: Longitudinal Physiological Data from a Wearable Device Identifies SARS-CoV-2 Infection and Symptoms and Predicts COVID-19 Diagnosis

: Background: Changes in autonomic nervous system function, characterized by heart rate variability (HRV), have been associated with and observed prior to the clinical identification of infection. We performed an evaluation of this metric collected by wearable devices, to identify and predict Coronavirus disease 2019 (COVID-19) and its related symptoms. Methods: Health care workers in the Mount Sinai Health System were prospectively followed in an ongoing observational study using the custom Warrior Watch Study App which was downloaded to their smartphones. Participants wore an Apple Watch for the duration of the study measuring HRV throughout the follow up period. Survey’s assessing infection and symptom related questions were obtained daily. Findings: Using a mixed-effect COSINOR model the mean amplitude of the circadian pattern of the standard deviation of the interbeat interval of normal sinus beats (SDNN), a HRV metric, differed between subjects with and without COVID-19 (p=0.006). The mean amplitude of this circadian pattern differed between individuals during the 7 days before and the 7 days after a COVID-19 diagnosis compared to this metric during uninfected time periods (p=0.01). Significant changes in the mean MESOR and amplitude of the circadian pattern of the SDNN was observed between the first day of reporting a COVID-19 related symptom compared to all other symptom free days (p=0.01). Interpretation: Longitudinally collected HRV metrics from a commonly worn commercial wearable device (Apple Watch) can identify the diagnosis of COVID-19 and COVID-19 related symptoms. Prior to the diagnosis of COVID-19 by nasal PCR, significant changes in HRV were observed demonstrating its predictive ability to identify COVID-19 infection. Funding


ABSTRACT:
Background: Changes in autonomic nervous system function, characterized by heart rate variability (HRV), have been associated with and observed prior to the clinical identification of infection.We performed an evaluation of this metric collected by wearable devices, to identify and predict Coronavirus disease 2019 (COVID-19) and its related symptoms.
Methods: Health care workers in the Mount Sinai Health System were prospectively followed in an ongoing observational study using the custom Warrior Watch Study App which was downloaded to their smartphones.Participants wore an Apple Watch for the duration of the study measuring HRV throughout the follow up period.Survey's assessing infection and symptom related questions were obtained daily.
Findings: Using a mixed-effect COSINOR model the mean amplitude of the circadian pattern of the standard deviation of the interbeat interval of normal sinus beats (SDNN), a HRV metric, differed between subjects with and without COVID-19 (p=0.006).The mean amplitude of this circadian pattern differed between individuals during the 7 days before and the 7 days after a COVID-19 diagnosis compared to this metric during uninfected time periods (p=0.01).Significant changes in the mean MESOR and amplitude of the circadian pattern of the SDNN was observed between the first day of reporting a COVID-19 related symptom compared to all other symptom free days (p=0.01).
Interpretation: Longitudinally collected HRV metrics from a commonly worn commercial wearable device (Apple Watch) can identify the diagnosis of COVID-19 and COVID-19 related symptoms.Prior to the diagnosis of COVID-19 by nasal PCR, significant changes in HRV were observed demonstrating its predictive ability to identify COVID-19 infection.
Funding: Support was provided by the Ehrenkranz Lab For Human Resilience, the BioMedical Engineering and Imaging Institute, The Hasso Plattner Institute for Digital Health at Mount Sinai, The Mount Sinai Clinical Intelligence Center and The Dr. Henry D. Janowitz Division of Gastroenterology.

INTRODUCTION
3][4] Health care workers (HCWs), characterized as any type of worker in a health care system, represent a vulnerable population with a threefold increased risk of infection compared to the general population. 5This increased risk of transmission is important in healthcare settings, where asymptomatic or presymptomatic HCWs can shed the virus contributing to transmission within healthcare facilities and their households. 6gital health technology offers an opportunity to address the limitations of traditional public health strategies aimed at curbing COVID-19 spread. 7Smart phone Apps are effective in using symptoms to identify those possibly infected with SARS-CoV-2, but they rely on ongoing participant compliance and self-reported symptoms. 8Wearable devices are commonly used for remote sensing and provide a means to objectively quantify physiological parameters including heart rate, sleep, activity and measures of autonomic nervous system (ANS) function (e.g., heart rate variability [HRV]). 9The addition of physiological data from wearable devices to symptom tracking Apps has been shown to increase the ability to identify those infected with SARS-CoV-2. 10 HRV is a physiological metric providing insight into the interplay between the parasympathetic and sympathetic nervous system which modulate cardiac contractility and cause variability in the beat-to-beat intervals. 113][14] Changes in this circadian pattern can be leveraged to identify different physiological states.Several studies have demonstrated that lower HRV, indicating increased sympathetic balance, is a reliable predictor of infection onset. 15,16However, HRV and its dynamic changes over time have not been evaluated as a marker or predictor of COVID-19.In response to the COVID-19 pandemic we launched The Warrior Watch Study™, employing a novel smartphone App to remotely enroll and monitor HCWs throughout the Mount Sinai Health System in New York City, a site of initial case surge.This digital platform enables remote survey delivery to Apple iPhones and passive collection of Apple Watch data, including HRV.The aim of this study is to determine if SARS-CoV-2 infections can be identified and predicted prior to a positive test result using the longitudinal changes in HRV metrics derived from the Apple Watch.

Study Design
The primary aim of the study was to determine whether changes in HRV can differentiate participants infected or not infected with SARS-CoV-  1).Participants carried out their normal activities throughout the study and were instructed to wear the Apple Watch for a minimum duration of 8 hours per day.
Wearable Monitoring Device and Autonomic Nervous System Assessment HRV was measured via the Apple Watch Series 4 or 5, which are commercially available wearable devices.Participants wore the device on the wrist and connected it via Bluetooth to their iPhone.The Watch is equipped with an enhanced photoplethysmogram (PPG) optical heart sensor that combines a green LED light paired with a light sensitive photodiode generating time series peaks that correlate with the magnitude of change in the green light generated from each heartbeat. 17Data are filtered for ectopic beats and artifact.The time difference between heartbeats is classified as the Interbeat Interval (IBI) from which HRV is calculated.The Apple Watch and the Apple Health app automatically calculate HRV using the standard deviation of the IBI of normal sinus beats (SDNN), measured in milliseconds (ms).This time domain index reflects both sympathetic and parasympathetic nervous system activity and is calculated by the Apple Watch during ultra-short-term recording periods of approximately 60 seconds. 11The Apple Watch generates several HRV measurements throughout a 24-hour period.HRV metrics are stored in a locally encrypted database accessible through the iPhone Health app which is retrieved through our custom Warrior Watch App.Data is transferred from the iPhone and Apple Watch upon completion of the e-consent and any survey in the App.Wearable data is stored locally allowing retrieval during the days when surveys are not completed by participants.

Statistical Analysis Heart Rate Variability Modelling
The HRV data collected through the Apple Watch was characterized by a circadian pattern, a sparse sampling over a 24-hour period, and a non-uniform timing across days and participants.These characteristics bias easily derived features including mean, maximum and minimum creating the need to derive methods that model the circadian rhythm of HRV.A COSINOR model was used to model daily circadian rhythm over a 24 hour period with the non-linear function Y where τ is the period ( =24h), M is the Midline Statistic of Rhythm (MESOR), a rhythmadjusted mean, A is the amplitude, a measure of half the extent of variation within a day and Φ is the Acrophase, a measure of the time of overall high values recurring in each day (Supplementary Figure 1).This non-linear model with 3 parameters has the advantage of being easily transformed into a linear model by recoding time (t) into two new variables x and z as  = sin(2t/),  = sin(2t/).HRV can then be written as We took advantage of the longitudinal structure of the data to identify a participant specific daily pattern and then measured departures from this pattern as a function of COVID-19 diagnosis or other relevant covariates.In order to do so we used a mixed-effect COSINOR model, where the HRV measure of participant i at time t can be written as HRVit = (M+.xit+ .zit ) +   .  + ei(t), ei(t)~N(0,s), and where M,  and  are the population parameters (fixed-effects) and i is a vector of random effects and assumed to follow a multivariate normal distribution  i~N(0,Σ).In this context the introduction of random effects intrinsically model the correlation due to the longitudinal sampling.To measure the impact of any covariate C on the participants' daily curve, we can introduce such covariates as fixed-effects as its interactions with x and z: HRVit = M+oCi+( + 2Ci).xit+ ( +  3Ci.)zit +   .  + ei(t) [equation 3].Model parameters and the standard errors of equation 3 can be estimated via maximum likelihood or reweighted least squares (REWL) and hypothesis testing can be carried out for any comparison that can be written as a linear function of  ′ ,    parameters.
However, to test if the COSINOR curve, defined by the non-linear parameters M, A and  in equation 1 differs between the populations defined by the covariate C, we proposed the following bootstrapping procedure where for each resampling iteration we: (1) Fit a linear mixed-effect model using REWL; (2) Estimated the marginal means obtaining the linear parameters for each group defined by covariate C; (3) Used the inverse relationship to estimate marginal means M, A and  for each group defined by C; and (4) Defined the bootstrapping statistics as the pairwise differences of M, A and  between groups defined by C. For such iterations, the confidence intervals for the nonlinear parameter was defined using standard bootstrap techniques, as well deriving the p-values for the differences of each non-linear parameter between groups defined by Ci.
Age and sex were included as a covariate in HRV analyses and admitted invariant and time-variant covariates.

Association and Prediction of COVID-19 Diagnosis and Symptoms
The relationship between a COVID-19 diagnosis and change in HRV curves were evaluated.To test this association, we defined the time variant covariate Cit for participant i at time t as: HRV metrics for the 14

Role of the Funding Source
The study sponsors played no role in the study design, data collection, analysis, writing or decision for publication.

RESULTS
Two hundred and ninety-seven participants were enrolled between April 29 th and September 29 th , 2020, when data was censored for analysis (  1d-e).

Identification and Prediction of COVID-19 Symptoms
Symptoms were frequently reported during the follow up period with the greatest number of participants reporting feeling tired or weak (n=87), followed by headaches (n= 82) and sore throat (n=60) (Table 4).Evaluating the days when participants experienced symptoms, we found that loss of smell or taste were reported the most with a mean of 138 days.This was followed by feeling tired or weak, reported a mean of 25 days and runny nose, reported a mean of 19.5 days (Figure 2).The mean MESOR, acrophase and amplitude observed in the circadian SDNN pattern in participants on the first day a symptom and on all other days of follow up are reported in  3a-c).
The mean MESOR, acrophase and amplitude observed in the circadian SDNN pattern in participants on the day before symptoms develop, on the first day of the symptom, on the day following the first day of the symptom and on all other days are reported in Excluded from the analysis was the day prior to symptom development and all other days.There were no other significant differences between the MESOR, amplitude, and acrophase of SDNNs circadian rhythm when comparing participants on the day before symptoms develop, on the first day of the symptom, on the day following the first day of the symptom and on all other days (Figure 3d-e).

DISCUSSION
In this prospective study, longitudinally evaluated HRV metrics were found to be associated with a positive SARS-CoV-2 diagnosis and COVID-19 symptoms.Significant changes in these metrics were observed 7 days prior to the diagnosis of COVID-19.To the best of our knowledge this is the first study to demonstrate that physiological metrics derived from a commonly worn wearable device (Apple Watch) can identify and predict SARS-CoV-2 infection prior to diagnosis with a SARS-CoV-2 nasal PCR swab.These preliminary results identify a novel easily measured physiological metric which may aid in the tracking and identification of SARS-CoV-2 infections.
9][20] However, the variable incubation period, high percentage of asymptomatic carriers, and infectivity during the pre-symptomatic period of COVID-19 have made containment challenging. 21This has further limited the utility of systematic screening technologies reliant on vital sign assessment or self-reporting of symptoms. 7vances in digital health provide a unique opportunity to enhance disease containment.Wearable devices are commonly used and well accepted for health monitoring. 9,22Commercially available devices are able to continually collect several physiological parameters.Unlike App-based platforms, wearable devices have the advantage of not requiring users to actively participate aside from regular use of the device.Prior to the COVID-19 pandemic, population level data from the Fitbit wearable device demonstrated effectiveness at real-time geographic surveillance of influenza-like illnesses through the assessment of physiological parameters. 23This concept was recently expanded during the COVID-19 pandemic by Quer and colleagues who demonstrated that the combination of symptom-based data with resting heart rate and sleep data from wearable devices was superior to relying on symptom-based data alone to identify COVID-19 infections. 10V has been shown to be altered during illnesses with several small studies demonstrating changes in HRV associated with and predictive of the development of infection.24 Ahmad and colleagues followed 21 subjects undergoing bone marrow transplant finding a significant reduction in root mean square successive difference metrics prior to the clinical diagnosis of infection.Furthermore, wavelet HRV was noted to decrease by 25% on average 35 hours prior to a diagnosis of sepsis in 14 patients.16 In another study in 100 infants, significant HRV changes were noted 3-4 days preceding sepsis or systemic inflammatory response syndrome with the largest increase being seen 24 hours prior to development.15 Building on these observations demonstrating that ANS changes accompany or precede infection, our team launched the Warrior Watch Study.
We demonstrated that significant changes in the circadian pattern of HRV, specifically SDNN's amplitude, was associated with a positive COVID-19 diagnosis.Interestingly, when we compared these changes over the seven days preceding the diagnosis of COVID-19 we continued to see significant alterations in amplitude when compared to individuals without COVID-19.This demonstrates the predictive ability of this metric to identify infection.Interestingly when we follow individuals 7-14 days after diagnosis with COVID-19, we find that the circadian HRV pattern starts to normalize and is no longer statistically different from an uninfected pattern.As an exploratory analysis we evaluated how HRV was impacted by symptoms associated with a COVID-19 diagnosis, since individuals may not be tested despite symptoms.We found significant changes in the amplitude of the circadian HRV pattern on the first day of symptoms, with a trend toward statistical significance on the day before and after symptoms are reported.Taken together, these findings highlight the possible use of HRV collected via wearable devices to identify and predict COVID-19 infections.
There are several limitations to our study.First, there was a small number of participants who were diagnosed with COVID-19 in our cohort limiting our ability to determine how predictive HRV can be of infection.However, these preliminary findings support the further evaluation of HRV as a metric to identify and predict COVID- where the linear coefficients ,  of the linear model in equation 2 are related to the non-linear parameters of the non-linear model in equation 1 by  = ()  = −().One can estimate the linear parameters ,  and then obtain the A and  as: days following the time of first positive SARS-CoV-2 nasal PCR test were used to define the positive SARS-CoV-2 infection window.To evaluate the predictive ability of changes in HRV prior to a COVID-19 diagnosis and to explore its changes during the infection period, the time variant covariate was used to characterize the following 4 groups: healthy uninfected individuals [t<t0-7], 7 days before COVID-19 diagnosis [t≥t0-7, t<t0], the first 7 days post COVID-19 diagnosis [t0≤t<t0+7] and the 7-14 days post diagnosis [t0+7≤t<t0+14].To determine the association between COVID-19 symptoms and changes in HRV metrics, we defined being symptomatic as the 1 st day of a reported symptom and compared this to all other days.To evaluate the predictive ability of HRV to identify upcoming symptom days and to explore its changes over time the time variant covariate was used to characterize the following 4 groups: healthy asymptomatic individuals for t<t0-1, one day before COVID-19 symptoms [t≥t0-1, t<t0], the first day of COVID-19 symptoms [t0≤t<t0+1] and one day post COVID-19 symptom development [t0+1≤t<t0+2].

Figure 2 .
Figure 2. Number of symptom days per participant when evaluating days when participants reported symptoms

Figure 3 :
Figure 3: Relationship between HRV circadian rhythm and symptom onset.Timeline (A) illustrates timing of symptom onset, HRV profiles of the first-symptom day (red) were compared to all other days (green).Daily HRV rhythm (B) on day of first symptom (red) and non/late-symptom (green) days, time (hours) is indicated by the xaxis and HRV (ms) is indicated by the y-axis.Plots (C) showing Mean and 95%CI for the parameters defining the circadian rhythm: Acrophase, Amplitude and MESOR on first symptom (red) and non/late-symptom (green) days.Daily HRV pattern (D) for non/late-symptomatic days (green), the day before first symptom (red), day of first symptom (orange) and day after first symptom (light green), time (hours) is indicated by the x-axis while HRV (ms) is indicated by the y-axis.Mean and 95% CI for the Acrophase, Amplitude and MESOR of the HRV measured on non/late-symptomatic days (green), the day before first symptom (red), day of first symptom (orange) and day after first symptom (light green), +p<0.1;*p<0.05;**p<0.01;***p<0.001;ns, not significant

Table 1 )
. The median age at enrollment was 36 years with 69% of participants being women.Twenty participants -129) were obtained per participant.Study compliance over the follow up period, defined as participants answering over 50% of daily surveys, was 70.4%.Identification and Prediction of COVID-19 DiagnosisThirteen participants reported a positive SARS-CoV-2 nasal PCR during the follow up period.The mean MESOR, acrophase and amplitude of the circadian SDNN pattern in participants diagnosed with and without COVID-19 are described in Table2.A significant difference in the circadian pattern of SDNN was observed in participants diagnosed with COVID-19 compared to those without COVID-19.There was a significant difference (p=0.006) between the mean amplitude of SDNNs circadian reported having a positive SARS-CoV-2 nasal PCR test prior to enrollment, while 28 participants reported having a positive blood antibody test prior to joining the study.The median duration of follow up was 42 days (range 0-152 days).A median of 28 HRV samples (range 1pattern in those with (1.23 ms, 95% CI -1.94-3.11)andwithoutCOVID-19(5.30 ms, 95% CI 4.97-5.65).No difference was observed between the MESOR (p=0.46) or acrophase (p=0.80) in these two infection states (Figure1a-c).The mean MESOR, acrophase and amplitude of the circadian SDNN pattern for those without COVID-19, those during the 7 days prior to a COVID-19 diagnosis, participants during the 7 days after a COVID-19 diagnosis and those during the 7-14 days after a COVID-19 diagnosis are described in Table3.Significant changes in the circadian pattern of SDNN were observed in participants during the 7 days prior and the 7 days after a diagnosis of COVID-19 when compared to uninfected participants.There was a significant difference between the amplitude of the SDNN circadian rhythm between

Table 5 .
There was a significant difference in the circadian SDNN pattern between participants on the first day a symptom was reported compared to all other days of follow up.Specifically, compared to all other days (43.48 ms, 95% CI 41.77-45.27).Similarly, there was a significant difference (p=0.01) between the mean amplitude of SDNNs circadian pattern on the first day of symptoms (2.58 ms, 95% CI 0.26-5.00)compared to all other days (5.30 ms, 95% CI 4.95-5.66)(Figure

Table 6 .
Significant changes in the circadian pattern of SDNN were observed, comparing the amplitude of the SDNN circadian pattern between participants during the first day of the symptom (3.07 ms, 95% CI 0.88-5.22)with the one day after the first symptom was reported (5.47 ms, 95% CI 3.16-7.76;p=0.56).
4.99-5.66;p=0.056).Again, excluded from the analysis was the first day of the symptom and the day after the first symptomatic day.Additionally, there was trend toward significance when

Table 1 .
19and warrant further study.An additional limitation is the sporadic collection of HRV by the Apple Watch.While our statistical was able to account for this a denser dataset would allow for expanded evaluation of the relationship between this metric and infections/symptoms.The Apple Watch also only provides HRV in one time-domain (SDNN), limiting assessment of the relationship between other HRV parameters with COVID-19 outcomes.Lastly, an additional limitation is that we relied on self-reported data in this study, precluding independent verification of COVID-19 diagnosis.Baseline demographics of participants at enrollment.
PCR, polymerase chain reaction; SD, standard deviation *Clinical trainee defined as a resident or fellow; clinical non-trainee defined as HCWs reporting at least one patient facing day during follow up, exclusive of resident and fellows; non-clinical staff defined as a HCW who did not report a patient facing day during follow up.

Table 2 .
HRV parameters in participants with and without COVID-19 diagnosed based on SARS-CoV-2 nasal PCR swabs.

Table 3 .
Comparison of HRV parameters based on the time period before and after diagnosis.

Table 4 .
Number of participants reporting each symptom.

Table 5 .
HRV parameters on the first day of reported symptoms compared to all other symptom free days.

Table 6 .
Comparison of HRV parameters based on symptom state and the time-period before and after the first day of reported symptoms.