key: cord-1053942-2bz9u8k0
authors: Marinsek, N.; Shapiro, A.; Clay, I.; Bradshaw, B.; Ramirez, E.; Min, J.; Trister, A.; Wang, Y.; Althoff, T.; Foschini, L.
title: Measuring COVID-19 and Influenza in the Real World via Person-Generated Health Data
date: 2020-05-30
journal: nan
DOI: 10.1101/2020.05.28.20115964
sha: 84284719be820f0924f88affce9bb040ec296c0f
doc_id: 1053942
cord_uid: 2bz9u8k0

Background: Since the beginning of the COVID-19 pandemic data from smartphones and connected sensors has been used to learn about symptoms presentation and management outside the clinic walls. However, reports on the validity of such data are still sparse, especially when it comes to symptom progression and relevance of wearable sensors. Objective: To understand the relevance of Person-GeneratedHealth Data (PGHD) as a means for early detection, monitoring and management of COVID-19 in everyday life. This includes quantifying prevalence and progression of symptoms from self-reports as well as changes in activity and physiological parameters continuously measured from wearable sensors, and contextualizing findings for COVID-19 patients with those from cohorts of flu patients. Design, Setting, and Participants: Retrospective digital cohort study of individuals with a self-reported positive SARS-CoV-2 or influenza test followed over the period 2019-12-02 to2020-04-27. Three cohorts were derived: Patients who self-reported being diagnosed with flu prior to the SARS-CoV-2 pandemic (N=6270, of which 1226 also contributed sensorPGHD); Patients who reported being diagnosed with flu during the SARS-CoV-2 pandemic (N=426, of which 85 also shared sensor PGHD); and patients who reported being diagnosed withCOVID-19 (N=230, of which sensor PGHD was available for 41).The cohorts were derived from a large-scale digital participatory surveillance study designed to track Influenza-like Illness(ILI) incidence and burden over time. Exposures: Self-reported demographic data, comorbidities, and symptoms experienced during a diagnosed ILI episode, including SARS-CoV-2.Physiological and behavioral parameters measured daily from commercial wearable sensors, includingResting Heart Rate (RHR), total step count, and nightly sleep hours. Main Outcomes and Measures: We investigated the percent-age of individuals experiencing symptoms of a given type (e.g.shortness of breath) across demographic groups and over time. We examined illness duration, and care seeking behavior, and how RHR, step count, and nightly sleep hours deviated from expected behavior on healthy days over the course of the infection episode. Results: Self-reported symptoms of COVID-19 present differently from flu. COVID-19 cases tended to last longer than flu(median of 12 vs. 9 days), are uniquely characterized by chest pain/pressure, shortness of breath, and anosmia. The fraction of elevated RHR measurements collected daily from commercial wearable devices rise significantly in the 2 days surrounding ILI symptoms onset, but does not appear to do so in a way specific to COVID-19. Steps lost due to COVID-19 persists for longer. Conclusion and Relevance: PGHD can be a valid source of longitudinal real world data to detect and monitor COVID-19-related symptoms and behaviors at population scale. PGHD may provide continuous, near realtime feedback to intervention effectiveness that otherwise requires waiting for symptoms to develop into contacts with the healthcare system. It has also the potential to increase pre-test probability of other downstream diagnostics. To effectively leverage PGHD for participatory surveillance it is crucial to invest in the creation of trusted, long-term communication channels with individuals through whichdata can be efficiently collected, consented, and contextualized,while protecting the privacy of individuals and ultimately facilitating the transition in and out of care.

The emergence of the novel SARS-CoV-2 (COVID- 19) and subsequent rapidly expanding pandemic has created significant gaps in our understanding of the prevalence of symptoms among individuals with COVID-19. Multi-disciplinary teams of physicians, data scientists, clinical informaticians, epidemiologists and many others around the world have engaged in using real-world data collected at point of care to help answer key questions around management of COVID-19 patients (see (1) for a US-centric overview of initiatives and open questions). At the same time, data from individuals in the context of participatory syndromic surveillance for COVID-19 (2) (3) (4) are being collected via smartphone apps around the world to perform hotspot detection and show promise in understanding symptom presentation and prevalence outside the clinic walls (5) . In addition to self-report, recent literature suggests that data from commercial sensors may be used for large scale surveillance of influenza outbreak, based on the fact that physiologic measures provided by the sensors (e.g., RHR) (6) (7) (8) and temperature (9) change in the presence of an infection. Several efforts are currently underway exploring the potential of wearable technology in support of syndromic surveillance aimed at the COVID-19 pandemic (10-12), however, no data yet exists on quantifying the link between COVID-19 disease and changes in physiological and behavioral parameters over the course of the disease. While specific clinical symptoms such as shortness of breath, fever, and dry cough have become hallmark characteristics of the disease (13) , symptoms presentation and patient behavior outside the clinic walls has received little attention. Lack of testing means we cannot build a canonical symptom presentation, which significantly undermines our ability to track, predict and control disease progression and manage critical care. Using a large-scale digital participatory surveillance study designed for the purpose of monitoring ILI over the 2019-2020 influenza season, we present data collected from a cohort of individuals who have self-reported being diagnosed by a medical provider with flu or COVID-19. A subset of this cohort provided daily physiological and behavioral data derived via wearable activity monitors (daily RHR, daily step count, and nightly sleep hours) allowing for the explicit linkage of illness onset and continuously measured physiological and behavioral parameters. Our contributions are twofold. First, we present data on the progression of COVID-19 symptoms in everyday life and contextualize those by comparing them with seasonal influenza. Second, we show that changes in physiological signals such as RHR are associated with the onset of symptoms, though they may not be specific to the type of ILI. To our knowledge, this is the first study that looks at longitudinal symptom reports of COVID-19 patients. It is also the first study presenting symptom reports linked to wearable data at the individual level for ILI (flu or COVID- 19) , enabling temporal alignment of symptom reports with correspondent changes in wearable data.

Evidation Health currently supports a mobile consumer application called Achievement (14) that rewards members based on completing health-related behaviors and participating in research by sharing Person-Generated Health Data (PGHD). Achievement members can connect activity trackers and health applications and authorize Achievement to automatically and continuously ingest connected data streams (15) . Achievement has an active user base of 3.7M individuals that are economically, demographically, and geographically diverse, and enables rapid (16) recruitment of participants. Since 2017, Achievement has been used to run a participatory ILI surveillance program, examining annual waves of Influenza virus infections (17) . The 2019-2020 version of the program consists of sending a weekly one-click survey to all Achievement members that asks if the individual experienced any flu-like symptoms in the past 7 days. Individuals who answer "yes" are immediately sent to a questionnaire, which contains questions about the dates of illness onset and/or recovery, symptom experiences in the previous 7 days, healthcare interactions and outcomes, medications, and household characteristics. On 2020-03-30, the questionnaire was updated to include questions that specifically address COVID- 19 , including questions about COVID-19 diagnosis, testing, and social distancing measures, as well as an expanded list of symptoms (consisting of shortness of breath, chest pain, and anosmia). The contents of the original and updated question-naires are included in Supplementary Note 2 in the Appendix. Since participants were sent a one-click survey every week, participants could submit multiple survey responses for the same ILI event. Responses to the original and updated surveys, collected between 2019-12-02 and 2020-04-27, comprise the initial survey dataset and include a total of 194,401 responses from 85,558 unique participants. In addition to agreeing to participate and sharing their survey responses, participants agreed to share their activity data from connected wearable sensors. The sensors data analyzed in this project consisted of minute-by-minute step counts, RHR recordings, and sleep states from 2019-11-01 to 2020-05-13 for the subset of participants with Fitbit sensors connected to the Achievement platform.

Survey Preparation. The pipeline for preparing the surveys for analysis is illustrated in Supplementary Figure S1 (a) and summarized here. Initial survey cleaning consisted of excluding all survey responses with self-reported illness onset dates or recovery dates that occurred 30 or more days before the survey completion date, excluding surveys with invalid illness onset and/or recovery dates (defined as dates occurring after the survey date or responses in which the illness recovery date occurred before the illness onset date), removing multiple survey responses from the same participant on the same day, and de-duplicating identical sets of survey responses. This filtering process reduced the dataset from 194,401 survey responses from 85,558 unique participants to 146,133 responses from 71,556 unique participants. Since participants could submit multiple survey responses for the same ILI event, distinct ILI events were inferred by merging survey responses from the same participant when the dates encompassing self-reported illness onset through recovery overlapped or were separated by no more than two days. The reconciliation process for merging individual question responses is described in Supplementary Note 3 in the Appendix. After excluding participants with five or more ILI events or multiple diagnosed ILI events (to remove participants with possible erroneous or fraudulent responses), the analysis set was reduced to a subset of 6,926 ILI events with self-reported confirmed diagnosis, each corresponding to a different participant. The analysis set was supplemented with demographic information (if available) from a different general health survey that included information about gender, age, body mass index (BMI), ethnicity, race, and pre-existing health conditions. Since the general health surveys were completed at different times as the ILI survey, there may be slight discrepancies in time-variant demographic data, such as age, BMI, and health conditions. Cohort Definition. All participants included in the analysis self-reported seeking medical care and being diagnosed with flu and/or COVID-19 by a healthcare provider (N=6,926). As shown in Figure 1 , these participants were divided into Statistical Testing. A two-step statistical testing procedure was used to test for differences in demographics, healthcare care-seeking behavior, medical outcomes, and symptoms among the three cohorts. First, for each sub-analysis (i.e., demographics, medical care-seeking, and symptom prevalence), a series of chi-squared tests of independence were performed to test for an association between the three cohorts and the different possible outcomes for each category. A Bonferroni correction was applied to adjust for running multiple chi-squared tests in each sub-analysis. Second, follow-up two-proportion z-tests were performed to test for differences in proportions for each outcome and each pair of cohorts. These follow-up tests were only performed for the categories with significant cohort differences as determined by the chi-squared tests. . Sensitivity analysis on the valid day thresholds was conducted and results did not change significantly when removing the requirement of having 10% of valid days for each day of week, or lowering the percentage of individual valid days to as low as 30%.

The pipeline for preparing the wearable data for analysis is illustrated in Supplementary Figure S1 (b).

Similarly to previous work (8), we examined the fraction of each cohort with elevated RHR in the days preceding and following ILI onset. First, days with no RHR recordings were imputed in order to ensure that the cohorts were the same across days of interest. Imputed RHR values were generated from predictions of a mixed effects regression model that was fit to all participant-days that RHR was recorded. The model specified fixed-effects for the week of the year to control for time of year effects (more specifically, this consisted of three terms for the 1st, 2nd, and 3rd expansions of an ordinal variable for week of flu season), a categorical fixed effect for the day of the week to account for differences in activity patterns by day of week, a fixedeffect for the average activity level in the participants' state of residence to control for different state-wide shelter-in-place and social distancing measures, and a random-intercept for each participant's baseline activity level to control for indi-vidual differences in activity levels. The model was fit to all participant-days with a RHR recording using the lme4 package for R (18) . Note that the imputed values were used only to fill days when RHR was not recorded, on all other days the observed value was used.

Next, in order to account for individual differences in RHR when defining thresholds for elevated RHR, RHR values were converted to z-scores using each participant's RHR mean and standard deviation across all days. The fraction of each cohort with elevated RHR was computed for the days surrounding the ILI event, defined as 10 days prior to 20 days after ILI onset. Elevated RHR was defined as being greater than 1 standard deviation above the participant's mean RHR. Two-proportion z-tests were performed to answer the following two questions: 1. Does a greater fraction of the COVID-19 cohort have elevated RHR in the days surrounding ILI onset (defined as Days -2 to 2 1 ) compared to baseline days prior to ILI onset (defined as Days -10 to -5 2 ) and 2. Does the fraction of participants with elevated RHR surrounding ILI onset differ between COVID-19 and Flu cohorts?

Events. In order to characterize daily changes associated with COVID-19 and flu events, we measured deviations from typical healthy measurements (RHR, step count, sleep hours) that occurred while participants were ill. We used a model on symptom-free days (conservatively assumed all days excluding the range 10 days before symptoms onset and within 20 days after symptoms onset) to generate individualized estimates of daily measurements that would have been recorded in the counter-factual scenario that the participant did not fall ill, and then computed the excess, defined as the difference (observed -estimated), on the days surrounding symptoms onset, and finally report the excess as a measure of deviations from expected typical measurements. The symptomfree day model was a mixed effects regression model with the same specification as what was used to impute missing RHR values in the previous analysis. The key difference in this analysis was that, in order to generate estimates based only on assumed symptom-free days, we excluded all data within 10 days before symptoms onset and within 20 days after symptoms onset when fitting the model. In order to visualize the time course of behavioral changes during COVID-19 and flu events, for each cohort we fit generalized additive mixed models with spline smoothing functions and random intercepts to the daily excess time series using the mgcv package for R (20) . This procedure was performed three separate times, once for each of the channels considered: daily total step counts, daily RHR, and total daily sleep minutes. 1 We conservatively allow 2 days for physiological changes before any symptom is reported, and up to 2 days after the onset of symptoms as the time horizon within which actions can be taken that would not be taken otherwise, without information from the wearable device) 2 Median time from exposure to development of symptoms is 5 days for COVID-19, (19) 

In the following section, we compare the presentation of diagnosed COVID-19 cases (N=230) to two groups of diagnosed flu cases: Non-COVID-19 Flu cases (N=426), which occurred in the same time frame as the COVID-19 cases, and Pre-COVID-19 Flu (N=6270), which occurred earlier in the 2019-2020 flu season before the outbreak of COVID-19.

Comparison of Demographic Trends. A demographic summary of the three cohorts is provided in Table 1 . A chi-squared test of independence was performed for each demographic category to test for significant differences across the full cohorts. Age-group and race were the only demographics that differed significantly among the cohorts after applying a Bonferroni correction to adjust for performing five comparisons (age group: p=.008, race: p<.001). Compared to the Pre-COVID-19 Flu cohort, the COVID-19 cohort was less likely to be White/Caucasian (63.9% vs. 70.0%) and more likely to be Asian or Pacific Islander (9.6% vs. 4.6%) (follow-up two-proportion z-tests, p=.047 and p<.001, respectively). The COVID-19 cohort was also more likely to prefer not to report race than both the Non-COVID-19 Flu and Pre-COVID-19 Flu cohorts (p=.018 and p<.001, respectively). In follow-up comparisons, the proportion of the COVID-19 cohort belonging to each age group was not significantly different than the Non-COVID-19 Flu nor the Pre-COVID-19 Flu cohorts, however a greater proportion of the Non-COVID-19 Flu cohort was aged 55 or older compared to the Pre-COVID-19 Flu cohort (7.5% vs. 3.8%, twoproportion z-test, p=.001). The demographics of the analyzed cohorts differs from what described in the literature for medically attended ILI events for the general US population (21) . The sample should be reweighed (22) before meaningful comparisons can be made with a target population with different demographics characteristics.

Care-seeking Behavior. Although all patients had to report seeking medical care and being diagnosed with either Flu or COVID-19 to be included in the analyses, locations of medical care, hospitalization rates, and medication prescription rates differed significantly across the three cohorts (chi-squared tests of independence with a Bonferroni correction, all p<.001), and are summarized in Table 2 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. 

In addition to differences in medical care-seeking and outcomes, we also observed differences in self-reported symptoms between COVID-19 and the two cohorts of flu patients. A summary of symptom prevalence across cohorts is reported in Table 3 . The most common symptoms across all ILI groups included cough, headache, body muscle ache, fatigue, and fever. The prevalence of symptoms was significantly different across the three cohorts (chi-squared test of independence, p<.001), confirming emerging literature and anecdotal reports (4, 10-13). All follow-up pairwise symptom comparisons were tested with two-proportion z-tests and a Bonferroni correction was applied for performing 33 tests. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. teristic of COVID-19 and have been reported to differentiate COVID-19 from seasonal flu, although it's important to note that, with the exception of cough, these symptoms are not necessarily common COVID-19 symptoms. For example, in the COVID-19 cohort, reports of chest pain were as common as reports of nasal congestion and sore throat, and reports of anosmia were as common as sneezing.

Compared to the Pre-COVID-19 Flu cohort, the COVID-19 cohort was significantly less likely to report experiencing body muscle ache, fever or feeling feverish, nasal congestion or runny nose, sneezing, chills or shivering, and sweats (all p<.001). The prevalence of several symptoms (i.e., shortness of breath, anosmia, and chest pain) could not be compared between the COVID-19 and Pre-COVID-19 Flu cohorts because they were not included in the original survey. Together, these findings are in line with recent work suggesting that the presentation of COVID-19 differs from other ILIs (2, 23), in particular with regard to anosmia, shortness of breath, coughing, fatigue, and muscle aches.

With the exception of headache, all symptoms were significantly less prevalent in the Non-COVID-19 Flu cohort relative to the Pre-COVID-19 Flu cohort. One possible reason for the difference in symptom presentations in the two flu cohorts is that the 2019-2020 flu season consisted of two waves of different flu strains: strain B (Victoria lineage) appeared earlier on and was followed by strain A (H1NI-pdm09) (24). According to the CDC, vaccines for the 2019-2020 season were well-matched against circulating strain A but not as well-matched against strain B (25) , which could account for the more mild symptom presentation in the more recent flu cases in the Non-COVID- 19 cohorts, we examined the prevalence of co-occurring sets of symptoms for the COVID-19 and Non-COVID-19 Flu cohorts ( Figure 2 ). The Pre-COVID-19 Flu cohort was excluded from this analysis since only a subset of symptoms were available for this cohort. To reduce the complexity of the possible symptom sets, we included only the 5 most prevalent symptoms in each cohort, which resulted in a reduced set of 7 symptoms: cough, headache, fever, fatigue, body muscle ache, chills or shivering, and shortness of breath. The two most common symptom sets consisted of all symptoms, which was predominated by COVID-19 cases, and all symptoms except for shortness of breath, which was predominated by Non-COVID-19 Flu cases. The other symptom sets that were more indicative of COVID-19 than Non-COVID-19 Flu included the symptom pair of shortness of breath and cough, and all symptoms other than chills or shivering.

Patients reported the dates of illness onset and illness recovery, which were used to compute the duration of each ILI event in days ( Figure 3 ). Duration of illness for COVID-19 cases tended to be longer than flu cases; COVID-19 illnesses lasted a median of 12 days, compared to 9 days for the Non-COVID-19 Flu cases and 7 days for the Pre-COVID-19 Flu cases. Symptom prevalence across each cohort for each day after illness onset is illustrated in Figure 4 and the days of peak symptom occurrence for each cohort are reported in Table  3 . In general, day-by-day symptom prevalence peaks later for the COVID-19 cases compared to the two groups of flu cases. With the exception of shortness of breath for Non-COVID-19 Flu cohort, all symptoms peak 2-3 days after illness onset in both flu cohorts. In contrast, COVID-19 symptoms peak 3-7 days after illness onset, with most symptoms peaking 4-5 days after illness onset. Some of the latest peaking symptoms are those that are most tightly associated with 6 | bioRχiv Luca Foschini et al. | covid-baseline . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. Figure 5 . Elevated RHRs were defined as RHRs that surpassed participants' mean RHR levels by 0, 0.5, or 1 standard deviation. In both the COVID-19 and Pre-COVID-19 Flu cohorts, we observe a spike in the prevalence of elevated RHR in the first few days after illness onset. We ran two-proportion z-tests to test 1) whether a greater proportion of COVID-19 patients had elevated RHRs around illness onset compared to before illness onset and 2) whether the prevalence of elevated RHR around illness onset differed between COVID-19 and Flu cases. The percent of COVID-19 patients with elevated RHRs (defined as a RHR >1 standard deviation above their personal mean RHR) was higher around the onset of COVID-19 (from Days -2 to 2) compared to the baseline period 5-10 days before illness onset (25% vs. 13%, p=0.005). The percent of the COVID-19 cohort with elevated RHR around illness onset was higher than that of the Non-COVID-19 cohort (16%, p=0.026), but did not differ from that of the Pre-COVID-19 cohort (22%, p=0.454).

Symptoms reports analyzed in previous sections can be understood as an assessment of the deviation from how the participant normally feels. In this section we present an analysis of wearable data to derive an analogous of the concept of deviations from the norm during the ILI event that symptoms reports capture. The objective impact of ILI (COVID-19 or flu) was quantified as the excess observed in daily steps lost, additional minutes of sleep, and increased RHR as compared to the expected measurements recorded in the counter-factual scenario where symptoms were not present for that day and individual, as estimated from a model fit only to symptom-free days. We present aggregate time series of these excesses for each of the . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Figure 6 . Results are shown in Figure 6 . Reductions in daily step counts are more marked and prolonged for COVID-19 as compares to Non-COVID-19 Flu and Pre-COVID-19 Flu. This is consistent with our observation of more extended symptoms and illness durations in the COVID-19 patients, but may also be explained by the adoption of more stringent self-imposed quarantine measures after COVID-19 diagnosis in addition to shelter-in-place. RHR has similar profiles for flu and COVID-19, suggesting the possibility to be used for real-time detection and monitoring of the time-course of the illness, but not necessarily forecasting before (self-reported) symptom onset or distinction of different ILIs. Sleep changes appear inconclusive, as the post-onset total sleep time increase observed for pre-stay-at-home flu may be explained by changes in sleeping schedule during sick days that would be less prominent when working from home.

We present the first read-out of PGHD including longitudinal symptoms reports and linked data from commercial wearables for 6,926 diagnosed flu and 230 diagnosed COVID-19 patients remotely collected in real-life settings. We describe symptoms across different ILIs, in order to build a better understanding of COVID-19 presentation and contextualize it with flu. Specific symptoms, including chest pain, shortness of breath or anosmia, as well as combinations of these symptoms (e.g., shortness of breath and coughing) were more prevalent in COVID-19 as compared to Non-COVID-19 Flu. Other symptoms, including fatigue and cough, were more pronounced later after illness onset. Generally, patients reported longer COVID-19 illnesses (median of 12 days) than Non-COVID-19 and Pre-COVID-19 Flu illnesses (9 and 7 days, respectively). Our results show that COVID-19 patients are far more likely to seek emergency care, and are more likely to be hospi-talized. This observation supports the severity with which COVID-19 infection is taken, and the scaling up of the medical response to the pandemic. We also observe however that COVID-19 patients are less likely to be prescribed medication, indicating the current dearth of available proven treatments, and further underlining the need for remote monitoring and containment (29) . Differences in self-reported symptoms are supported by data from the subset of our cohort who also contributed wearable sensors PGHD. We observed that activity (measured as lost steps) was reduced longer and more prominently for COVID-19 patients as compared to other groups. This is consistent with the observed longer illness durations for COVID-19 generally, but could also reflect orders to self-quarantine. Consistent with this line of reasoning, 89% of the COVID-19 cohort reported being told by a medical provider to selfquarantine compared to only 57% of the Non-COVID-19 Flu cohort. We observe a significantly increased fraction of participants with elevated RHR measurements in the 2 days surrounding ILI symptoms onset. This has previously been observed for other ILIs (7), and is also observed for COVID-19 patients. If validated in large and representative populations (30), our findings suggest the potential for PGHD to support remote monitoring of infectious disease patients, with opportunities ranging from improve resource allocation for further remote diagnostic, to inform population-level early-warning systems based on geo-localized aggregate symptoms. The prominence of passively collected sensor data among other PGHD data sources requires further investigation. The key observation from our findings is that although sensor data can be a general, real-time trigger for ILI recognition and tracking, subjective symptom information is required to provide specificity in distinguishing COVID-19 from flu and likely other ILI. While disease-specificity may be improved by additional data sources such as pulse oximetry, respiration rate, electrodermal activity, but may still ultimately require further interactions with the individual. This underlines that . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. . to assess any clinical utility.

The studied cohorts come from convenience samples that are not representative of the US population at large, in particular we note that African-Americans, alongside older individuals, are underrepresented in our cohort therefore our findings may not generalize. Increasing access and usage of these tools in these risk groups (31) is therefore of critical importance. Differences in the rate of the occurrence across demographic groups and disease severity levels have not been investigated, though preliminary findings point to possible differences with other ILIs (23, 32) . Largescale connected populations offer the ability to reduce this divide and examine the impact COVID-19 is having across demographic and geographic groups, helping to highlight vulnerable populations and target care delivery (4) . It is also clear that even our approach may underestimate severity, due to participants not reporting symptoms or not wearing sensors in days when symptoms are most severe, or during hospitalization events (see Figure S3 ). The conclusions presented here may therefore even be conservative. Additionally, we recognize that our analyses do not immediately translate to real-time implementation, due to lag in data collection that comes from sensor synchronization and data synchronization. Lags in our dataflow are nevertheless small compared to the gains in symptom detection and reporting compared to canonical practice. Clearly the COVID-19 pandemic is still in its infancy, and although our work on surveying annual waves of ILI infections is well established, testing is currently reserved for patients with severe symptoms, or strongly suspected to be infected with COVID-19. This could create a self-fulfilling prophecy where patients that match the current assumptions about presentation are more likely to be tested, and therefore more likely to confirm current thinking in our symptom reporting. For example, a patient with non-canonical COVID-19 symptoms may not get tested but simply be told they have another ILI. This further underlines the criticality of widespread testing to develop a canonical presentation. We are also aware that in our selection of ILI events, for each participant we select COVID-19 events or the most recent ILI event. This biases our analysis towards later calendar dates when sensor data is most affected by social distancing. For this reason we have included a chronologically parallel group of Non-COVID-19 Flu patients. A second issue is that we could be missing participants' most severe ILI events, which could have happened earlier in the season. We will continue to monitor symptomatic and behavioral changes associated with COVID-19 and non-COVID-19 ILIs as more events are captured and as guidance on social distancing and stay-athome measures are relaxed. Further analysis will focus on how strongly these measures confound our observations. Outlook. These subjective (self-reported) and objective (via commercial wearables) PGHD allow us to learn about symptom presentation, care-seeking behavior, and contextualize COVID-19 as compared to flu. As more data is being collected, further work will focus on increasing the breadth of participation, and examining other PGHD signals which may support early detection and differentiation of ILIs. To accelerate learnings and findings generalizability it would be desirable that the many ongoing study initiatives could participate in the creation of a data consortium to facilitate access to cross-study harmonized datasets to a broader audience of qualified researchers, following the example of data consortia that are being created for Real-World Data (33) . As care becomes more decentralized and telehealth becomes more prominent (34) , PGHD can become a valuable tool on an individual level as patients transition in and out of care (35) . Equally, taken in aggregate, PGHD can provide insights into public health, such as hotspot detection and real time-monitoring for public health interventions, crucial in monitoring effectiveness of reopening and enabling contact tracing (29, 36) . The vast majority of learnings about COVID-19 have come from real world data sources such as EHR, claims, etc. PGHD can be an crucial addition to the real world data arsenal, adding a large-scale understanding of early signals, several days before impact is seen at centers of care. As the COVID-19 pandemic continues to develop, and as future annual ILI waves arrive, understanding and correctly reacting to symptom presentation will be critically important. These results support not only an emerging picture that COVID-19 has a distinct presentation, but highlight the power of PGHD, digital health, and connected populations in broadly and remotely monitoring health status. Luca Foschini et al. | covid-baseline . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. . https://doi.org/10.1101/2020.05.28.20115964 doi: medRxiv preprint 6. We'd like to know more about the symptoms you experienced. Looking back over the past 7 days, please indicate on which days you felt the following symptoms. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. . 24. Which of the following over-the-counter (non-prescription) medications did you personally decide to take to treat or manage your current symptoms in the past 24 hours? Select all that apply. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. . . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 30, 2020. . https://doi.org/10.1101/2020.05.28.20115964 doi: medRxiv preprint during a given event. These are highly subjective, thus all responses were retained, with a given date coded as "one of the worst days" if the participant indicated as such in any survey.

For event-level categorical features, the algorithm described in Figure S1 was used to collapse surveys to a single response. Numerical event-level features, for example the number of household members who have experienced ILI symptoms, were aggregated by taking the maximum value reported. All other features which could not be reconciled were simply aggregated as concatenated unique values. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 30, 2020. Figure S2 describes the percentage of each ILI cohort reporting daily symptoms between one week prior and 4 weeks post symptom onset. Figure S3 describes the percentage of observed symptom reporting for hospitalized and non-hospitalized COVID-19 cohorts, between one week prior and 4 weeks post symptom onset.

Non-COVID-19 Flu Pre-COVID-19 Flu Non-hospitalized COVID-19 Fig. S3 . Percentage of the Hospitalized (N=83, purple) and Non-hospitalized (N=147, gray) COVID-19 sub-cohorts with symptom reports for days -7 to 28 since illness onset.

A summary of coverage of wearable sensor data over the course of the study is visualized in Figure S4 . . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 30, 2020. . https://doi.org/10.1101/2020.05.28.20115964 doi: medRxiv preprint Fig. S4 . Coverage of Fitbit steps, sleep, and RHR data on each calendar date of the study, color-coded by cohort. Each row is one participant (ordered by date of ILI-onset) and each column is one calendar date. Shaded days indicate that wearable data was recorded on that day from that participant. Days highlighted in yellow indicate the ILI onset dates.

De-identified study data will be made available to qualified researchers on the Sage Synapse platform (37) in September 2020.

Luca Foschini et al. | covid-baseline bioRχiv | 23

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 30, 2020. . https://doi.org/10.1101/2020.05.28.20115964 doi: medRxiv preprint

Friends of Cancer Research Reagan-Udall Foundation for the FDA. Covid-19 evidence accelerator

Zoe covid symptom tracker

Harvard Boston Children's Hospital. Covidnearyou symptom tracker

Evidation Health. Covid-pulse

Real-time tracking of self-reported symptoms to predict potential covid-19

Flu near you: Crowdsourced symptom reporting spanning 2 influenza seasons

Influenza surveillance using wearable mobile health devices

Harnessing wearable device data to improve state-level real-time surveillance of influenza-like illness in the usa: a population-based study. The Lancet Digital Health

Fever and mobility data indicate social distancing has reduced incidence of communicable disease in the united states

Tempredict study

Presenting characteristics, comorbidities, and outcomes among 5700 patients hospitalized with COVID-19 in the new york city area

Longitudinal data from activity trackers show that those with greater inconsistency in activity levels are more likely to develop more severe depression

Accelerating research with technology: Rapid recruitment for a large-scale web-based sleep study

Influenza surveillance using wearable mobile health devices

Fitting linear mixed-effects models using lme4

The incubation period of coronavirus disease 2019 (COVID-19) from publicly reported confirmed cases: Estimation and application

Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models

Estimated influenza illnesses, medical visits, hospitalizations, and deaths in the united states -2018-2019 influenza season, 2020. 22. National Center for Health Statistics. National health and nutrition examination survey

Profiling presenting symptoms of patients screened for sars-cov-2, 2020. 24. Centre for Disease Control. Influenza positive tests reported to cdc by u.s. clinical laboratories

Centre for Disease Control. Frequently asked influenza (flu) questions: 2019-2020 season

Twitter report on covid-19 symptoms

Association between resting heart rate and inflammatory markers (white blood cell count and highsensitivity c-reactive protein) in healthy korean people

Association between resting heart rate and inflammatory biomarkers (high-sensitivity c-reactive protein, interleukin-6, and fibrinogen) (from the multi-ethnic study of atherosclerosis)

Data ineroperability an exchange to support covid-19 containment

Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (tripod): the tripod statement

COVID-19 and African Americans

Covid-19 by race and ethnicity: A national cohort study of 6 million united states veterans. medRxiv

Covid-19 research database

Remote digital monitoring in clinical trials in the time of COVID-19

Michelle Colder Carras, and Alain Labrique. Global preparedness against covid-19: We must leverage the power of digital health

The authors would like to thank Raghu Kainkaryam for technical input, and Christine Lemke, Malay Gandhi, Stephanie Jones for feedback on the manuscript. The authors would also like to thank Dr. Stephen Friend and the team at 4YouandMe for the invaluable partnership and discussions.

This study received expedited review and IRB approval from Solutions IRB (Protocol ID #2018/11/8). Waiver of informed consent was granted by the IRB. Prior to each questionnaire, participants were notified about how their survey responses and behavioral data will be used for research purposes through a disclosure.

Weekly 1-Click Item 1. Have you experienced flu-like symptoms in the past 7 days (such as fever, chills, cough, shortness of breath, and/or headache)? If you had flu-like symptoms in the past 7 days, but have recovered, please still answer YES. 4. When did you feel you were completely recovered from your illness? If you don't recall the exact date, please provide the best estimate.(a) calendar date selection 5. We'd like to know more about the symptoms you experienced. Looking back over the past 7 days, did you have any of the following symptoms? Please select all that apply. Have you been in close contact with anyone outside your household (e.g., family members, friends, coworkers, acquaintances) who has experienced flu-like symptoms recently? Close contact can include direct physical contact, face-to-face contact for longer than 15 minutes, exchange of bodily fluids, or being within 6 feet of the person for more than 15 minutes.