Wearable Finger Pulse Oximetry for Continuous Oxygen Saturation Measurements During Daily Home Routines of Patients With Chronic Obstructive Pulmonary Disease (COPD) Over One Week: Observational Study

Background Chronic obstructive pulmonary disease (COPD) patients can suffer from low blood oxygen concentrations. Peripheral blood oxygen saturation (SpO2), as assessed by pulse oximetry, is commonly measured during the day using a spot check, or continuously during one or two nights to estimate nocturnal desaturation. Sampling at this frequency may overlook natural fluctuations in SpO2. Objective This study used wearable finger pulse oximeters to continuously measure SpO2 during daily home routines of COPD patients and assess natural SpO2 fluctuations. Methods A total of 20 COPD patients wore a WristOx2 pulse oximeter for 1 week to collect continuous SpO2 measurements. A SenseWear Armband simultaneously collected actigraphy measurements to provide contextual information. SpO2 time series were preprocessed and data quality was assessed afterward. Mean SpO2, SpO2 SD, and cumulative time spent with SpO2 below 90% (CT90) were calculated for every (1) day, (2) day in rest, and (3) night to assess SpO2 fluctuations. Results A high percentage of valid SpO2 data (daytime: 93.27%; nocturnal: 99.31%) could be obtained during a 7-day monitoring period, except during moderate-to-vigorous physical activity (MVPA) (67.86%). Mean nocturnal SpO2 (89.9%, SD 3.4) was lower than mean daytime SpO2 in rest (92.1%, SD 2.9; P<.001). On average, SpO2 in rest ranged over 10.8% (SD 4.4) within one day. Highly varying CT90 values between different nights led to 50% (10/20) of the included patients changing categories between desaturator and nondesaturator over the course of 1 week. Conclusions Continuous SpO2 measurements with wearable finger pulse oximeters identified significant SpO2 fluctuations between and within multiple days and nights of patients with COPD. Continuous SpO2 measurements during daily home routines of patients with COPD generally had high amounts of valid data, except for motion artifacts during MVPA. The identified fluctuations can have implications for telemonitoring applications that are based on daily SpO2 spot checks. CT90 values can vary greatly from night to night in patients with a nocturnal mean SpO2 around 90%, indicating that these patients cannot be consistently categorized as desaturators or nondesaturators. We recommend using wearable sensors for continuous SpO2 measurements over longer time periods to determine the clinical relevance of the identified SpO2 fluctuations.


Introduction
Chronic obstructive pulmonary disease (COPD) is a highly prevalent lung disease that is characterized by persistent airflow limitation due to a mixture of obstructive bronchiolitis and emphysema [1]. Morbidity and mortality of COPD are high and still increasing [2], leading COPD to become the third-leading cause of death worldwide by 2030 [3]. COPD patients can suffer from low blood oxygen concentrations due to gas exchange abnormalities [1]. Hypoxemia during the night (ie, nocturnal desaturation) is also common in patients with COPD [4][5][6] due to nocturnal alveolar hypoventilation and ventilation-perfusion mismatching [7]. Hypoxemia can worsen with increasing disease severity [8]. Furthermore, it has been shown that hypoxemia is associated with lower exercise tolerance; decreased quality of life; increased risk for exacerbations, defined as "a sustained worsening of the patient's condition, from the stable state and beyond normal day-to-day variations" [9]; and higher risks of death [8].
The presence of hypoxemia can be assessed by invasive blood gas analyses. A noninvasive method to assess peripheral blood oxygen saturation (SpO 2 ) is pulse oximetry. Spot check SpO 2 measurements with thresholds of 88%-92% have been suggested for the detection of hypoxemia [1,10,11]. In telemonitoring applications, daily SpO 2 spot checks are used to raise alerts for exacerbations when SpO 2 spot check values drop below predefined SpO 2 thresholds [12]. Nocturnal desaturation is usually defined as having an SpO 2 value below 90% for more than 30% of the time in bed, measured during one [6,[13][14][15][16] or two nights [5,[17][18][19][20].
The current SpO 2 monitoring strategies do not take into account natural fluctuations in SpO 2 . Sampling SpO 2 with such a low time frequency or during such a short time period may thus lead to classification errors (ie, hypoxemic or nonhypoxemic and nocturnal desaturator or nondesaturator) or false alerts in telemonitoring applications [12]. Wearable finger pulse oximeters provide the possibility to collect SpO 2 data at higher time frequencies and over longer time periods. This makes it possible to assess and account for oxygen saturation fluctuations in patients with COPD. Therefore, the objective of this study was to use wearable finger pulse oximeters to continuously measure SpO 2 during daily home routines of COPD patients and assess natural SpO 2 fluctuations. We hypothesized that significant natural SpO 2 fluctuations are present within and between multiple days and nights, which may lead to classification errors (ie, nocturnal desaturator or nondesaturator) or false alerts in telemonitoring applications. Figure 1 provides a general overview of the methods applied in this single-center, 1-week observational study. COPD patients at Global Initiative for Chronic Obstructive Lung Disease stages II-IV (GOLD II-IV) were recruited at the Centre of Expertise for Chronic Organ Failure (CIRO), a COPD treatment center located in Horn, the Netherlands, during a standard baseline assessment prior to pulmonary rehabilitation [21]. The target sample size was set a priori to 20 patients. COPD patients were eligible to enroll in the study based on the following criteria: (1) clinically stable (ie, no exacerbation in the past 4 weeks), (2) no rollator use, and (3) no long-term oxygen therapy. Patients that were interested in participating were called a few days after the baseline assessment to schedule a home visit for the delivery of the wearable sensors. During this visit, the functioning of the sensors and data acquisition protocol were explained, written informed consent was obtained, and a new visit was planned for collection of the sensors at the end of the study period. During the 7-day study period, which took place before the start of pulmonary rehabilitation, two phone calls were made to resolve potential technical difficulties. Demographics, oxygen partial pressure in arterial blood (PaO 2 ), postbronchodilator pulmonary function data (ie, forced expiratory volume in 1 second [FEV1], forced vital capacity [FVC], and transfer factor for carbon monoxide [TLCO]), 6-minute walking distance (6MWD), and COPD assessment test (CAT) results were collected during the standard baseline assessment at CIRO. The study was approved by the Medical Research Ethics Committees United (MEC-U) (study approval number: NL58079.100. 16) in the Netherlands and executed between December 2016 and April 2018.

Wearable Sensors and Protocol
Continuous SpO 2 measurements were performed for 1 week using a wearable finger pulse oximeter: WristOx 2 3150 (Nonin Medical). Nonin oximeters have frequently been used for home monitoring of patients with COPD [12] and the WristOx 2 3150 model complies with the International Organization for Standardization (ISO) standards ISO 10993-1 and ISO 80601-2-61. The manufacturer reports an accuracy of ±2% for SpO 2 measurements [22]. Sampling frequency was fixed at 1 Hz. Only the cumulative measurement time was visible for the participants. SpO 2 values were not shown on the WristOx 2 display to prevent patients from changing their behavior when deviating SpO 2 values would occur. Every participant received three WristOx 2 devices to deal with the limited battery life (ie, 48 hours of continuous measurements with one WristOx 2 ). Participants were instructed to wear the WristOx 2 on the index finger of their nondominant hand every night and as much as possible during the day, depending on their daily routines and comfort of wearing the finger clip. Raw data were stored on the internal memory of the WristOx 2 and downloaded at the end of the week using nVISION software, version 6.4.0.10 (Nonin Medical).
Simultaneously, actigraphy was performed with the SenseWear Armband (BodyMedia) for obtaining contextual information about physical activity levels and when the participants were lying down and/or asleep. The SenseWear Armband is a multisensory triaxial accelerometer, combining accelerometry with measurements of heat flux, galvanic skin response, and skin temperature. Based on these measurements, the armband provides information about, for example, energy expenditure (EE), expressed as metabolic equivalent of task (MET), or steps taken, while also indicating when the wearer is lying down and sleeping, at a standard sampling time of 1 sample per minute. Measurements of MET were used for classifying physical activity levels, whereas the indications of lying down and sleeping were used to separate daytime measurements from nocturnal measurements, as further described below. The SenseWear Armband has been shown to be accurate for measurements of both physical activity [23,24], except when using a rollator [24], and sleep estimations [25]. Patients were asked to wear the armband on the left upper arm, except when there was contact with water (eg, when taking a shower). Battery life of one SenseWear Armband was sufficient for continuous 24-hour measurements with a 1-minute sampling time during the whole week. Data were stored on the internal memory and downloaded at the end of the week using the BodyMedia SenseWear 8.1 software (BodyMedia). Figure 2 visualizes the preprocessing steps. The SenseWear Armband indications about lying down and sleep were used to determine the time of going to bed in the evening and the time of getting out of bed in the morning. These time stamps were used to divide SpO 2 data into nocturnal and daytime data (see Figure 2A). Only full-night nocturnal SpO 2 measurements were retained for further analyses and daytime SpO 2 measurements were retained if there was at least one hour of SpO 2 measurements during that day. One hour of daytime measurements was considered sufficient to examine whether significant fluctuations occurred in SpO 2 values during the day. Previous studies examining intraday fluctuations only examined measurements of one hour or less [26][27][28][29]. No days or nights had to be excluded due to lacking SenseWear Armband measurements. Days with at least one hour of measurements in both the afternoon and evening were used for the comparison between afternoon (ie, 13:00-18:00) and evening (ie, 18:00-going to bed) SpO 2 values. No comparison was performed between SpO 2 values in the morning (ie, before 13:00) and the afternoon and evening, because SpO 2 measurements were often not performed before 13:00 (see Multimedia Appendix 1).

Figure 2.
Visualization of the different preprocessing steps. Panel A shows all original data, containing error values, that are divided into daytime and nocturnal data. Panel B zooms in on the effect of data exclusion on a specific part of daytime peripheral blood oxygen saturation (SpO 2 ) data. Panel C zooms in on the effect of down-sampling and interpolating on the same part of daytime SpO 2 data.
The raw SpO 2 data contained error values (ie, the number 500 was provided in the raw data file) when measurements were considered invalid by the algorithms of the WristOx 2 manufacturer (ie, orange dashes in Figure 2A). However, close inspection of the time series indicated that invalid data (eg, sudden low values) surrounded these error values. A data-cleaning algorithm was developed to exclude these invalid SpO 2 values as follows. First, small blocks of data (ie, less than 20 samples) between error values were excluded. Second, bigger blocks of data (ie, between 20 and 100 samples) between error values were excluded, only when the mean SpO 2 value of this data block was deviating more than 6% from the mean SpO 2 value of the full day or night under consideration: both steps are shown in Figure 2B. Thorough visual and raw data examination of both valid and invalid data led to the choice of this 6% threshold for excluding invalid blocks of data (eg, Figure   2B: around 17:00-18:00). Third, remaining outliers were excluded by down-sampling the data (see Figure 2C). Autocorrelation analyses indicated that a sampling time of 20 seconds was appropriate (see Multimedia Appendix 2) and all data were thus down-sampled by taking the median SpO 2 value of each consecutive 20-second block. By taking the median value, the effect of outliers was excluded. As a last step, small gaps of SpO 2 data (ie, 3 or fewer missing samples) were filled using linear interpolation (see Figure 2C). After application of the data-cleaning algorithm, all data were visually checked to ensure valid SpO 2 values were retained, while invalid values were removed.

Data Analyses
The SenseWear Armband indications about lying down and sleep were used to calculate the total night sleeping time (TNST) (ie, sum of all minutes indicated as sleep), wake time after sleep onset (WASO) (ie, sum of all minutes spent awake during the time in bed, after the first onset of sleep), and sleep efficiency (Seff) (ie, the ratio between TNST and time in bed) for every separate night. Weekly averages of TNST, WASO, and Seff were calculated to describe the sleep quality of the included patients.
Daytime SpO 2 data were divided into daytime data during rest (EE≤1.5 MET, while the patient was still awake), during low-intensity physical activities (LIPAs) (1.5 MET<EE≤3 MET), and during moderate-to-vigorous intensity physical activities (MVPAs) (EE>3 MET). Data quality of the continuous SpO 2 measurements was assessed for both nocturnal and daytime data based on the amount of valid data (ie, excluding error values and cleaned values). Furthermore, the effect of physical activity on the amount of valid data was examined by comparing data quality during MVPA with data quality during rest and LIPA (EE≤3 MET).
Mean SpO 2 , SpO 2 SD, and cumulative time spent with SpO 2 below 90% (CT90) were calculated for every separate (1) day, (2) day in rest, and (3) night. Hereafter, the weekly average and weekly range (ie, the difference between the maximum and minimum value over the different days or nights of the same patient) of these features were calculated for every patient.
Intraday and intranight fluctuations were quantified as the weekly average of the standard deviation of the SpO 2 measurements. The range of SpO 2 values during the day in rest, calculated as the difference between the maximum and minimum SpO 2 values of that day in rest, was determined to indicate how much spot-check values in telemonitoring applications could differ depending on the moment of the measurement. Only daytime SpO 2 in rest was considered because only at these moments could spot checks have been performed in a telemonitoring application. In addition, the difference between mean SpO 2 values in rest in the afternoon (ie, between 13:00 and 18:00) and in the evening (ie, between 18:00 and going to bed) was calculated for every day separately to examine differences in SpO 2 baseline levels during the day.
Night-to-night and day-to-day SpO 2 fluctuations were quantified as the weekly ranges of the three features (ie, mean SpO 2 , SpO 2 SD, and CT90). Furthermore, we examined how many of the included patients changed category between nocturnal desaturator and nondesaturator.

Statistical Analysis
Patient characteristics, weekly averages, and weekly ranges were summarized for all patients as mean and SD. Paired-sample t tests were used to test for differences between weekly averages of nocturnal and daytime SpO 2 in rest. Pearson correlations assessed the relationship between mean SpO 2 and patient characteristics, intraday SpO 2 fluctuations, or intranight SpO 2 fluctuations. A P value of <.05 was considered statistically significant. All analyses were carried out in Jupyter Notebooks (Project Jupyter) [30] using the Python 3.5 programming language (Python Software Foundation) [31].

Continuous SpO 2 Measurements, Preprocessing, and Data Quality
An overview of the amount of SpO 2 measurements that were included for further analyses is provided in Table 2 and Multimedia Appendix 1. A total of 2 days and 2 nights of measurements from patient 4 were missing due to battery issues. A total of 3 days and 3 nights of measurements from patient 11 were excluded because correct time stamps were missing, due to the patient unintentionally resetting the time indication by taking out the batteries. The last day and night of patient 11's measurements could not be analyzed because no clear distinction could be made between daytime and nocturnal data. Still, 3 days and 3 nights of patient 11's measurements were used for further analyses.
The SpO 2 dataset contained 1.83% error values and an additional 1.49% were excluded during the first two steps of the data-cleaning algorithm (see Figure 3)

Discussion
Continuous SpO 2 measurements with wearable pulse oximeters identified significant SpO 2 fluctuations between and within multiple days and nights of patients with COPD. COPD patients are known to be a heterogeneous group [32,33], which leads to large differences in mean SpO 2 values between patients. Day-to-day SpO 2 fluctuations have not yet received a lot of attention in scientific literature and night-to-night fluctuations have not yet been studied over more than two nights in patients with COPD. Some studies examined intraday fluctuations, however, these studies only examined short-term SpO 2 fluctuations (ie, measurements of one hour or less) in healthy individuals [29] or infants [26][27][28]. Our results showed that significant SpO 2 fluctuations occurred between and within multiple days and nights for the included patients. This is often not taken into account in telemonitoring applications [12]. When using the current definition of nocturnal desaturation, 50% (10/20) of the included patients changed category between desaturator and nondesaturator over the course of 1 week.
Continuous SpO 2 measurements with high amounts of valid data could be obtained during a 7-day monitoring period, except during MVPA (EE>3 MET). Nocturnal measurements were well-perceived by the patients, leading to the high compliance. These measurements only had a small amount of data that was considered invalid (0.69%). This indicates that it is feasible to perform multi-night continuous SpO 2 measurements. Most patients also performed daytime measurements during rest and LIPA (ie, 17 patients measured, on average, 7.6 hours per day). These had a limited amount of invalid data (6.10%). However, very few measurements were performed during MVPA (ie, 0.2 hours per day) and almost one-third of these measurements produced invalid data due to motion artifacts. Therefore, it might not be recommended to analyze continuous SpO 2 measurements during periods of MVPA.
Two types of preprocessing steps for continuous SpO 2 data have been reported: (1) removing deviating samples or (2) down-sampling the data. Deviating samples have been removed visually [29,34] by using an unspecified data-cleaning algorithm [35] or by removing samples with a sample-to-sample deviation of more than 4% [36] or 8% [37]. The latter method, however, does not remove artifacts that last over longer periods of time (eg, see Figure 2B: around 17:00-18:00). Therefore, artifacts in this study were removed by removing blocks of deviating values instead of separate samples. After sample removal, Morillo et al excluded remaining artifacts by down-sampling the data from 8 Hz to 1 Hz [37]. However, autocorrelation analyses in our study showed that a sampling time of 1 sample per second still led to oversampling of the data (see Multimedia Appendix 2). SpO 2 could be down-sampled to 1 sample per 20 seconds for improved artifact removal, without losing information about the SpO 2 dynamics.
Mean nocturnal SpO 2 was lower than mean daytime SpO 2 in rest. By taking into account the SpO 2 fluctuations over different days and nights, the results of this study generalize the findings of Soguel Schenkel et al, who only performed SpO 2 measurements during a single day and night that were up to five days apart [38]. Other studies used awake measurements of SpO 2 to predict nocturnal mean SpO 2 or CT90. They used the first 15-30 minutes of the nocturnal measurements to calculate the mean awake SpO 2 values [15,[39][40][41]. For these studies, it should be noted that slightly lower SpO 2 values were observed in the evening compared to the afternoon, on average a 0.8% difference, and awake SpO 2 values depend on the time of day when the measurements are performed. It is unlikely that this difference is solely due to the specified oximeter accuracy of ±2%, as no consistent bias has been reported for the WristOx 2 when measuring over a longer time period.
The observed patient-to-patient, day-to-day, and intraday differences can have implications for telemonitoring applications that are based on daily SpO 2 spot checks [12]. In several telemonitoring setups, alerts were raised when daily spot checks of SpO 2 dropped below a generic threshold value (eg, 90% for all patients) [12]. However, the high patient-to-patient differences point out the shortcomings of these generic thresholds. Other telemonitoring setups used personalized thresholds, but these were still fixed on one specific threshold value for every patient [12]. For these setups, alerts can be merely a consequence of natural day-to-day and intraday fluctuations, instead of being triggered by the onset of an exacerbation. A recent paper better dealt with the day-to-day fluctuations by suggesting a day-to-day decrease of more than 4% to alert for exacerbations, however, a more thorough examination of this method is needed [42]. Intraday SpO 2 values during rest ranged over 10.8% (SD 4.4) within one day. Moreover, statistical interpretation of weekly average of SpO 2 SD (1.6%, SD 0.6), which is a measure of intraday fluctuations, indicates that during 5% of the day in rest, SpO 2 fluctuates more than 3.2% (ie, 2 times SpO 2 SD) beyond the mean daytime-in-rest value. In comparison, previous studies only reported a 1%-2% decrease in SpO 2 spot checks around exacerbation onset compared to stable periods [43][44][45]. These natural intraday fluctuations can thus easily result in false alerts for exacerbations. Patients with lower SpO 2 values might experience a higher number of false alerts as, similar to healthy individuals [29], intraday fluctuations (ie, SpO 2 SD) increased with decreasing mean SpO 2 . Altogether, personalized alerts based on intelligent algorithms will be necessary to cope with all of these natural fluctuations in daytime SpO 2 in rest.
Preferably, these alerts should be based on continuous measurements over longer time periods, in contrast to the currently used daily spot checks, to account for the identified SpO 2 fluctuations and to exclude the potential effect of the ±2% oximeter accuracy. In addition, the slightly lower SpO 2 values in the evening compared to the afternoon show that the daily measurements should always be performed at the same moment of the day. A clearly defined measurement protocol, which is often not specified [12], can thus further attempt to limit the effect of natural SpO 2 fluctuations in these telemonitoring applications.
The night-to-night differences of nocturnal mean SpO 2 (ie, average weekly range of 2.0% over the different nights) led to highly varying CT90 values over the different nights of the measurement week (ie, average range of 28%). These high variations in CT90 resulted in 50% (10/20) of the included patients changing category between desaturator and nondesaturator, due to the fact that these occasional desaturators all had a mean SpO 2 value around the threshold of 90%. A similar finding has been reported by Lewis et al, who concluded that 35% of the included patients changed category over two consecutive nights of measurements [20]. Later studies then tackled this problem by performing measurements over two nights, categorizing a patient as a desaturator if desaturation occurred in at least one of both nights [5,[17][18][19]. However, the results of this study indicate that even two nights are insufficient to make a consistent separation between desaturators and nondesaturators, as 6 out of 10 occasional desaturators (60%) were desaturators or nondesaturators in the first two nights and only changed category afterward. Our results suggest that it might be impossible to consistently categorize COPD patients with a mean nocturnal SpO 2 value around 90% as desaturator or nondesaturator. Based on measurements over multiple nights, these patients could thus be referred to as occasional desaturators. Further research is needed to assess the clinical relevance of identifying these three different nocturnal desaturation profiles.
Some limitations should be taken into consideration when interpreting the results of this study. The main limitation, as is often the case in similar studies, is the small sample size. Consequently, no comparison could be made between mild, moderate, and severe hypoxemic patients. Nevertheless, the increasing SpO 2 SD with decreasing mean SpO 2 suggests an increase in SpO 2 fluctuations for more hypoxemic patients. No control group was included because this study aimed to perform continuous SpO 2 measurements in COPD patients for identification of SpO 2 fluctuations that could affect SpO 2 applications, rather than comparing SpO 2 between COPD patients and healthy controls. In addition, daily home routines can greatly differ between patients and healthy controls [46], impeding a proper comparison. As shown in Multimedia Appendix 1, daytime measurements were only seldom performed during full days due to the impracticalities of wearing the finger clip during activities that require manual actions (eg, during morning routines). This, in combination with the high amount of invalid data during MVPA, suggests that continuous SpO 2 measurements with a finger clip will have more profound limitations in a population that is more physically active than the target population of this study. Reliable wearable oximeters that do not require a finger clip could thus increase compliance. Moreover, the limited battery life and the inability of real-time data transmission of the used oximeter can further complicate the integration of continuous SpO 2 measurements into practice. Technological advances are thus needed to allow for long-term, continuous monitoring of SpO 2 . This study, however, mainly intended to show the potential of prolonged continuous measurements to identify SpO 2 fluctuations. Therefore, a certified wearable pulse oximeter with finger clip was preferred over more user-friendly, watch-type oximeters that have not yet been proven to be accurate. The resulting, more fragmented, daytime measurements were sufficient to identify large natural fluctuations occurring within one day (ie, SpO 2 in rest ranged over 10.8% [SD 4.4] within one day), confirming the a priori posed hypothesis. The limited battery life was addressed by providing the patients with multiple sensors to cover the 7-day monitoring period.
This study was the first to use wearable finger pulse oximeters for prolonged continuous SpO 2 measurements in COPD patients, as opposed to only performing spot checks or continuous measurements during one or two nights. These measurements showed that spot checks or one-or two-night measurements should be interpreted with caution, as the conclusions based on these measurements might change depending on the moment of the measurement. Measurements were performed at home during daily life routines of COPD patients, which provides a more natural SpO 2 profile compared to supervised measurements. By adding actigraphy measurements, the necessary contextual information could be gathered for more accurate analyses of the continuous SpO 2 measurements.
In conclusion, continuous SpO 2 measurements with wearable pulse oximeters identified significant SpO 2 fluctuations between and within multiple days and nights of patients with COPD. Continuous SpO 2 measurements during the daily home routine of patients with COPD generally had high amounts of valid data, except for motion artifacts during MVPA. The continuous measurements showed that mean nocturnal SpO 2 was lower than mean daytime SpO 2 in rest, and significant SpO 2 fluctuations occurred between and within multiple days and nights. The large fluctuations of daytime SpO 2 in rest indicate that clear measurement protocols and personalized alerts, based on intelligent algorithms, will be needed to increase the performance of telemonitoring applications that make use of daily SpO 2 spot checks. Lastly, it was shown that CT90 values can vary greatly from night to night in patients with a nocturnal mean SpO 2 around 90%, indicating that these patients cannot be consistently categorized as desaturators or nondesaturators. We recommend using wearable sensors for performing continuous SpO 2 measurements over longer time periods to determine the clinical relevance of the identified SpO 2 fluctuations.