Real-time alerting system for COVID-19 and other stress events using wearable data

Abstract

Early detection of infectious diseases is crucial for reducing transmission and facilitating early intervention. In this study, we built a real-time smartwatch-based alerting system that detects aberrant physiological and activity signals (heart rates and steps) associated with the onset of early infection and implemented this system in a prospective study. In a cohort of 3,318 participants, of whom 84 were infected with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), this system generated alerts for pre-symptomatic and asymptomatic SARS-CoV-2 infection in 67 (80%) of the infected individuals. Pre-symptomatic signals were observed at a median of 3 days before symptom onset. Examination of detailed survey responses provided by the participants revealed that other respiratory infections as well as events not associated with infection, such as stress, alcohol consumption and travel, could also trigger alerts, albeit at a much lower mean frequency (1.15 alert days per person compared to 3.42 alert days per person for coronavirus disease 2019 cases). Thus, analysis of smartwatch signals by an online detection algorithm provides advance warning of SARS-CoV-2 infection in a high percentage of cases. This study shows that a real-time alerting system can be used for early detection of infection and other stressors and employed on an open-source platform that is scalable to millions of users.

Conflict of interest statement

M.P.S is a co-founder and member of the scientific advisory board of Personalis, Qbio, January, SensOmics, Protos, Mirvie, NiMo, Onza and Oralome. He is also on the scientific advisory board of Danaher, Genapsys and Jupiter.

© 2021. The Author(s).

Figures

Fig. 1. Study overview.

a, Participants with a Fitbit and/or Apple Watch were asked to share their wearable and survey data using the study mobile app MyPHD. The app securely transfers the de-identified data (heart rates, steps and survey events) to the back-end for real-time analysis. On the back-end, three online infection detection algorithms were deployed, and the results from one of the algorithms (online NightSignal) were returned to the participants using the app: red alerts indicate abnormal changes in overnight RHR; green alerts indicate normal overnight RHR. b, A real-world example of real-time pre-symptomatic detection of COVID-19 using the online NightSignal algorithm for a participant using an Apple Watch. Alerts were triggered 2 d before the symptom onset date and continued until 3 d after the diagnosis date.

Fig. 2. Examples of COVID-19 real-time pre-symptomatic and asymptomatic detection.

a, Pe-symptomatic detection for participants who tested positive for SARS-CoV-2 using a Fitbit (top) and an Apple Watch (bottom). For the participant using a Fitbit, the panels show data derived using the online NightSignal, online RHRAD and online CuSum algorithms. Alerts generated by the NightSignal algorithm initiated 3 d before symptom onset and persisted for the next 15 d. For the participant using an Apple Watch, the panel shows data derived using the online NightSignal algorithm. Alerts appeared 10 d before symptom onset and continued until 10 d after that. b, Asymptomatic detection for participants who tested positive for SARS-CoV-2 using an Apple Watch (top) or a Fitbit (bottom). For the participant using an Apple Watch, the panel shows data derived using the online NightSignal algorithm. For the participant using a Fitbit, the panels show data derived using the online NightSignal, online RHRAD and online CuSum algorithms.

Fig. 3. Association of red alerts with COVID-19 symptoms and diagnosis.

a, Association of the initiation of red alerts in the NightSignal algorithm with COVID-19 symptom onset in 66 participants who tested positive for SARS-CoV-2 with symptoms using a Fitbit or an Apple Watch, with respect to a window of time centered around symptom onset (21 d before to 21 d after symptom onset). The NightSignal algorithm achieved pre-symptomatic detection in 53 participants and post-symptomatic detection in five participants; eight participants did not receive any red alert associated with their COVID-19 symptoms during the detection window. b, Association of the initiation of red alerts in the NightSignal algorithm with a SARS-CoV-2-positive test in asymptomatic participants using a Fitbit or an Apple Watch. The plot shows 21 d before and 21 d after the COVID-19 diagnosis date. The NightSignal algorithm achieved asymptomatic detection in 14 participants; four participants did not receive any red alert associated with their COVID-19 diagnosis during the detection window. c, The distribution of scores of red alerts with respect to a window of 21 d before and 21 d after the COVID-19 symptom onset date in participants who tested positive for SARS-CoV-2. Red bars indicate the cumulative scores of red alerts, as described in the Methods.

Fig. 4. Association of clustered red alerts with symptom progression.

a, Bubble plot showing day-by-day frequency counts of individuals reporting symptoms during the second half of the infection detection window (from symptom onset to 21 d later). Bubble size and shading are indicative of the relative magnitude of the frequency count and the median severity, respectively. The percentage in the brackets alongside each symptom indicates the total number of individuals reporting that symptom over all 21 d as a fraction of the total number of symptomatic participants who tested positive for SARS-CoV-2. b, Illustrative example tracing the symptoms of a participant who tested positive for SARS-CoV-2 from symptom onset to 21 d later and continuing thereafter intermittently for an additional 2 months. For each day, an aggregate symptom score was computed as the sum of the relative severity of the symptoms, each weighted by its specificity to individuals who tested positive for SARS-CoV-2. The scale for computing the symptom score was based on individual symptom intensities, each ranging from mild (score = 1) to very severe (score = 5) as reported by the participant. The aggregated score shown in this bar plot is a measure of the overall severity. c, The bar plots show the percentages of red alert periods (from NightSignal algorithm) associated with each symptom (left) or activity (right) as annotated by participants who tested positive for COVID-19 as well as by participants who tested negative for COVID-19. Source data

Fig. 5. Association between RHR elevation with symptoms and activities.

a, Example of a SARS-CoV-2-positive case. Alerts began before symptom onset and continued until the diagnosis date. Elevated RHR was associated with severe fatigue, fever and headache. b, Example of a SARS-CoV-2-negative case. Even though alerts were present throughout the symptom period, the magnitude of RHR elevation is noticeably lower compared to the SARS-CoV-2-positive case in a. c, Example of alerts associated with M. pneumonia infection. On the 4th day after symptom onset, the participant received a negative test for SARS-CoV-2. M. pneumonia was detected on the 8th day after symptom onset, and both symptoms and alerts were receded on the 14th day after symptom onset (5 d after antibiotic therapy). d, e, Examples illustrating associations of stress, poor sleep and mood change with the occurrence of alerts in the individuals who tested negative for SARS-CoV-2. f, Example illustrating the association between repeated alcohol consumption and alerts in an individual who tested negative for SARS-CoV-2. g, Example illustrating extended altitude change and alerts in an individual who tested negative for SARS-CoV-2.

Fig. 6. Association of red alerts with COVID-19 vaccination.

a, Examples of the association of COVID-19 vaccination with alerts from the online NightSignal algorithm. b, Effects of COVID-19 vaccination on average RHR overnight in the cases of (from left to right) the first dose of the Moderna vaccine, the first dose of the Pfizer-BioNTech vaccine, the second dose of the Moderna vaccine and the second dose of the Pfizer-BioNTech vaccine. c, Distribution of symptoms reported and alerts received for 1 week after the first and second doses of the Pfizer-BioNTech and Moderna vaccines.

Extended Data Fig. 1. Wearable devices distribution.

Of the 2,155 participants who had a smartwatch: 1,031 wore Fitbits, 970 wore Apple Watches, 98 wore Garmin, and 56 had other devices. Note that we consider the device with the most amount of data as the main device in case of having more than one device.

Extended Data Fig. 2. Flow chart of participants.

A flow chart over the participants in the study including the number of participants with sufficient wearable data in each category — in COVID-19 positive cases, if there are more than two days missing in a week before symptom onset (for the symptomatic cases) or more than a week missing data in 21 days before the diagnosis date (for the asymptomatic cases), we consider the case as insufficient data — and number of participants in each sub-category of COVID-19 positive cases: prospective/retrospective, Fitbit/Apple Watch, and symptomatic/asymptomatic. As shown in the figure, 66 participants were symptomatic and 18 were asymptomatic. 45 participants (5 asymptomatic) confirmed diagnosis of COVID-19 via written documentation or verbal confirmation of their test result.

Extended Data Fig. 3. Deterministic Finite State Machine in NightSignal algorithm.

The state machine consists of six states, each labeled with an alert color and three symbols for transition between states based on the current average RHR overnight and the deviation level from the baseline (streaming median of averages of RHR overnight). For example, a red alert gets triggered if for two consecutive nights, the average RHR overnight is at least four bpm above the calculated baseline.

Extended Data Fig. 4. Comparing distribution of RHR overnight data in Fitbit vs. Apple Watch.

To show the distribution of RHR data in Fitbit and Apple Watch, RHR overnight data points have been divided into 16 ranges (1-10, 10-20, etc.) and each bar depicts the total number of nights that falls into each group. Each participant is presented with a color. Note that for the majority of participants, for most of the nights, there are 300 to 500 RHR overnight data points (that is, almost minute resolution) in the case of Fitbit. However, the range differs considerably in Apple Watch; the reason is that Apple Watch takes heart rate and step counts readings with different resolutions based on the activities. Note that in the NightSignal algorithm, for each night, we use the RHR overnight data regardless of the amount of data. The reasons for not setting a threshold for the minimum amount of data points required are as follows: the first reason is that in most cases, even a very few data points are sufficient to get a proper average RHR overnight because we only consider HR records where the corresponding time interval (for example, few minutes) for step count is zero. The second reason for that is if we do so (for example, set the threshold to 40 data points), we will miss a significant number of nights (for example, first four bars). It is important to note that there were 15 participants who had data collected from both Fitbit and Apple Watch. With respect to the average RHR overnight that is used in NightSignal algorithm, except for one participant who had significant different values between the Fitbit and Apple Watch (one possible reason can be that the watch has been used by another person), for other 14 participants, the median delta between the Fitbit and Apple Watch data for joint nights (total of 1,006 nights) is only two bpm. Hence, despite the difference between the distribution of RHR overnight data points, there is no significant difference in the average RHR overnight point of view that is being used in the NightSignal algorithm between two devices. It is noteworthy to mention that unfortunately we did not have many COVID-19 positive cases with Garmin data to properly evaluate the algorithm for Garmin watches, and thus we did not include Garmin in our analysis.

Extended Data Fig. 5. Thresholds and parameters in the NightSignal algorithm.

As we discussed previously, the NightSignal algorithm uses the streaming median of average RHR overnight as the baseline. We believe that it is a proper individual healthy baseline as we show that the fluctuation is insignificant and it usually deviates due to a long-term abnormal event (for example, infection, medication consumption, vaccination). (A) The minimum number of nights required to hit a baseline close to the baseline over three months (within ± two bpm from the baseline - the reason behind choosing the threshold of two is that the baseline would still remain in the green zone) for the majority of participants. As depicted in the figure, for over 80% of the participants, the proper baseline was observed after only seven nights. (B) The range (max-min) of the medians of average RHR overnight during three months of data with the median value of only three bpm. Similarly, Quer et al. studied the variability in individual resting heart rate and showed that most participants had a median weekly fluctuation in RHR of only three bpm. Given the fact that we only consider RHR overnight, the median fluctuation in RHR overnight is still three bpm even for a duration of three months. (C) The corresponding distribution of the medians of average RHR overnight over three months indicates a standard deviation that is very low (0.8).

Extended Data Fig. 6. Additional examples of alerts for COVID-19 positive, COVID-19 negative, and untested participants.

(a) Signals for a COVID-19 positive case with mild and short-lasting symptoms. Note that this participant received zero red alerts during a healthy period of over seven months and only received two red alerts starting the night of the symptom onset date (the only period where the RHR overnight is much higher than baseline for three consecutive nights). This example shows missing only two days of data can lead to missing a whole cluster of COVID-19 related alerts. (b) A COVID-19 positive case shows that triggering alerts based on deviations in two nights does not lead to delayed alert. (c) and (d) Both cases show participants who reported confirmed diagnosis of COVID-19 but with consecutive negative tests after the positive test, seven and two negative tests, respectively. In both cases, the NightSignal algorithm correctly did not trigger any red alert. (e) An example of a sick participant with moderate symptoms followed by a COVID-19 negative test. Note that RHR overnight period began to increase three nights before the symptoms developed. (f) An example of an untested participant who reported no illness or symptoms of any kind during the study (only some poor sleep and alcohol consumption points for P174112). This participant received no red alerts for over a three months healthy period, however alcohol consumption affects the RHR overnight but the impact is either only for one night or not severe enough to trigger a red alert. (g) This example indicates the impact of lifestyle changes (here weekdays vs. weekends) on RHR overnight in a COVID-19 positive participant. Clearly, repeated patterns occur before COVID-19 positive test and reoccurrence of the similar patterns begin two weeks after that (likely due to recovery from the illness and starting the same lifestyle).

Extended Data Fig. 7. Impact of other non-infectious events on NightSignal alerts (all day vs. overnight).

Examples of the impact of different events (for example, home and work stress, travel, and intense exercise) on the alerts based on two configurations: all day vs. overnight. Comparing the plots for each participant shows that the NightSignal algorithm reduces possible false positives due to non-infectious events by analyzing RHR overnight.

Extended Data Fig. 8. Impact of winter holidays on NightSignal alerts.

As shown previously, there is a noticeable increase in the number of alerts during winter holidays — particularly late December and beginning of January — (‘holiday bump’) due to the higher rate of travel, alcohol, entertainment, stress, and illness compared to other times of the year.

Extended Data Fig. 9. Sensitivity and false positive rate changes in the NightSignal algorithm for different thresholds.

Under different thresholds, we report the sensitivity and false positive rate for the NightSignal algorithm. For the sake of clarity, it is important to mention that for anomaly detection problems like the one addressed in this study, having a perfect ground truth is very hard (here, due to the facts that the exact time of virus exposure is unknown in majority of the cases and also other non-infectious events can trigger the alerts - usually smaller clusters). Hence, defining the exact COVID-19 illness period and consequently correct alert is difficult. In other words, all the alerts within the assumed detection window might not necessarily be related to COVID-19, so a higher sensitivity does not by itself mean a greater accurate detection rate. Therefore, besides the aforementioned reasons in Supplementary Fig. 5, to keep the false positive rate sufficiently low, as well as achieving a high sensitivity, we use the threshold of three for the warning (yellow) level. Moreover, this threshold can be easily configured by each individual for a more personalized and precise analysis. This modification will require IRB approval and is outside of the scope of the present study and is thus part of a future work.