A prospective observational study was conducted to meet the study objectives. Participants were asked to wear a wearable device for a minimum of 8 days. At the end of the study, participants were assessed for frailty, activities of daily living, and level of comorbidity.

Home care clients in the Greater Toronto Area were recruited through VHA Home Healthcare from August 2018 to September 2019. VHA Home Healthcare is a home care agency that serves over 3000 clients throughout the Greater Toronto Area and other metropolitan areas in Ontario, Canada. Patients 55 years or older who had been receiving personal support service for more than 3 months were eligible for the study. Patients who were diagnosed with primary neuromuscular pathology, dependent on wheelchair, in an end-of-life program, or had cognitive impairments that could interfere with the use of wearable devices were excluded. Eligible home care patients were identified using VHA’s electronic medical record system.

The Xiaomi Mi Band Pulse 1S (Mi Band, hereafter) is a commercially available wearable device that is worn on the wrist. It uses a triaxial accelerometer to capture motions to approximate step count and sleep events. It is equipped with an optical heart rate sensor (photoplethysmography) to measure minute-by-minute heart rate. While the Mi Band can be worn on either the wrist or neck (as a pendant), its placement was limited to the wrist for the study. The reliability and internal consistency of Mi Band’s performance for measuring step count when walking and jogging has been validated [32,33]. Wrist-worn wearable devices displayed systematically lower heart rate during exercise, but the Mi Band demonstrated the highest accuracy [32].

We collected daily step count, light sleep time, deep sleep time, total sleep time, awake time, sleep quality, mean heart rate, and heart rate standard deviation. Sleep quality was calculated as the percentage of sleep duration over total sleep time; sleep duration was determined by subtracting awake time from total sleep time [34,35]. Heart rate was measured in beats per minute. A pool of 10 devices were used in rotation and sanitized throughout the study. The adherence to wearing the device was defined as 10 hours or more of wear time per day [36].

Frailty was assessed using the Fried Frailty Index, a tool that has been developed for and used widely with community-dwelling older adults [1]. The Fried Frailty Index assesses phenotypic frailty based on 5 criteria: weight loss, exhaustion, slowness, weakness, and low physical activity. The index categorizes frailty into 3 stages based on the number of criteria that are met: nonfrail, prefrail, and frail corresponding to scores of 0, 1-2, and 3-5, respectively [1]. We dichotomized the Fried Frailty Index into a frail group for those with a score of 3 or higher and a nonfrail group for those with a score of 2 or lower [1].

Sociodemographic variables were collected using a short background questionnaire and through review of the patient’s medical chart. These sociodemographic variables included age, sex, weight, height, ethnicity, level of education, income, and marital status. The level of comorbidity was assessed using the Charlson Comorbidity Index (CCI) [37]. The level of activities of daily living was assessed with the Katz index of independence [38]. The number of hours of service received per week was collected by reviewing the patient’s medical chart.

Descriptive statistics and univariate comparisons of means, medians, and proportions were performed to describe the sociodemographic information and patient assessments according to their frailty status. The level of education was condensed into 2 levels: high school (some or completed) and postsecondary. Household income was categorized into a lower income, those who earned $30,000 (approximately US $22,653) per year or less, and higher income, those who earned $30,000 or higher per year. Ethnicity was categorized into 2 levels: Caucasian and others which included aboriginal identity, Latin American, African American, South Asian, Southeast Asian, East Asian, Filipino, Arab, and West Asian.

Wearable device data were examined for participants adherence level, and days with less than 10 hours of wear time were excluded. Heart rate measurements of zero were generated when the device failed to have good skin contact. Such measurements were treated as missing and were removed from the analyses.

The Shapiro-Wilk test was performed to check for normality. To check for significant differences between patients who were frail and patients who were nonfrail, when the assumption of normal distribution was met, a two-tailed independent t test was used, while the Mann-Whitney U test was performed otherwise. The chi-square test was performed for categorical variables. The posthoc chi-square test was performed when significance was observed.

Pearson and Spearman correlation statistics were used to examine the relationship between frailty, sociodemographic information, patient assessments, and the data collected from the wearable devices.

Multiple logistic regression models were generated to predict frailty status. A sequential stepwise feature selection method was used to select the variables to be fitted into the models. The feature selection was used on the pool of sociodemographic and patient assessment variables to determine the features to be included in model 1. Model 2 was built by applying feature selection to the variables derived from the wearable device data. Model 3 used all available variables in a feature selection algorithm; the selected variables were used to build the logistic regression model. The Hosmer-Lemeshow test was performed to test the goodness-of-fit for each model. The predictive performance of each model was evaluated and compared using the area under the receiver operating characteristics curve (AUROC).

Statistical significance was set at α=.05 for all statistical results. The significance level for posthoc tests was corrected using the Bonferroni method. All statistical analyses were performed using R (version 3.6.0) in R studio (version 1.2.1335; R Studio Inc). Stepwise feature selection was performed using the function (stepAIC, version 7.3-51.4) from the MASS library [39].

This study received ethics approval from the Office of Research Ethics Board at the University of Waterloo (ORE22842).

A total of 72 older adults responded to the mailed recruitment brochure. All 72 older adults were contacted, and 45 agreed to participate in the study; 4 participants withdrew before completion of the 8-day study period. Data attrition due to technical issues resulted in data from 4 participants not being included. In total, 37 older home care clients were included in the study.

Participants were 57 to 96 years of age, with a mean age of 82.23 (SD 10.84) years and 76% (28/37) were female (Table 1). The prevalence of frailty among the study population was 35% (13/37). On average, participants were observed for 9.43 (SD 1.99) days. Participants who were frail (mean age: 83.92 years) were significantly older (P<.001) than those who were nonfrail (mean age: 80.61 years). There was a significant difference in the income level between older adults who were frail and those who were nonfrail (P=.03). Posthoc comparisons within each of the 3 income levels showed no statistical significance (low income: P=.93; mid to high income: P>.999) after correcting the α level with the Bonferroni method. Frail patients received significantly greater hours of home care services per week compared to the hours received by patients who were nonfrail (P=.01). The resulting P values of the Shapiro-Wilk normality tests are presented in Multimedia Appendix 1. The results of group difference tests are presented in Multimedia Appendix 2.

On average, older adults wore the device for 20.03 (1.64) hours per day (Table 2). Home care clients who were frail reported significantly lower daily step counts than their nonfrail counterparts did (mean steps per day: 367.11 vs. 1023.95, respectively; P=.04). Total sleep time (P=.01) and deep sleep time (P<.01) were significantly longer for older adults who were frail, but no difference was found for light sleep time (P=.04). No difference was found for heart rate measures. Box plots corresponding to Table 2 are presented in Multimedia Appendix 3.

The correlation between wearable data and frailty is summarized in Table 3. Daily step count was negatively correlated with frailty level (r=–0.52, P<.001). All 5 sleep measures were moderately correlated with frailty. Education level was moderately correlated with frailty status (r=–0.40, P=.02). No relationship was found between heart rate measures and frailty status.

The growing aging population in Canada and the emphasis on aging-in-place call for innovative ways to improve efficiency in the home and community health care sector. There is an increasing interest in integrating information and communication technology such as consumer-grade wearable devices into health care delivery due to their rising popularity, ease-of-use, and the potential usefulness of continuously collected data [40]. The aim of this study was to investigate the possibility of assessing and predicting frailty using a wearable device.

We observed 37 older home care clients for a minimum of 8 days. The prevalence of frailty in the study sample, 35% (13/37), was similar to that found in other research studies examining home care clients [3,41]. Many research studies [9,42] reported a significantly higher prevalence of frailty in older women compared to prevalence in older men, but this was not observed in our study sample. However, another study [3] that examined the same population did not find any significant difference between the sexes. Overall, the study sample seemed reasonably representative of the home care population. Previous research studies [1,43] reported an association between income and education level and frailty. Our study sample had significantly different income levels between the 2 frailty groups (P=.03). However, the posthoc chi-square analysis results did not reach statistical significance (low income: P=.93; mid to high income: P>.999). Education level was moderately correlated with frailty level (r=–0.40, P=.02). Overall, our study sample displayed the general characteristics of frail populations [1,43].

Our study found a significantly higher utilization of home care service by older adults who were frail compared to utilization by older adults who were nonfrail (mean hours per week: 5.15 vs 2.77; P=.01). Unfortunately, the current system fails to meet all care needs of home care clients as indicated by the increased hours of informal care and caregiver distress for the home care clients with more severe frailty [13]. Resulting adverse health outcomes and increased health care utilization [3] highlight the need for a better allocation of home care service to those who stand to benefit the most.

In our study sample, older adults who were nonfrail walked significantly more than the older adults who were frail. This result is in line with the findings of previous research studies where reduced daily step count and physical activity were observed for frail community-dwelling older adults [44] and ICU patients [45]. In a previous study [46], daily step count was significantly related to frailty. Our study extended this evidence outside the controlled settings and beyond 24-hour monitoring period [47,48], and demonstrated the relationship in an unsupervised setting.

Sleep measures including longer total sleep, deep sleep, and light sleep durations; awake time; and sleep quality were shown to be related to more severe frailty. This is contrary to the common knowledge of deterioration of sleep quality and quantity with aging [49]. However, in epidemiological studies, a longer sleep duration was associated with an increased risk of heart disease and all-cause mortality [50]. The lowest mortality risk was found for those who sleep about 7 hours a night [51], while men who slept more than 8 hours per day had a tripled risk of heart disease [52]. This relationship was shown in our study sample where older adults who were nonfrail and older adults who were frail had significantly different total sleep durations (P=.01). Older adults who were nonfrail had a mean total sleep duration of 6.48 hours (close to 7 hours), while their frail counterparts slept for 8.16 hours. These findings demonstrate the additional information wearable devices provide over conventional sleep quality assessments.

In this study, we built logistic regression models using a sequential stepwise feature selection method. Feature selection in general can help improve predictive performance [53]. It minimizes the number of features needed in a model, which was critical given the small sample size of this study. While manual feature selection based on expert knowledge could have been a feasible alternative, our goal was to maximize frailty prediction performance in our data set by utilizing an empirical feature selection method. The analysis of multiple logistic regression models showed that wearable device data were a superior source of information for predicting frailty than sociodemographic information and patient assessments. However, the highest AUROC of 0.90 was achieved with the model that used wearable device data, sociodemographic, and patient assessment information. Previously, a similar study [16] that used a neck-worn wearable device to obtain step count and physical activity-related variables achieved an AUROC of 0.88 in discriminating the prefrail group from the frail and nonfrail groups. Another study [48] used 2 research-grade wearable devices concurrently and achieved an AUROC of 0.86 in discriminating 3 frailty states using stride length. Both studies were limited due to their short 48-hour observational period and being conducted in a laboratory setting. Our study demonstrated that unsupervised monitoring of frailty at home using a wearable device is possible. Our results corroborate that wearable technology should complement, rather than replace, the existing practice [54].

Many mobile health and telehealth apps have been successful at delivering health care while improving efficiency [55]. A study [56] that examined telehealth for frail older adults found the most cost-effective telehealth program used automated monitoring of vital signs to reduce health service use and facilitate remote follow-up. Wearable devices are becoming increasingly affordable and are capable of offering a similar use case as telehealth apps with their automated monitoring of physical activities, sleep, and heart rate. The range of information collected from wearable devices are also increasing with the advancement of sensor technology such as electrocardiogram, blood glucose level, oxygen saturation level, and electrodermal activity. When coupled with well-calibrated algorithms that enable early detection of health deteriorations such as frailty, cost savings can be further increased. The added value of wearable devices in assessing frailty for home care clients and community-dwelling older adults should be carefully evaluated for their feasibility in real-life settings. Each home or community health care system is unique, including but not limited to their target population, geographical area, and funding structure. Future research should consider these factors when evaluating the clinical value and cost savings of wearable devices.

Future research should confirm the predictive power of data derived from wearable devices and extend it beyond the home and community care sector. Our results indicated that wearable devices are a valid tool when an adequate analytical process is used. We recommend that future home care research studies leverage the potential of consumer-grade wearable devices to help identify vulnerable and frail groups who may benefit from additional home care services and increased access to health care.

Our study has several limitations. First, the small study sample prevented us from stratifying patients into nonfrail, prefrail, and frail groups. A third frailty state could have helped us demonstrate gradient measures of wearable data. The small sample size also limited the number of variables that could be used in developing multiple logistic regression models. The 3 logistic regression models were each fitted with 2 to 4 features. They exceeded the common rule of 1-in-10 and which may have increased the risk of overfitting [57]. The small sample size precluded partitioning our data into training and test sets. As a result, the reported predictive performance overestimated the performance that would be found on a different sample of older adults. A further caution should be taken when interpreting the results of the Hosmer-Lemeshow test due to the small sample size.

Our research used an 8-day observation period. While this was longer than the observation periods of most other studies using wearable devices, an even longer observational period may be required to reveal new patterns that are not observable within 8 days such as weekdays versus weekends and seasonal differences. Lastly, the validation studies that examined the Mi Band [32,33] were conducted in younger participants, limiting their generalizability to older adults of this study.

In this study, we proved the concept of using a wrist-worn consumer-grade wearable device to assess frailty among older home care clients. Data collected from the wearable device, such as total sleep time and deep sleep time, were associated with frailty. The frailty prediction model based on variables selected from wearable devices, sociodemographic variable, and patient assessment variables achieved the highest AUROC of 0.90, compared to the AUROC of the other predictive models that either used only sociodemographic and assessment variables or only wearable device–derived variables.