VALIDATION OF WEARABLE BIOFEEDBACK TECHNOLOGY FOR HEART RATE TRACKING VIA REFLECTIVE PHOTOPLETHYMOGRAPHY. A Thesis. Presented to the.

Transcription

1 VALIDATION OF WEARABLE BIOFEEDBACK TECHNOLOGY FOR HEART RATE TRACKING VIA REFLECTIVE PHOTOPLETHYMOGRAPHY A Thesis Presented to the Faculty of California State Polytechnic University, Pomona In Partial Fulfillment of the Requirements for the Degree Master of Science In Kinesiology By Kiana Liane Lewis 2016

2 SIGNATURE PAGE THESIS: VALIDATION OF WEARABLE BIOFEEDBACK TECHNOLOGY FOR HEART RATE TRACKING VIA REFLECTIVE PHOTOPLETHYSMOGRAPHY AUTHOR: Kiana Liane Lewis DATE SUBMITTED: Fall 2016 Kinesiology and Health Promotion Department Dr. Edward Jo Thesis Committee Chair Kinesiology and Health Promotion: Exercise Science Dr. Michael Liang Committee Member Kinesiology and Health Promotion: Exercise Science Dr. Bonny Burns-Whitmore Committee Member Human Nutrition & Food Science ii

3 ABSTRACT The rising awareness for personal health and fitness through self-monitoring is largely enabled by recent advancements in wearable biosensor technology. Commercially available physical activity trackers incorporate sophisticated algorithms and multi-sensor technology capable of providing users with real-time biometrics including heart rate (HR) and energy expenditure (EE). The Basis Peak (BPk) and Fitbit Charge HR (FB) are two such activity trackers currently available for consumer purchase. With growing interest and use of these trackers for personal health and fitness, the need to validate their accuracy becomes especially important. The purpose of this study was to test the validity of HR measurements by BPk and FB in comparison to criterion measures acquired by electrocardiography (ECG). Twenty-four healthy subjects (12 males and 12 females) (age= years) were recruited for this study and underwent the same 77-minute testing protocol during a single visit. Each participant completed an initial rest period (lying down) of 15 minutes followed by 5 minute periods of each of the following activities: low cycling (60W resistance), intense cycling (120W), walk ( mph), jog ( mph), run ( mph), arm raises with self-selected resistance (12 reps), lunges with self-selected resistance (12 reps), and plank (60-second hold). A five-min rest period was administered between each exercise task. Each subject wore a BPk on one wrist and a FB on the opposite wrist. Wrist assignment (i.e. dominant and non-dominant) was alternated between subjects. Care was taken to follow proper wear guidelines as suggested for each device. Criterion measurements of HR were measured by 12-lead ECG. Time synced data from each device was concurrently and continuously acquired second-by-second throughout the entire 77-minute protocol for each subject. When iii

4 examining data in aggregate, there was a strong correlation between BPk and ECG for HR (r= 0.92, p<0.001) with a mean bias of -2.5 bpm (95% LoA 19.3, -24.4). The FB demonstrated a moderately strong correlation with ECG for HR (r= 0.83, p<0.001) with an average mean bias of -8.8 bpm (95% LoA 24.2, -41.8). During higher exercise intensities, the BPk demonstrated an r=0.77 and mean bias= -4.9 bpm (95% LoA 21.3, ) while the FB demonstrated an r=0.58 and mean bias= bpm (95% LoA 28.6, ). The BPk satisfied validity criteria for HR monitors, however showed a marginal decline in accuracy with increasing physical effort. The FB failed to satisfy validity criteria and demonstrated a substantial decrease in accuracy during higher exercise intensities. iv

5 TABLE OF CONTENTS Signature Page... ii Abstract... iii List of Tables... vii List of Figures... viii Chapter One: Introduction... 1 Statement of the Problem... 3 Purpose Statement... 3 Specific Aim... 3 Hypothesis... 3 Significance of the Study... 4 Limitations... 4 Delimitations... 4 Operational Definitions... 4 Chapter Two: Literature Review... 6 Multi-Sensor Activity Trackers... 6 Laboratory to Practical Methods for Heart Rate Monitoring... 7 Future Research and Conclusion Chapter Three: Methodology Experimental Design Participants Experimental Procedures v

6 Screening Anthropometry Criterion Measure of Heart Rate Activity Tracker Procedures Statistical Analysis Chapter Four: Results Aggregate Heart Rate Data Basis Peak vs. ECG Fitbit Charge HR vs. ECG Heart Rate Data above Mean ECG Heart Rate (>116 bpm) Basis Peak vs. ECG Fitbit Charge HR vs. ECG Heart Rate Data below Mean ECG Heart Rate (<117 bpm) Basis Peak vs. ECG Fitbit Charge HR vs. ECG Task Specific Comparison Chapter Five: Discussion References vi

7 LIST OF TABLES Table 1 Results for Basis Peak. *p< Table 2 Results for Fitbit (FB) Charge HR. *p< Table 3 Activity-specific results for Basis Peak and Fitbit Charge HR. Standard error (SE). *p< vii

8 LIST OF FIGURES Figure 1 Figure 2 Figure 3 Figure 4 Reflective Photoplethysmography for Heart Rate Detection. Adapted from: 12 Results for aggregated data set. Correlation between each test device and ECG and Bland-Altman Plots indicating mean bias scores and 95% limits of agreement (LoA) Results for data set above mean ECG heart rate (>116 bpm). Correlation between each test device and ECG and Bland-Altman Plots indicating mean bias scores and 95% limits of agreement (LoA) Results for data set below mean ECG heart rate (<117 bpm). Correlation between each test device and ECG and Bland-Altman Plots indicating mean bias scores and 95% limits of agreement (LoA) viii

9 CHAPTER ONE INTRODUCTION Digital health, specifically wearable biosensor technology, has recently undergone tremendous advancements in the sport, fitness, and health industries. Initially developed to augment personal fitness and performance with basic quantitative feedback, the newest generation of devices, i.e. activity trackers, provides real-world, immediate feedback on multiple biometrics related to the consumers physical activity, health, and exercise quality (Lyons, Lewis, Mayrsohn, & Rowland, 2014). With continuous technological advancements in wearable biofeedback devices, the potential applications have also expanded to include medical surveillance, non-invasive medical care, and mobile healthwellness monitoring (Lyons et al., 2014). The pursuit for practical and accurate methods to assess personal health and physical activity continues to lay emphasis on wearable and sophisticated biosensor technologies. It has been previously suggested that integrative biometric processing through multiple sensors may support activity-specific prediction algorithms to accurately compute real-life energy expenditure (EE) and overall activity level (Chen & Bassett, 2005; Doherty & Oh, 2012; Gao, Bourke, & Nelson, 2014). This has stimulated the adoption of multi-sensor technology in nascent activity trackers which has shown to outperform more rudimentary monitors that utilize basic accelerometer data to infer movement, EE, and subsequently overall activity level. Moreover, with the inclusion of sophisticated reflective photoplethysmography (PPG) technology, contemporary activity trackers such as the Basis Peak TM (BPk) and Fitbit Charge HR TM (FB) have the capacity to use heart rate-derived algorithms to support EE (Keytel et al., 2005; Luke, Maki, 1

10 Barkey, Cooper, & McGee, 1997; Wallen, Gomersall, Keating, Wisloff, & Coombes, 2016). Besides predicting EE, PPG-based heart rate (HR) detection allows for unobtrusive objective monitoring of physical exertion as well as systematic exercise prescriptions. The use of PPG technology for HR monitoring has shown acceptable validity, however there is inherent variability, indicating that the accuracy of these trackers is dependent on the specific device used and the type and intensity of activity. Therefore, PPG-based activity trackers remain under scrutiny due to the number of extrinsic factors that may interrupt proper HR detection (i.e. ambient light, sweat, anatomical placement, movement, skin contact force). As the use of wearable biosensors become more prevalent in modern-day society, the accuracy of the physiological data they provide, such as HR, increases in importance. With the recent development of new types of activity trackers incorporating PPG-based HR sensors, there has been a steady focus on improving overall device performance. With that said, validation of HR measurement accuracy for such multi-sensor activity trackers becomes a critical step in improving the consumer experience and at times, their safety. Notwithstanding, there is a present lack of rigorous, scientifically-based validation studies on the accuracy of multi-sensor PPG activity trackers when compared to gold-standard methods for HR measurements. The aforementioned commercial devices are no exception. Thus, future studies to assess the validity of HR measurements from commercially available activity trackers like the BPk and FB against a criterion measure ECG would be highly warranted. 2

11 Statement of the Problem Valid and thereby accurate measurement of HR is one of the central foci of commercially available activity trackers. Recent studies have shown varying degrees of reliability and accuracy in HR monitoring with multi-sensor activity trackers incorporating reflective PPG (Parak & Korhonen, 2014; Wallen et al., 2016). In particular, there is a lack of data supporting the accuracy of the BPk and FB, two commercially available activity trackers incorporating PPG technology. Thus, an extensive validation study examining the accuracy of the BPk and FB for HR measurements would provide critical information to the consumer regarding the efficacy of these activity trackers. Purpose Statement The purpose of this study is to evaluate the validity of HR measurement by the BPk and FB against the criterion measure ECG across an array of structured laboratory-based exercise tasks. Specific Aim Specific Aim: In 24 healthy male and female subjects, we will test the accuracy of the Basis Peak and Fitbit Charge HR wrist-worn activity trackers and their respective proprietary PPG-based sensors for the measurement of heart rate across a number of structured laboratory-based exercise tasks. Hypothesis Alternative Hypothesis (HA): Across all exercise tasks, the activity trackers will accurately measure heart rate when compared to simultaneous criterion 12-lead (electrocardiograph) ECG. 3

12 Significance of the Study The results of the proposed study are projected to offer further practical evidence concerning the validity of PPG-based activity trackers, in particular the BPk and FB. Moreover, the data from the planned study may contribute to the ever-advancing fitness, sport, and medical industries and stimulate further advancement of activity tracker technology for the purpose of monitoring heart rate, i.e. health, performance, and fitness progress. Limitations The limitations to the proposed study include the following: 1) the algorithms used for each device will be proprietary and will not be disclosed to the investigators by the proprietor, 2) performance of PPG-sensors may be altered by wrist placement (dominant vs. non-dominant) however counterbalancing methods will be implemented to address this limitation to the best ability, and 3) various sizes of the FB will be utilized to ensure proper fit and thus the same device will not be used for all subjects. Delimitations The measure of analysis will be delimited to healthy male and female subjects between the ages of years. Participation will be denied if the subject has any major chronic diseases or physical conditions in which exercise or proper interpretation of data would be contraindicated. Operational Definitions Endurance Exercise- exercise conditions that are low-force and long-duration in nature and involves continuous repeated muscle contractions. 4

13 Recreationally Trained- those individuals that exercise regularly at least 3 hours/week within the past 6 months. Reflective Photoplethysmography- an optical technique that captures the change in arterial blood volume by means of the measurement of light that is reflected with each beat (Fukushima, Kawanaka, Bhuiyan, & Oguri, 2012). Biosensor- a device which converts a biological response into an electrical quantifiable signal (Asare, Kviesis-Kipge, Ozols, Spigulis, & Erts, 2012). Heart Rate- the number of times a person s heart beats per minute. Electrocardiography- The process of recording the electrical activity of the heart over a period of time using electrodes placed on the skin. These electrodes detect the electrically acting changes on the skin that arise from the heart muscle's electrophysiologic pattern of depolarizing during each heartbeat Activity Tracker- A consumer-use device that monitors activity-related biometrics like heart rate and energy expenditure 5

14 CHAPTER TWO LITERATURE REVIEW Multi-Sensor Activity Trackers With increasing global prevalence of lifestyle-mediated health epidemics, such as obesity and heart disease, today s society has shown concomitant awareness for healthy and active living. Accordingly, biosensor technology for personal biometric tracking has undergone tremendous advancements in the recent years in efforts to provide consumers an effective means for monitoring various health-related variables (Mercer et al., 2016). This technology has been incorporated into wearable devices often referred to as activity or fitness trackers. With easy access to basic quantitative data, activity trackers provide useful and immediate feedback on multiple biometrics related to the consumers physical activity, exercise quality, and occasionally nutritional status and sleep behavior (Doherty & Oh, 2012). The pursuit for practical and accurate methods to assess personal health and physical activity level continues to lay emphasis on wearable and sophisticated biosensor technologies. It has been recently postulated that integrative biometric processing through accelerometers together with other types of biosensors (e.g. body temperature, skin galvanic response, heart rate, etc.) may support activity-specific prediction algorithms to accurately compute real-life energy expenditure (EE) and overall activity level (Chen & Bassett, 2005). This has stimulated the adoption of multi-sensor technology in modern activity trackers which has shown to outperform more rudimentary monitors that utilize basic accelerometer data to infer movement and subsequent EE (Doherty & Oh, 2012). Besides EE, recent advancements in wearable activity trackers allow for the measurement of heart rate (HR), a biometric traditionally used as a proxy for physical 6

15 exertion. Customarily, consumer-level HR monitors require an external receiver/monitor (typically a wrist watch) and a signal transmitter (typically a chest strap). With most wearable technologies, a more simplified and versatile device may be most conducive to consumer use. Thus, optical HR sensors have been incorporated into the latest multisensor activity tracker technologies to allow for a more streamlined and multi-function experience for personal fitness and health monitoring (Mercer et al., 2016). The Basis Peak TM (BPk) and Fitbit Charge HR TM (FB) wrist-worn activity trackers are an example of current generation devices that integrate multi-sensor technology including reflective photoplethysmographic (PPG) HR detection to offer a more practical, consumer-friendly experience. However, multi-sensor activity trackers incorporating PPG-based optical sensors remain under much scrutiny in regards to its accuracy for personal HR tracking, especially considering the number of extrinsic factors that may interrupt proper biometric computation (i.e. ambient light, sweat, anatomical placement, movement) (Allen, 2007). Therefore, it is imperative that commercial activity trackers, such as the BPk and FB, compare favorably to gold standard laboratory or clinical methods of measurement for HR. Laboratory to Practical Methods for Heart Rate Monitoring The electrocardiogram (ECG) has been the gold-standard for cardiac testing in the clinical and research settings since the early 1900s (Fye, 1994; Kligfield, 2002; Kligfield et al., 2007). Thus, ECG has been utilized as the criterion measure for most validation studies on clinical to consumer-based heart rate detecting devices. The ECG incorporates biosensing technology to directly detect the neuroelectrical activity responsible for the contraction of cardiac muscle and therefore the heart. It may be considered the most 7

16 direct method of detecting the heartbeat among scientists. The orderly series of electrical signals during cardiac depolarization and repolarization are represented as cardiac waveforms on the ECG, i.e. P-wave, QRS complex, and T wave. The greatest ECG amplitude signal in a healthy individual is the R-wave of the QRS complex representing ventricular depolarization and contraction. On an ECG, the QRS complex would represent the heart rate. The HR would then be extrapolated from the time interval between each R-wave, i.e. R-R interval. The 12-lead ECG is considered the standard method although 3- and single-lead approaches have been validated previously (Fye, 1994). In terms of HR measurements, a single-lead ECG is sufficient as the R-R interval does not vary among leads. Regardless of ECG methodology, testing procedures require the use of multiple silver/silver-chloride electrodes adhered to specific anatomical sites on the upper torso. The electrodes detect the electrical activity of the heart during contraction and carries the signals through wires to an analog to digital converter. Then, the signals are translated to visual ECG cardiac waveforms for further analysis. Adding to the cumbersome nature of ECG procedures as described, the anatomical electrode sites must be pre-prepared by cleaning, shaving, and abrading the skin to avoid signal artifact and noise. Although the ECG is an important and most accurate tool for medical level cardiac testing, it is certainly not the most practical for normal consumer-level use for heart rate monitoring. For applications in sport, fitness, and general health tracking, ECG technology must be streamlined for practical use, however, with that, questions of accuracy emerge. Given the need for consumer-based HR monitoring for various subclinical applications like sport and exercise, Polar technology was developed in the early 1990s to 8

17 essentially incorporate the ECG concept into more streamlined consumer-friendly devices (Achten & Jeukendrup, 2003). Polar monitoring initially utilized the fingers and hands as electrode sites for HR detection, but later moved closer to the heart for more direct and less interrupted cardiac signal detection. Thus, modern generation Polar monitors require the use of a chest strap integrated with signal acquiring electrodes. The signal is then digitized and subsequently transmitted, often wirelessly, to a receiver/monitor reflecting a HR value. Therefore, Polar HR monitoring involves at least 2 separate and wirelessly integrated devices, i.e. the transmitter and receiver. The Polar monitor is arguably the most utilized and practical method for HR tracking in an array of exercise to laboratory applications. This may be due to the substantial pool of data that supports the validity and thus the accuracy of Polar devices in reference to the criterion measure ECG (Vanderlei, Silva, Pastre, Azevedo, & Godoy, 2008). The advantage of using a Polar monitor would be the mobility aspect in a more exercise- or sport-oriented setting. It would also allow researchers to record HR accurately and efficiently in the laboratory setting (Achten & Jeukendrup, 2003) without the cumbersome procedures of ECG. Polar devices, however, poses a disadvantage for practical everyday use for the consumer and fails to provide a useful method for long-term health and fitness tracking. Although Polar technology remains one of the most practical as well as accurate means of tracking HR acutely, new generation activity trackers are trending towards a single multisensor device that can monitor various biometrics, like HR, for extended periods and provide consumers with more inclusive or daily feedback on their health and fitness. Thus, modern-day activity trackers incorporate alternate biosensor technology for HR detection that does not require the use of an external electrode-based device. 9

18 Compared to Polar technology, a more indirect method of HR monitoring known as reflective PPG has recently been introduced to the activity tracker industry (Elgendi, 2012; Fukushima et al., 2012; Parak & Korhonen, 2014). Reflective PPG is a method of detecting and measuring HR by estimating the change in blood capillary volume per heartbeat using infrared or green LED lights (Elgendi, 2012; Fukushima et al., 2012). The fundamental form of PPG technology requires minimal optical and electronic components including a light emitting diode (LED) and a photodetector which receives reflected LED light (Allen, 2007; Lee et al., 2013). These components may be easily incorporated into a wearable device such as a wrist watch. During PPG-based HR monitoring at the skin, the LED emits a green light into the skin and superficial blood capillaries (Lee et al., 2013; Maeda, Sekine, & Tamura, 2011). Green light with a wavelength of approximately 500nm is highly absorbed by blood and thus poorly reflected (Figure 1). The amount of green light reflected back to the sensor is detected by the photodetector. Also, the amount of reflected light changes with altering blood volume. The capillary blood volume decreases temporarily during a heartbeat (systole), thus decreasing the absorption and increasing reflectance of the green light. The photodetector senses this altering reflectance between systole and diastole, and the signal is subsequently used to compute HR through specialized algorithms. With the available technology in PPG HR detection, the latest multi-function/sensor activity trackers allow for HR measurements to be accessible in a single device. Wallen et al. (2016) conducted a study in which four commercially available activity trackers were assessed for reflective PPG-based HR measurement accuracy. In this study, investigators examined the validity of HR measurement of the Apple Watch, 10

19 FB, Samsung Gear S, and Mio Alpha wearable devices in twenty-two healthy male or female participants. The participants completed a one-hour protocol involving supine and seated rest, treadmill walking and running, and stationary cycling. Two devices were tested per session and worn according to manufacturer instructions. The Samsung Gear S demonstrated the weakest correlation with ECG (r=0.67) while the Apple Watch exhibited the strongest (r=0.95). The Apple Watch presented with the least error (-1.3 bpm difference) while the FB showed the greatest error (-9.3 bpm difference). There is reference to these devices as accurate, however, authors do not specify whether the devices can be deemed valid. Moreover, despite each device incorporating similar PPGbased HR detection methods, there is clear inherent variability in performance or accuracy amongst these activity trackers. This may likely be due to the unique aspects of each device, such as algorithms, physical characteristics of the tracker, and quality of sensor. Thus, not all PPG-based trackers produce equivalent measurements of HR. To our knowledge, Wallen et al. (2016) remains the only comprehensive validation study on PPG activity trackers. However, the limitations of this study include crude methodology for HR data acquisition and relatively low sampling frequency and thus, a limited pool of HR data for analysis. Moreover, HR data was not simultaneously collected from all devices clearly due to the number of devices tested. Also, possibly due to the limited data pool, researchers failed to examine the accuracy of the activity trackers in a task-specific manner (per exercise, high vs. low intensity, etc.). Thus, further validation studies on commercially available activity trackers for reflective PPG HR measurements must address the aforementioned limitations to add strength to the evidence concerning validity and accuracy. 11

20 Figure 1. Reflective Photoplethysmography for Heart Rate Detection. Adapted from: Future Research and Conclusion Valid and thereby accurate measurement of HR is one of the central foci of commercially available activity trackers. As described above, recent studies have shown varying degrees of performance and accuracy in HR monitoring with multi-sensor activity trackers incorporating reflective PPG (Wallen et al., 2016). Thus, continuous validation work must be performed as newer generation PPG-based activity trackers become available for consumer use. In particular, future validation studies must incorporate sophisticated, high-sampling frequency methods of data acquisition to provide an adequate data pool for comprehensive analysis. Also, testing device accuracy under specific exercise tasks would be especially insightful as it is currently unknown 12

21 how movements may affect device performance. In so doing, results would be anticipated to provide critical information to the consumer regarding the efficacy of these activity trackers. 13

22 CHAPTER THREE METHODOLOGY Experimental Design Twenty-four healthy subjects (12 males and 12 females) (age= years) participated in this crossover design study. Each subject underwent the same 77-minute protocol during a single visit to the Human Performance Research Laboratory at California State Polytechnic University, Pomona (CPP). Initially, participants completed and signed an informed consent form and underwent assessment for basic anthropometric measures (i.e. height, bodyweight). Subsequently, subjects underwent preparatory steps for 12-lead ECG testing procedures for criterion measurement for heart rate (HR) (described below). Next, researchers placed the Basis Peak (BPk) and Fitbit Charge HR (FB) on opposite wrists. Wrist assignment (i.e. dominant or non-dominant) alternated between subjects for counterbalance and to contend with any confounding effects associated with wrist assignment. Care was taken to follow proper use guidelines as suggested by the manufacturer for each device. Time synced HR data from each device (test devices and ECG) were concurrently and continuously acquired second-by-second throughout the entire 77-minute protocol for each participant. During data acquisition, participants completed an initial rest period (lying down) of 15 minutes followed by 5 minute periods of each of the following activities: low cycling (60W resistance), intense cycling (120W), walk ( mph), jog ( mph), run ( mph), arm raises with self-selected resistance (12 reps), lunges with self-selected resistance (12 reps), and plank (60-second hold). In between each exercise task a 5-minute rest period was implemented. 14

23 Participants We recruited 24 healthy subjects (12 males and 12 females) (age= 24.8 ± 2.1 years, weight= 71.3 ± 9.9 kg, height= ± 6.8 cm). Recruitment of potential study subjects was performed by posting flyers on the CPP campus as well as by mass solicitation. Interested individuals were provided with a full overview of the study procedures as well as the study consent form. An incentive for study participation included providing the study participant with a complimentary BPk smart watch at the conclusion of the study. Informed consent was obtained after discussing the study procedures in detail, including the voluntary nature of participation and notification that the subject can withdraw at any time. Upon the subject s agreement to participate, a signed copy was given to the subject. The study was approved by the CPP Institutional Review Board. Participants were included in the study if they were physically able to perform the exercise requirements of the study protocol with no to minimal risk for injury or health complications. Participants were excluded from the study if they reported or exhibited any significant medical conditions, including cardiovascular or pulmonary disease that may limit ability to exercise or increase the cardiovascular risk of exercising. Experimental Procedures Screening All subjects completed a pre-participation medical questionnaire (PAR-Q) and a habitual physical activity questionnaire. 15

24 Anthropometry Body weight was measured to within 100 grams on a calibrated scale. Height was determined to the nearest centimeter using a precision stadiometer. Body Mass Index was calculated as weight in kg divided by height in meters squared. Criterion Measure of Heart Rate A standard 12-lead electrocardiograph (ECG) system (Cosmed C12x; Concord, CA, USA) integrated to a laboratory-grade metabolic testing system (ParvoMedics TrueOne 2400; Salt Lake City, Utah, USA) was used to provide criterion measures of heart rate and energy expenditure (energy expenditure will not be included in analysis). For ECG testing, the electrode placement sites were prepared by standardized procedures of cleaning, shaving, and abrading the skin to improve signal acquisition and to minimize noise artifact. Ten silver/silver-chloride self-adhesive electrodes (RA, LA, RL, LL, V1, V2, V3, V4, V5, and V6) were placed on the upper torso according to the Mason-Likar lead placement configuration for the 12-lead ECG (Mason and Likar, 1966). The ECG wires were secured to the body using elastic foam wrap and a standard ECG stress test belt to minimize electrode movement artifact. The ECG and indirect calorimetry system was integrated and thus metabolic and heart rate data acquisition was simultaneous and time-synced. Although, metabolic data was excluded in the analysis, the integration of testing equipment was necessary to obtain exportable second-by-second HR records. Heart rate data per second was converted to beats per minute (BPM) automatically by the data acquisition software program prior to analysis. 16

25 Activity Tracker Procedures The BPk and FB was attached to opposing wrists on the subject according to manufacturer instructions. Half of the subject pool wore the BPk on the dominant wrist and the FB on the non-dominant wrist. The other half of the subject pool wore the BPk on the non-dominant wrist and the FB on the dominant wrist. We implemented this counterbalancing strategy to avoid any potential confounding factors associated with the wrist on which the devices are placed. Data acquisition from each device along with ECG and IC was time-synced according to a single master clock. Heart rate and EE data was exported from the BPk using a data extraction software program customized specifically to sync with the BPk via Bluetooth transmission upon completion of the testing protocol. For the FB data acquisition, the track exercise function on the mobile application interfaced with the FB device was used. This function allowed for time-synced GPS and HR data acquisition. Upon completion of the testing protocol, the exercise metrics during the tracked exercise were uploaded to the Fitbit servers on a personal online dashboard. Subsequently, the GPS (.tcx) file linked to the tracked exercise was downloaded from the Fitbit online dashboard and imported into a Microsoft Excel spreadsheet. The spreadsheet displayed time-synced, second-by-second GPS and HR data. The GPS data was discarded while the HR data was subsequently used for analysis. The mobile application settings for the FB was adjusted appropriately for each subject. Each device was confirmed to have full battery charge prior to testing. Statistical Analysis Three levels of statistical analysis were implemented to assess the level of accuracy of the consumer devices in reference to ECG: i) Pearson Product-Moment 17

26 Correlation analysis was used to determine the strength of relationship between ECG and each of the consumer devices and whether the relationship was statistically significant. A significant correlation was determined if the p-value was less than 0.05 while the strength of correlation was determined by the correlation coefficient (r-value). The strength of correlation was determined as follows: = strong, = moderately strong, = moderate, = moderately weak, <0.59= weak; ii) the Bland-Altman method was used to further assess the agreement between the consumer devices and ECG for heart rate measurements and whether the differences varied in a systematic or ambiguous way over the range of measurements (Bland & Altman, 1986). The mean bias between consumer heart rate device and ECG (=FB HR ECG HR and =BPk HR ECG HR) and the 95% limits of agreement (LoA; LoA = mean bias/difference ± 1.96 standard deviation of the difference) were identified. Bland-Altman plots graphed the individual difference scores against the individual averages of the heart rate values obtained by the consumer device and ECG; and iii) Mean absolute differential between the consumer device and the ECG were computed. This represented the average difference score regardless of direction of the difference (i.e. regardless of under- or overestimation). All three levels of analysis were implemented on aggregate HR data, HR data above the mean ECG HR, HR data below the mean ECG HR, resting or recovery HR data, and task-specific HR data. Previous validation studies (Dolezal et al., 2014; Terbizan, Dolezal, & Albano, 2002) provided validity criteria for heart rate measurement as: 1) a correlation between ECGderived heart rate and the heart rate measured by the test device of r=0.90 or greater, 2) a mean bias less than 3 beats/min, and 3) a standard error less than 5 bpm. Satisfaction of at least the first and second criteria together or all criteria was used to determine validation 18

27 of the consumer devices for accurate HR measurement. However, results from all three aforementioned analyses were considered and discussed to provide a more inclusive examination on the accuracy of each device. Based on the assumption that the true correlation between ECG-derived HR and device-measured HR is associated with ρ=0.85, a sample size of 24 subjects was determined (Wallen et al., 2016). 19

28 CHAPTER FOUR RESULTS Aggregate Heart Rate Data Basis Peak vs. ECG When examining time-synced ECG and Basis Peak HR data in aggregate (inclusion of all resting and exercise conditions) (n= 87,340), there was a significant (p<0.0001) and strong positive correlation between ECG and BPk (r= 0.92) (Table 1 and Figure 1). The mean HR from the BPk and ECG was ± 28.0 bpm and ± 29.0 bpm, (p<0.05) respectively. There was a mean absolute differential of 5.9 ± 9.8 bpm or 5.3 ± 8.3% between BPk and ECG. The BPk demonstrated a mean bias of -2.5 ± 11.1bpm (95% CI -2.6, -2.5, LoA 19.3, -24.4) in reference to the ECG criterion measure (Table 1 and Figure 1). Fitbit Charge HR vs. ECG When examining time-synced ECG and FB HR data in aggregate (n= 87,340), there was a significant (p<0.0001) and moderately strong positive correlation between ECG and FB (r= 0.83) (Table 2 and Figure 1). The mean HR from the FB and ECG was ± 29.0 bpm and ± 29.0 bpm, (p<0.05) respectively. There was a mean absolute differential of 10.6 ± 15.8 bpm or 9.8 ± 14.0% between FB and ECG. The FB demonstrated a mean bias of -8.8 ± 16.9 bpm (95% CI -8.9,-8.7, LoA 24.2, -41.8) in reference to the ECG criterion measure (Table 2 and Figure 1). 20

29 Table 1. Results for Basis Peak. *p< Parameter Basis Peak Mean HR (bpm ± SD) ECG Mean HR (bpm ± SD) Mean Absolute Difference (bpm ± SD) Mean Absolute Difference (% ± SD) Aggregate Data (n=87,340) Data above ECG HR >116bpm (n=41,315) Data below ECG HR <117bpm (n=46,025) ± ± ± ± ± ± ± ± ± ± ± ± 7.0 Correlation (r) 0.92* 0.77* 0.84* Mean Bias -2.5 ± ± ± 8.2 (bpm ± SD) (95% CI -2.6, -2.5) (95% CI -5.0, -4.7) (95% CI -0.5, -0.4) 95% Limits of Agreement (Upper, Lower) Standard Error of the Estimate (SEE) 19.3, , , Table 2. Results for Fitbit (FB) Charge HR. *p< Parameter FB Charge HR Mean HR (bpm ± SD) ECG Mean HR (bpm ± SD) Mean Absolute Difference (bpm ± SD) Mean Absolute Difference (% ± SD) Aggregate Data (n=87,340) Data above ECG HR >116bpm (n=41,315) Data below ECG HR <117bpm (n=46,025) ± ± ± ± ± ± ± ± ± ± ± ± 10.4 Correlation (r) 0.83* 0.58* 0.73* Mean Bias -8.8 ± ± ± 10.7 (bpm ± SD) (95% CI -8.9, -8.7) (95% CI -12.9, -12.5) (95% CI -5.4, -5.2) 95% Limits of Agreement (Upper, Lower) Standard Error of the Estimate (SEE) 24.2, , ,

30 a. b. c. d. Figure 2. Results for aggregated data set. Correlation between each test device and ECG and Bland-Altman Plots indicating mean bias scores and 95% limits of agreement (LoA). HR Data above Mean ECG HR (>116 bpm) Basis Peak vs. ECG Time-synced HR data pairs above the mean ECG HR (>116 bpm; n= 41,315 pairs) were analyzed. During conditions in which the ECG HR (true HR) exceeded 116 bpm, there was a significant (p<0.05) and moderately strong positive correlation between BPk and ECG (r= 0.77) (Table 1 and Figure 2). Furthermore, the mean HR from the BPk and ECG was ± 20.4 bpm and ± 18.5 bpm, respectively. There was a mean absolute differential of 6.9 ± 12.4 bpm or 5.3 ± 9.5% between BPk and ECG. The BPk 22

31 demonstrated a mean bias of -4.9 ± 13.4 bpm (95% LoA 21.3, -31.0) in reference to the ECG criterion measure (Table 1 and Figure 2). Fitbit Charge HR vs. ECG During conditions in which the ECG HR (true HR) exceeded 116 bpm, there was a significant (p<0.05) and weak positive correlation between FB and ECG (r= 0.58) (Table 2 and Figure 2). Furthermore, the mean HR from the FB and ECG was ± 25.3 bpm and ± 18.5 bpm, respectively. There was a mean absolute differential of 13.8 ± 20.3 bpm or 11.3 ± 17.0% between FB and ECG. The FB demonstrated a mean bias of ± 21.1 bpm (95% LoA 28.6, -54.0) in reference to the ECG criterion measure (Table 2 and Figure 2). a. b. c. d. Figure 3. Results for data set above mean ECG heart rate (>116 bpm). Correlation between each test device and ECG and Bland-Altman Plots indicating mean bias scores and 95% limits of agreement (LoA). 23

32 HR Data below Mean ECG HR (<117 bpm) Basis Peak vs. ECG Time-synced HR data pairs below the mean ECG HR (<117 bpm; n= 46,025 pairs) were analyzed. During conditions in which the ECG HR (true HR) was below 117 bpm, there was a significant (p<0.05) and moderately strong positive correlation between BPk and ECG (r= 0.84) (Table 1 and Figure 3). Furthermore, the mean HR from the BPk and ECG was 93.9 ± 15.0 bpm and 94.3 ± 14.2 bpm, respectively. There was a mean absolute differential of 4.9 ± 6.6 bpm or 5.3 ± 7.0% between BPk and ECG. The BPk demonstrated a mean bias of -0.4 ± 8.2 bpm (95% LoA 15.6, -16.4) in reference to the ECG criterion measure (Table 1 and Figure 3). Fitbit Charge HR vs. ECG During conditions in which the ECG HR (true HR) was below 117 bpm, there was a significant (p<0.05) and moderate positive correlation between FB and ECG (r= 0.73) (Table 2 and Figure 3). Furthermore, the mean HR from the FB and ECG was 89.0 ± 15.2 bpm and 94.3 ± 14.2 bpm, respectively. There was a mean absolute differential of 7.6 ± 9.2 bpm or 8.5 ± 10.4% between FB and ECG. The FB demonstrated a mean bias of -5.3 ± 10.7 bpm (95% LoA 15.7, -26.3) in reference to the ECG criterion measure (Table 2 and Figure 3). 24

33 a. b. c. d. Figure 4. Results for data set below mean ECG heart rate (<117 bpm). Correlation between each test device and ECG and Bland-Altman Plots indicating mean bias scores and 95% limits of agreement (LoA). Task-Specific Heart Rate Comparison Correlation coefficient, mean bias, and SEE for the BPk and FB for each exercise task and rest/recovery periods are indicated in Table 3. Across all exercise tasks and resting/recovery conditions, the BPk demonstrated significant (p<0.05) moderate to strong correlations with ECG while the FB exhibited significant (p<0.05) weak to moderately strong correlations. Both devices performed best during resting/recovery conditions as indicated by relatively strong correlations to ECG and small mean bias scores (BPk: r= 0.96, mean bias= -0.2 bpm; FB: r= 0.83, mean bias= -3.7 bpm). During exercise conditions, the BPk provided accurate HR readings during low intensity cycling, walking, jogging, and running. The accuracy for the BPk was slightly compromised 25

34 during higher intensity cycling, resisted arm raises, and resisted lunges while the greatest detriment to performance occurred during the isometric plank (r=0.78, mean bias= -4.7 bpm). During walking, jogging, and running tasks, the FB performed with strong correlation (r=0.94 to 0.95) and excellent agreement to ECG (about -4 to -3 bpm bias). However, performance of the FB was exceptionally poor during cycling, resisted arm raises, lunges and isometric plank. In particular, the FB and ECG had the weakest correlation and agreement during the resisted lunges (r=0.28, mean bias= bpm) and isometric plank (r=0.26, mean bias= bpm). Table 3. Activity-specific results for Basis Peak and Fitbit Charge HR. Standard error (SE). *p<

35 CHAPTER FIVE DISCUSSION The main objective of this study was to determine if commercially available wrist-worn activity trackers like the Basis Peak and Fitbit Charge HR measures HR accurately in reference to the criterion measure ECG. The use of reflective PPG in fitness wearables is understood to be a relatively new application for consumer level HR tracking, and inherent variability in accuracy may likely exist among various devices. In previous findings, conditions of low physical exertion elicited the least variability in error values among various trackers as similarly demonstrated by the present investigation (Stahl, An, Dinkel, Noble, & Lee 2016). Stahl et al. (2016) also showed that the TomTom Runner Cardio (3.3%) and the BPk (3.6%) performed on average with greater accuracy than the Scosche Rhythm (4.0%), Mio Alpha (4.6%), Microsoft Band (4.8%), and FB (6.2%). Similar to the present findings, the FB appeared to perform worse than the BPk. Interestingly, correlation data conflicted with our findings. It was stated that both BPk and FB had very strong correlation with ECG (r=0.9) for HR measurements across all exercise tasks (walking and running) (Stahl et al., 2016). Comparatively, our correlation results are in agreeance with the BPk (r=0.92), but not with the FB (r=0.83). This difference between the present study and Stahl et al. (2016) may be explained by the abundance of HR data collected at higher physical exertions in the current protocol. As stated earlier, the accuracy of HR measurements by the FB was largely underestimated with higher levels of physical exertion. Also, the protocol implemented by Stahl et al. (2016) allowed for only manual data collection every minute. This, in turn, would decrease the data pool per exercise and limit investigators to a smaller, less extensive 27

36 analysis as well as increase the propensity for human error in data acquisition. Having a larger data pool with second by second electronic data acquisition may eliminate chance for human error and afford a more comprehensive and inclusive analysis. Thus, to our knowledge, the present study was the most comprehensive analysis of PPG-based wearables for HR accuracy. When reviewing all acquired time-paired data (n=87,340 pairs), the BPk met the proposed validity criteria for HR detecting and monitoring devices (r 0.90 and mean bias < 3 bpm). We observed a strong correlation (r=0.92) between the BPk and ECG with an acceptable mean bias of -2.5 bpm and an absolute differential from criterion measurements of approximately 5%. Based on the Bland-Altman analysis and the 95% limits of agreement (+19.3 to bpm), the bias of the BPk may be reasonably described as systematic and, therefore, may be used interchangeably with ECG for accurate HR measurements. Interestingly, the accuracy of the BPk declines slightly with an increase in physical exertion. This was observed in the analysis of data pairs associated with an ECG HR above the mean ECG HR of 116 bpm (n=41,315). Specifically, the correlation between BPk and ECG weakened per coefficient of 0.77, and the mean bias slightly increased to -4.9 bpm. Even so, the data set above the ECG mean showed a strong correlation between the BPk and the criterion measure ECG. When examining the data acquired during resting and recovery stages and most exercise tasks specifically, the BPk performed comparably accurate with an ideal agreement to ECG (especially during rest/recovery) (Table 3). The BPk demonstrated a marginal decrease in performance only during the intense cycling and isometric plank portions of the protocol. These outcomes are consistent with previous results for other PPG-based devices (Mio 28

37 Alpha and Scosche myrhythm) that revealed a task-specific variation in device performance. For instance, the Mio Alpha showed the greatest error (-4.8%) during cycling while the Scosche myrhythm was during the walking exercise (-3.1%) (Parak & Korhonen, 2014). The FB presented with weaker correlation (r=0.83) and less agreement (mean bias= -8.8 bpm, 95% LoA 24.2, -41.8) to ECG than the BPk when examining the aggregated data set (n=87,340). Although a moderately strong correlation to ECG was found, both the mean bias score and the r-value failed to meet validity criteria. Furthermore, when examining the Bland-Altman analysis for the aggregate data, the FB showed tendencies to underestimate HR. This is especially evident during tasks inducing heightened physical exertion. Hence, the FB cannot be considered valid and, therefore, interchangeable with ECG for accurate HR detection and monitoring. Our data corroborate findings of previous FB results by Wallen et al. (2016) who reported a -9.3 bpm bias, 95% LoA (7.4, bpm), and a correlation coefficient of 0.81 (Wallen et al., 2016). As with the BPk, we observed a decline in the device s performance during exercises eliciting higher ECG heart rates (i.e. >116 bpm). The very weak correlation (r= 0.58) together with the large mean bias (= bpm) and high standard error (= 20.1) strongly suggest the FB to be an inaccurate means of measuring HR during higher physical exertion. The FB seemed to perform best during low physical exertion based on a small average bias of -5.3 bpm. However, even during the lighter physical exertion the correlation strength (r= 0.73) fails to suggest that the FB produces accurate HR measurements. Specifically, during resting or recovery situations, the FB demonstrated a moderately strong correlation with ECG (r= 0.83) and underestimated HR by only

38 bpm on average. The FB performed with moderate accuracy and agreement to ECG during the walk, jog, and run as reflected by strong correlations and relatively low mean bias scores (~ -4 bpm). During the isometric plank, resisted lunges, and cycling, however, the FB demonstrated very weak to moderately weak correlations and large mean bias scores. Thus, in addition to activities of high physical exertion, specific exercise tasks also appear to dramatically limit the performance of the FB. It is evident based on our data as well as others that PPG-based HR monitors experience reduced accuracy during elevated physical exertion and specific exercise tasks, two conceivably interrelated factors. Moreover, the degree of performance detriment and the exercise type related to poor performance varies among the different PPG devices. To investigate the potential factors obstructing accurate measurements during exercise, especially for the FB, we examined the commonalities of the exercise tasks eliciting the worst device performance, i.e. cycling, resisted arm raises, resisted lunges, and isometric plank. Each of these exercises involve sustained or repetitive contractions of forearm skeletal muscles which may influence the attributes of PPG signals necessary for accurate HR detection. Previous evidence suggests that the amount of pressure or compression exerted on the device by the consumer may play a role in the signal waveform and thereby, PPG quality (Allen, 2007; Rafolt & Gallasch, 2004; Teng & Zhang, 2004). Moreover, an increased compression of the device against the wrist may illicit noise artifacts which would ultimately disrupt adequate signal processing. This, in turn, would interpose proper HR detection and an accurate reading (Teng & Zhang, 2004). Accordingly, manufacturer instructions for wrist-worn PPG devices often recommend the user to refrain from overtightening the strap as to avoid large sensor to 30

39 skin contact force or compression. During exercises that involve sustained or repeated forearm muscle contractions, there is an imminent increase in the compression force between the wrist and the device s PPG sensor. Elevated contact force between the FB and the skin may be a plausible explanation to the lack of performance evident during these specific exercises. However, this speculation is unsubstantiated by our data since the effects of contact force on device accuracy was not the scope of the present investigation. Nonetheless, both test devices were worn according to manufacturer recommendations throughout the entire protocol. Hence, the activity trackers under investigation, particularly the FB, may require closer examination as it relates to 1) whether or not the amount of force applied to the device by muscular contractions varies among different exercises, and 2) if so, whether or not altering compression force on the device s sensor during exercise would affect the quality of the PPG signal and overall performance of HR tracking. Overall, HR measurements on the dorsal wrist via reflective PPG sensors has apparent and evident limitations (Rafolt & Gallasch, 2004; Teng & Zhang, 2004). Long established methods of detecting and monitoring heart rate via electrocardiac chest strap sensors and external receivers have continuously demonstrated superb accuracy and agreeability to ECG (Terbizan et al., 2002). However, the technology industry has made a substantial push for a more practical and consumer-friendly wrist-worn biofeedback device. Thus, efforts to improve the technology behind PPG-based HR tracking in multifunction activity trackers require continual monitoring and resolution of extrinsic factors that may disrupt PPG signals and overall accurate HR detection. 31