted to hospital or stayed in the community or in their residence in long-term care.

Participants were eligible if older than 60 years of age, English speaking, and able to sign a consent form approved by University of California, Los Angeles, Institutional Review Board (IRB# 16-000166 entitled Sensing in At-Risk Populations). Exclusion criteria were movement disorders or complete paralysis of the upper or lower extremities. The diversity of cohort comprised patients who were postsurgical and poststroke and had functional limitations because of medical illnesses.

Patients were given a smartwatch every morning at 9 am, and the watches were collected from them at around 6 PM daily. Sensors placed throughout the facility collected data passively without any interaction required from patients. Patients normally stayed in the resident room (bedroom) and were scheduled for an hour of daily exercise and activity in the therapy area of the nursing home.

The core of SARP is comprised of the following: hardware—(1) commercially available Sony SmartWatch 3 with built-in EM7180 ± 2 g triaxial accelerometer, 420mA battery, and BCM43340 Bluetooth module; (2) proximity beacons (MCU ARM Cortex-M4 32-bit processor) mounted at locations of interest within resident rooms (bedrooms) and therapy area, shown with red color dots in Figure 1; clinically validated software—activity recognition, indoor localization, and data visualization algorithms, all encompassed within a Health Insurance Portability and Accountability Act–compliant infrastructure.

Details of the system architecture can be found in [16-20], and the patent is described in [21]. Activity tracking and indoor localization models were built, validated, and refined prior to this study on a separate cohort of patients [17].

Analysis inclusion criteria were defined to ensure all patients satisfy a minimum amount of daily sensor data and collected PT and OT assessments. Analysis criteria include patients with the following data: (1) ≥3 days of watch data; (2) each day ≥4 hours of watch wear time; and (3) ≥3 sessions of PT or OT or a combination of both PT and OT.

Cohort data were agglomerated for analyses according to the consort diagram shown in Figure 2.

Visualization of prior analysis was generated to unveil any longitudinal patterns. The time trends of sensor-based features appeared to be approximately linear; hence, we decided to use linear models for longitudinal analysis.

Descriptive statistics (medians and IQR) were computed for clinical assessments (ie, PT and OT) at each session. Generalized linear mixed effect model was used to understand the longitudinal relationships between the clinical measures and the sensor-based features [23-26]. Due to the frequency difference in which sensor and clinical assessments were collected, we merged a day of clinical assessment data with its corresponding day or closest day containing the sensor data (SD 3 days). Note that a valid day of sensor data should satisfy the analysis inclusion criteria 1 and 2.

Three models, each with different sets of sensor-based features, were constructed for each clinical outcome. Model 1 included overall energy intensity as covariate. Model 2 considered energy intensity at resident room and energy intensity at therapy area as covariates. Additionally, sensor-based activity parameters (eg, energy intensity of sitting) were used in model 3. Linear time indicates the number of weeks since the enrollment day. Interaction effects of sensor features with time were also included.

The Ethics Board reviewed this study. The following was their determination: “The UCLA Institutional Review Board (UCLA IRB) has approved IRB#16-000166 entitled ‘Sensing At Risk Populations (SARP).’ UCLA's Federal wide Assurance (FWA) with Department of Health and Human Services is FWA00004642. The UCLA IRB waived the requirement for HIPAA Research Authorization to identify potential research participants. The UCLA IRB waived the requirement for informed consent for the review of medical records to identify potential research participants under 45 CFR 46.116(d). The UCLA IRB waived the requirement for signed informed consent for participants admitted to the BECH for acute care under 45 CFR 46.117(2).”

From 184 consented patients, 110 (60%) met the watch wearing time protocol with mean age of 79.4 (SD 5.9) years. Moreover, 97 (88%) patients were included in PT-watch paired analysis and 60 (54%) in OT with watch analytics. Most participants were female (n=79, 72%) and of White race or ethnicity (n=84, 76%). Additionally, 62% (n=69) of the patients had pain, 99% (n=109) of them needed some level of assistance with functional mobility activities (transfer activity), and 75% (n=83) needed assistive devices for walking. Table 2 presents detailed sociodemographic and clinical characteristics of the 110 patients. ADL parameters and their significance in determining the outcome are presented based on initial assessments, at the time of admission, or within one day.

The community group spent higher overall energy intensity and energy intensity at the resident room compared to the hospital group, as seen in Figures S1 (a) and S1 (b) of Multimedia Appendix 1. However, energy intensity during therapy sessions tends to have similar values between two groups, especially toward the end of the rehabilitation period, as seen in Figure S1 (c) of Multimedia Appendix 1.

The descriptive statistics of clinical parameters are summarized in Table 3. It shows that “gait distance feet” increases over time (median and IQR after the first week), and “activity tolerance” increases (IQR after first week and median after second week). The table indicates no clear improvements in other clinical-based measures gauged by PT and OT functional levels within 3 weeks.

The associations of repeated PT, OT, and sensor-based measurements are modeled through three generalized linear mixed models. On PT and sensor associations, according to Table 4, the results of model 1 revealed that gait distance feet (β=.28; SE=0.06; P<.001), gait level surface β=.17; SE=0.04; P<.001, and bed mobility including supine to sit (β=.26; SE=0.05; P<.001) improved over time. Higher overall energy intensity indicates a higher score of transfer activity (β=.22; SE=0.08; P=.03).

In model 2, energy intensity at the therapy room was positively associated with transfer activity (β=.19; SE=0.08; P=.02). In addition, gait distance feet (β=.28; SE=0.05; P<.001), gait level surface (β=.17; SE=0.04; P<.001) and bed mobility including supine to sit (β=.26; SE=0.05; P<.001) improved every week.

In model 3, sitting energy intensity showed positive association with transfer activity (β=.16; SE=0.07; P=.02). Meanwhile, according to model 3, participants showed weekly improvements in gait distance (measured in feet; β=.27; SE=0.06; P<.001), gait level surface (β=.16; SE=0.05; P<.001), and bed mobility including supine to sit (β=.26; SE=0.05; P<.001).

On OT and sensor associations, Table 4 shows that lower body dressing, toileting activity, and activity tolerance in general improved every week in all three models. The higher value of overall energy intensity in model 1 implied a higher functional score of lower body dressing (β=.19; SE=0.09; P=.03) and toileting activity (β=.23; SE=0.09; P=.01).

The occurrence of a location is equal to the number of times a patient spends more than 40 continuous seconds within that specific location. In other words, if the smartwatch receives Bluetooth low energy signal of a beacon corresponding a location for 40 seconds, the occurrence of that location increases by one unit. Figure 3 (a and b) shows total occurrences of patients in various nursing facility locations (daily) normalized by the number of patients in each category. Darker colors indicate higher frequency of patients visiting a particular location. In short, patients in outcome category “home” traveled within the facility (resident and therapy area) much more frequently than patients eventually admitted to a longer-term care or the “hospital” group. Additionally, no patient in the hospital category used upper body exercise (SciFit), Endorphin, and stair equipment in the therapy area.

To the best of our knowledge, this paper and what we described in [13] are first to explore a combination of indoor localization and physical activity tracking to assess older residents. Following baseline investigations [13], in this paper, we highlight significant findings in longitudinal analyses of clinical and sensor-based features.

One of the principal findings of this study is that the energy intensity spent in therapy sessions, unlike in resident room, tend to have similar values in both outcome groups, more significantly toward the end of the rehabilitation period (Figure S1 in Multimedia Appendix 1). Perhaps the therapists in both patient groups are encouraged to complete their therapy activities and are part of an individually designed therapeutic program that aimed to improve functional activity. Moreover, energy intensity spent in the resident room is very similar to overall energy intensity in that patients generally spend most of their time in the resident room. Resident room activity levels are likely to be crucial in determining the outcome of patients, even at early stages of their rehabilitation. Further understanding of the therapeutic skills learned during therapeutic intervention and carryover into the resident room warrants further study.

Based on Table 3, the PT and OT features investigated in this study all improved over time along with the sensor-based feature, energy intensity. However, improvements are more distinguishable between admission day and weeks 1 and 2. On week 3, the mean value for sensor-based features such as overall energy intensity declines. Similarly, OT and PT features show less change compared to week 1 and admission day. One possible reason could be the drop in sample size after week 2 as patients are likely to be discharged earlier. Note that despite the steady PT and OT functional scores in later times, the interquartile range decreases over time, which indicates less variations in functional levels. This could mean that residents achieved their functional goals or plateaued in functional progression. Other aspects that limit a resident’s functional ability need to be examined to determine if nonmotor parameters are limiting a resident’s progress. Cognition, vision, and psychological factors are some of the areas that may limit functional progression.

Table 3 also shows that except the “gait distance in feet,” the improvement of features was not evident after the 2nd and 3rd week. Further exploration of therapy treatment intensity or type of intervention is warranted. Significant improvements in “gait distance in feet” suggest the importance of this feature in clinical assessment. The rest of the gait measures showed they were less likely to change over time. Dynamic gait parameters and their relation to mobility in daily activities need more investigation.

The captured sensor-based longitudinal changes such as lying down, sitting, and overall energy intensity reflect changes in PT and OT features (Table 4). This finding confirms the benefit of remote patient monitoring systems as adjunct tools to further reveal patients’ daily story lines. Such systems can bear valuable information in further understanding the type and intensity of therapy interventions that impact overall functional outcome. Brisk features remained surprisingly unchanged over time when patients were expected to become less sedentary during recovery of functional abilities, at least partially. Average sedentary time among all patients was more than 99.8% and remained unchanged. In other words, the cohort was walking less than 0.2% of the time, measured objectively by the SARP wrist-worn sensor. This finding strongly suggests that focusing on sedentary features among elderly patients is beneficial, confirming the studies in [27-29], contrary to the emphasis many patient monitoring systems place on using activity trackers to count steps [30,31]. This study shows the importance of translating all movements into measurements such as energy, or energy intensity, rather than solely relying on steps. This may shed light on the type of intervention needed for improving the mobility of the elderly resident population.

This study had some limitations. Wrist-worn accelerometers used for activity recognition are popular due to their ease of use and ability to capture a comprehensive set of activities. However, interpreting users’ data in sedentary positions such as sitting or standing can be quite challenging. Movements (or lack thereof) in sedentary positions are hard to be distinguished by wrist-worn sensors [32]. Compliance to technology is another obstacle faced in this study. Patients accepting to use the technology is a challenge expected to be generic and present in similar studies.

Battery consumption of smart watches can be problematic when trying to transmit data, hourly or daily. Battery lifetimes are normally insufficient in almost all smartwatch manufacturing brands. Their operating systems are designed to perform sophisticated tasks, many of which are not needed for patient remote monitoring such as receiving messages and calls. Furthermore, consumer-grade wearables have wide variability in their accuracy across a range of functional activities depending on their placement, the individuals’ movement characteristics, speed of walking, using assistive devices, and so on. The best way to tackle this problem is to use wearable sensors specifically designed (hardware and software) for patient monitoring. However, commercially available research-grade sensors are very expensive and not yet clinician and patient friendly [33].

The study cohort had two outcome groups that were not equally presented. The data set predominantly comprised majority class instances and contained only a few instances of patients who were re-admitted to a long-term care. Akin to most imbalanced medical data sets, analyzing such data poses a great challenge [34].

This study aimed to show that wearable activity trackers, despite raising concerns about their efficacy in quantifying residents’ health, can result in a better understanding of patients’ well-being when tailored for a specific cohort. Such studies can hopefully pave the way in early prediction of hospitalization, developing intervention alerts and improving overall quality of care. As discussed, our remote patient monitoring system, SARP, captures a combination of indoor localization and physical activity features. SARP information on daily and longitudinal activity patterns can be incorporated into mobile health technology platforms to provide a better assessment of underrepresented, particularly frail, populations.