Consumer-grade wearables identify changes in multiple physiological systems during COVID-19 disease progression

Abstract

Consumer-grade wearables are needed to track disease, especially in the ongoing pandemic, as they can monitor patients in real time. We show that decomposing heart rate from low-cost wearable technologies into signals from different systems can give a multidimensional description of physiological changes due to COVID-19 infection. We find that the separate physiological features of basal heart rate, heart rate response to physical activity, circadian variation in heart rate, and autocorrelation of heart rate are significantly altered and can classify symptomatic versus healthy periods. Increased heart rate and autocorrelation begin at symptom onset, while the heart rate response to activity increases soon after symptom onset and increases more in individuals exhibiting cough. Symptom onset is associated with a blunting of circadian variation in heart rate, as measured by the uncertainty in the phase estimate. This work establishes an innovative data analytic approach to monitor disease progression remotely using consumer-grade wearables.

Keywords: COVID-19; circadian rhythms; disease monitoring; heart rate; mathematical modeling; wearables.

Conflict of interest statement

D.B.F. is the CSO of Arcascope, a company that makes circadian rhythms software. Both he and the University of Michigan are part owners of Arcascope. S.W.C., D.B.F., C.M., and M.T. receive research funding from an Arcascope NIH SBIR grant for a different research project. However, Arcascope did not sponsor the research presented herein. S.S. received Fitbit devices at reduced cost for the Intern Health Study. J.T., C.M., D.B.F., S.W.C., M.T. S.S., Y.F., and C.F. are inventors of intellectual property related to this work, for which the University of Michigan is pursuing intellectual property protections.

© 2022 The Authors.

Figures

Graphical abstract

Figure 1

Basal HR and autocorrelation increased at symptom onset in COVID patients (A) Sample fit of the HR algorithm ±5 days around COVID symptom onset (yellow day) in one individual. The red points correspond to heart data as measured by a wearable device, while the blue points correspond to the fit from the HR algorithm. The magenta line is a plot of the daily basal HR fit from the algorithm. Note that the lower HR data typically occurred during sleep, which is removed by the HR algorithm. (B) Daily basal HR estimates and SE of the mean from participants in each of the three study groups (blue, Roadmap-CS; red, IHS; yellow, Mishra et al. participants). (C and D) Plot of the percentage of participants with increased basal HR (C) and autocorrelation (D) on the respective days around symptom onset compared with the basal HR and autocorrelation in the baseline period. The red bars indicate the percentage of participants with a significantly increased parameter value compared with the participant’s baseline distribution. (E) Autocorrelation of the residuals from the HR algorithm. We computed total noise at every time point (see STAR Methods) as the difference between the daily circadian fit and the data values. Then, we fit a linear autocorrelation model to the total noise at time t + 1 versus the total noise at time t fixing the intercept at 0 (uncorrelated noise should be normally distributed around 0). The dashed dark blue lines are the mean individual linear fits during the baseline period for each individual. The dashed orange lines are the mean individual linear fits on day 1 (the day after symptom onset) for each individual. The solid lines plot the population means of the linear fits. The mean slope on day 1 (0.78) is significantly higher than the population mean slope from the baseline period (0.73, p = 0.03).

Figure 2

HRpS increased around COVID symptom onset (A) Schematic of the calculation of the HRpS residual, for sample individuals. For each individual, we fit a linear relationship between the daily step counts and the HRpS parameter estimated for all days up to 10 days before symptom onset (blue dots). The red dots represent the daily step count versus HRpS pairs from days 0 to 14 after symptom onset. Then, for each daily step count and HRpS estimate pair, we compute the residual as the difference between the HRpS parameter estimate and the HRpS estimate from the predicted linear relationship. The left panel corresponds to a participant that reported cough as a symptom while the right panel participant did not. (B) Mean HRpS residuals and SE of the mean for the respective days around COVID symptom onset in the whole population. The mean residual peaks on days 5 and 6 and is significantly greater than 0 on those days (day 5, p = 0.021; day 6, p = 0.0092). (C) Percentage of participants with increased HRpS residual on the respective days around symptom onset compared with the baseline period. The red bars indicate the percentage of participants with a significantly increased parameter value compared with the participant’s baseline distribution. (D) The distribution of HRpS residuals in participants that reported cough (red, mean = 0.0198) versus the distribution from those participants that did not (blue, mean = −0.0043). The mean distributions are significantly different (p = 0.018). (E) Mean residual and SE of the mean for subjects who did report cough (red) and did not report cough (blue) on the respective days around symptom onset.

Figure 3

Circadian phase uncertainty increased around COVID symptom onset (A) The HR algorithm samples the circadian phase from the posterior distribution a user-specified amount of times. The blue bars plot the histogram of circadian phase samples from 1 day of data in one example individual. The red line (5:54 a.m.) is the mean circadian phase from the sampled distribution and is taken as the estimate of the circadian phase for the day. The dashed red lines are the uncertainty bounds that correspond to the number of hours on either side of the phase estimate containing 80% of the samples (in this case, 5.61 h). (B) Sample actogram for one participant from days −50 to 5 around COVID symptom onset. The black histogram-like bars represent HR throughout the day, i.e., thicker bars correspond to higher HR values in the 5-min bin. The actogram is double plotted; that is, the first row plots HR on day −50 and then day −49, the second row days −49 and −48, etc. The red line plots the circadian phase estimate from the HR algorithm with the shaded region representing the hours of uncertainty in that estimate. See STAR Methods and (A). (C) The mean phase uncertainty in hours for days −35 to 14 around COVID symptom onset in the whole population. The shaded region corresponds to the SE of the mean. The uncertainty was increased when compared with our previously published algorithm because of the shorter window of data.

Figure 4

A linear support vector machine model successfully classifies early pre-symptomatic periods versus infection periods (A) Schematic for our machine learning analysis. Two classes were passed into a linear support vector machine classification learner: class 1 consisting of individual parameter estimates on days −10 to −6 (P-10, …, P-6) and class 2 consisting of individual parameter estimates on pre-symptomatic days −5 to −1 (P-5, …, P-1). A total of 55 of the 89 individuals met our requirements for the pre-symptomatic machine learning classification. (B–E) Receiver operating characteristic (ROC) curve when using only the amplitude (B), circadian phase uncertainty (C), basal HR (D), or autocorrelated noise (E) parameter as a feature. (F) ROC curve when using all parameters (basal HR, HRpS residual, autocorrelated and uncorrelated noise, amplitude, and circadian phase uncertainty) as features.