Objective: To determine the accuracy of heart rate, physical activity and sleep by wearable devices in hospitalized medical patients.
Design: Prospective cohort feasibility study Clinicaltrials.gov NCT03646435)
Participants: Patients were recruited from two major teaching hospitals– Toronto General Hospital and Toronto Western Hospital – of the University of Toronto that both have General Internal Medicine services. The General Internal Medicine patients of the University of Toronto hospitals are known to have a median age of 73, a median of 6 comorbidities and are admitted with a wide variety of medical diagnoses.13
Subject inclusion criteria were English-speaking adults (aged 18 years or older) admitted to a General Internal Medicine service and who were able to provide consent. Subject exclusion criteria: 1) patients in whom routine vital sign monitoring was not indicated, such as those with palliative, comfort-oriented goals of care, 2) patients with significant cognitive impairment limiting their ability to complete surveys, and 3) patients at risk of vascular compromise of the arm on which the wearable device was to be placed (e.g., dialysis fistulas, peripherally inserted central catheters). Patients isolated under contact precautions were also excluded to reduce the risk of transmitting infection. A convenience sample size of 50 participants was chosen for this pilot study similar to a recent feasibility wearable study.12 The Research Ethics Board of University Health Network approved the study (ID# 18-5621). All participants provided written informed consent.
Intervention: We used Fitbit Charge 2 and Charge 3, wrist bands that can measure activity, sleep and heart rate (cost $150–160 CAD). Participants were asked to wear the band continuously until discharge or for up to one week, whichever came first.
Data Collection: The following information was collected from each participant at enrollment: age and sex. The following information was collected after enrollment: most responsible diagnosis, comorbidities, discharge status (alive, left against medical advice or died in hospital), discharge location (home, home with support services, rehabilitation facility, or supportive housing), length of stay, duration of wearing device, and documentation of atrial fibrillation in previous documented comorbidities or by any ECG performed during the visit.
For each participant we collected the following information during the time of study participation: daily sleep questionnaire completed by patients, heart rates documented by nursing staff as part of vital signs, activity assessments recorded daily by nurses, and Fitbit recordings for heart rate, sleep and activity. At both hospitals involved in this study, vital signs are measured by nurses and are manually entered into the electronic health record.
We collected daily sleep quality from patients using the Richards-Campbell Sleep Questionnaire (RCSQ), a validated survey instrument for measuring sleep quality in hospitalized patients.14 This survey uses 0-100 mm visual analog scales to assess sleep depth, latency, awakenings, percentage of time awake, and overall quality of sleep. Item values are summated and divided by 5 providing a mean score between 0-100 reflecting the patient's perception of their sleep quality with higher values representing better sleep. Participants were also asked to report the total hours of sleep they experienced.
As part of routine practice, nurses twice daily recorded their patients’ activity levels on the following scale:
Bedfast (confined to bed)
Chairfast (ability to walk severely limited or non-existent)
Walks occasionally (walks occasionally during day, but for very short distances, and spends majority of each shift in bed or chair)
Walks frequently (walks outside the room at least twice a day and inside room at least once every two hours during waking hours)
Fitbit data was transmitted from the wearables via Bluetooth to a mobile device using the Fitbit app and then uploaded to Fitbit servers. Fitbit heart rate, sleep, and activity data were recorded minute-by-minute and downloaded from Fitbit servers using a custom, in-house Python program using the Fitbit application programming interface.15
Fitbit Data Processing: To process data from Fitbits, determining wear time can be done through rules based on activity or heart rate data.16 Since hospitalized patients are known to have low mobility, we used a heart rate-based algorithm instead of a steps-based algorithm.16 With heart rate-based algorithms, devices are assumed to be not worn if significantly less than usual heart rate recordings are recorded.16 Typically, when the device was worn consistently, there would be 50–60 heart rate recordings per hour. We assumed the device was not worn when there were less than 30 heart rate recording per hour. This was to avoid incorrectly considering the device being on the patient due to known problems of occasional spurious heart rate measurements that can be recorded even when devices are not worn.17,18
Analysis: Continuous demographic variables were analyzed using descriptive methods; categorical variables using contingency tables and histogram plots. Primary endpoints were the correlation between the following measures: 1) Nurse-reported vital signs matched using the timestamp documented in the electronic health record to the closest Fitbit heart rate data point within a one-minute tolerance, 2) Fitbit sleep and the RCSQ completed by patients for the same night, and 3) Fitbit activity and nurse-recorded patient activity over the same day. Because atrial fibrillation reduced the accuracy of Fitbit heart rates in another study,10 we conducted a predefined subgroup analysis for heart rate in patients who did or did not have atrial fibrillation. Atrial fibrillation was determined based on patient history and ECGs performed during admission. Statistical analyses were performed using Python SciPy libraries.19