Rest-activity rhythmicity as digital biomarkers of postoperative recovery and complications in children

Patient selection and recruitment

This study was approved by the Institutional Review Board at the Ann and Robert H. Lurie Children’s Hospital of Chicago (IRB: #2018-1836). Children 3-18 years old who underwent laparoscopic appendectomy for complicated appendicitis from 2019 to 2022 were approached for inclusion. Patients were excluded if any pre-existing mobility limitations, comorbidities, or doctor-ordered physical activity limits of more than 48 hours post-surgery would impact their postoperative recovery. Patients and/or their caregivers who do not speak English or Spanish were also excluded due to the cost and time associated with translation services.

A research coordinator identified eligible hospitalized patients through the Electric Health Record (EHR) and then approached and recruited pediatric patients and their caregivers after appendectomy. Written consent was obtained from the caregiver for all children under 18 years old. In addition, verbal assents were obtained from children under 12 years old, and written assents from children 12-17. After the consent/assent process, patients were given a wearable fitness tracker (Fitbit: Inspire HR or Inspire 2). The Fitbit application was installed on the caregiver’s smartphone, and a deidentified Fitbit account was set up, ready for wearable data collection via a cloud platform (Fitabase) to occur through postoperative days (POD) 21. Data were collected from POD 1-21, where POD 1 is defined as the first calendar day after surgery, and each day runs from 00:00 to 23:59. The Fitbit was given to the participant as compensation at the end of this period, marking study completion.

Collection of clinical information

The following clinical information was extracted from the patient’s EHR: age, sex, race/ethnicity, postoperative length of stay, and the type and date of any postoperative complications during the recovery period. Phone surveys were conducted on POD 3, 7, 10, 14, and 21 to collect information regarding any new symptoms and healthcare utilization during this period. A panel of five clinicians confirmed the diagnosis of postoperative complications and their corresponding subtypes through EHR and survey results review.

Collection of wearable data

The Fitbit and app automatically synchronized wearable data to the Fitabase cloud platform, including per-minute HR and per-minute step count, as well as daily summaries of resting HR, activity (e.g., total step counts), and sleep (e.g., total sleep minutes asleep). After the study period, per-minute HR, per-minute step count, and daily sleep metrics were downloaded from the cloud platform (Fitabase) and stored on the secure hospital server. Existing studies have validated these Fitbit data with acceptable accuracies for children (3-18 years old) for health monitoring applications.^48–54 Fitbit data were collected over a 21-day study period, from POD 1 to POD 21. For all participants, this period included a brief hospital stay following surgery and subsequent discharge to the community. The day of having surgery is defined as POD 0, which is excluded from the study period. The study period starts the first day after surgery day (POD1), which is defined as postoperative day 1, at 0:00 (midnight, or 12:00AM).

Patient groups

After the study period, patients were retrospectively assigned to one of two groups: those who were diagnosed with one or more postoperative complications (e.g., intra-abdominal abscess, surgical site infection, etc.), ⁴⁴ and those who were not diagnosed with any complication (normative recovery).

Outcome measures of 24-hour activity rhythms

Based on top-ranking Fitbit features in our previous study, ⁴² we selected Periodicity and Amplitude as the primary RAR measures of interest for the 24-hour Activity Rhythm to maximize clinical interpretability and specifically capture behavioral movement contrast (detailed selection rationale is provided in eMethods 1). Low periodicity and reduced amplitude of 24-hour activity rhythm are significantly linked to disease and disease-related deaths. ⁵⁵ These two methods and the metrics of activity rhythms were chosen since there is previous evidence to support their use as digital biomarkers for remote patient monitoring, such as for Alzheimer’s disease, ⁵⁶ cancer, ⁵⁷ and depression. ⁵⁸ A moving average was applied to the per-minute step count data using a window size of 15 minutes, as is commonly used by circadian studies.^59,60 The moving average starts at time zero for the day (00:00) and stops at the end of the day (23:59). This resulted in up to 96 data points per 24-hour period from which the outcome measures were computed. The Periodicity of the 24-hour Activity Rhythm was extracted from the Periodogram of these data, a non-parametric method for estimating the power spectral density of a time series.^61,62 A value above 1.0 indicates the presence of a 24-hour rhythm, whereas a value below 1.0 implies a weaker or less-defined 24-hour rhythm. The greater the value, the stronger the 24-hour rhythm appears to be. Amplitude of the 24-hour Activity Rhythm was derived using Cosinor Regression, a parametric method that estimates the maximum difference in the day from mean activities over the 24-hour period.^62,63 A reduction in this measure would indicate circadian disruption.^62,64 The methodology for processing per-minute step counts data and detailed metric extraction are presented in the Supplementary (eMethods 2-4 and eFigures 1a, 2a, 3, 4a).

Outcome measures of 24-hour heart rate rhythms

Based on previous studies,^42,62 the Circadian Quotient (which our prior study termed relative amplitude) ⁴² was selected as the primary RAR measure of interest for the 24-hour HR Rhythm to provide a clinically interpretable, standardized measure of autonomic regulation (detailed selection rationale in eMethods 1). It is a simple measure of the difference between periods of peak HR and rest HR, allowing for more consistent comparisons across different patients within the study cohort. The same data preprocessing approach of moving average and modeling approaches, including Periodogram and Cosinor Regression, were applied to per-minute HR data (presented in the Supplementary [eMethods 2,5 and eFigures 1b, 2b, 4b]. Then, the Circadian quotient of 24-hour HR rhythm was calculated as the Amplitude of the 24-hour HR Rhythm normalized by mean HR across the day. Lower values of this metric indicate circadian disruption.^62,64

Statistical analysis

For analysis, we only included PODs on which the Fitbit computed the daily sleep metrics and resting HR to ensure periods of sleep and wake were both incorporated in the RAR metrics; all other days were excluded. To compare demographic and clinical characteristics between the groups with and without complications, two-sample t-tests were conducted for continuous variables and Chi-square tests for categorical variables. The three RAR metrics were summarized using mean and standard deviation (SD) or the median with interquartile range (IQR) by the groups. Outliers were defined as values exceeding 1.5 times the IQR below the first quartile or above the third quartile to present data distribution and individual differences. ⁶⁵ Outliers on each POD were incorporated into the curve-fitting processes to account for individual differences. To compare the changes over time, each RAR metric was modeled using smoothing spline regression over the 21-day study period, with the mean of each POD as a knot. Separate models were created for groups with and without complications, and smoothing parameters were estimated using Generalized Cross-Validation (GCV). ⁶⁶ Next, a nonlinear, cubic-plateau regression was applied to the group without complications to estimate the POD to plateau, along with its 95% confidence interval (CI). Multivariable mixed-effects models were used to compare the main effect of each RAR metric across PODs between the groups with and without complications, accounting for within-subject random effects. The models were also adjusted for sex, age, and race/ethnicity. Statistical analyses were conducted with PyCharm (Python 3.10) and RStudio (R 4.3.3) software (details about packages in eMethods 6). All statistical tests were two-sided, and statistical significance was set as 0.05.

Recruitment, clinical outcomes and wearable data

Of the 127 patients approached during study recruitment, 24 (18.9%) declined participation (eFigure 5). Of the 103 enrolled patients, nine were excluded from the analysis because they did not have any sleep metrics, step count, or HR data available from their Fitbit data. Of the 94 patients included for analysis, 81 had no complications during recovery (86%) and 13 had complications (10 with intra-abdominal abscess, 2 with superficial surgical site infections, and one with a small bowel obstruction).

There were no statistical cohort differences between the groups with and without complications (Table 1) in age (p=.093) and sex (p=.96). There was a difference in race/ethnicity distribution (p=.006), wherein 61% of the patients who experienced complications were Hispanic and 64% patients were Hispanic within 94 patients. The number of days with Fitbit data available was not significantly different between the groups with and without complications (mean [SD], 13.3 [5.9] vs. 13.8 [6.1] days; p=.80). The group with complications had a longer hospital postoperative length of stay than the group without complications (mean [SD], 3.6 days [1.8] vs. 6.4 [4.8], p<.001).

OPEN IN VIEWER

Table 1.

Cohort characteristics and wearable data availability.

Characteristics	Patients, No. (%)		P-value
	Without complications	With complications
Total	81 (86%)	13 (14%)	N/A
Demographics
Age, mean (SD)	10.0 (3.9)	11.8 (2.5)	.093
Female	36 (44%)	8 (61%)	.962
Race/Ethnicity
White	14 (17%)	4 (31%)	.006
Hispanic	52 (64%)	8 (61%)
Others	15 (19%)	1 (8%)
Postoperative Information
Length of Stay, mean (SD)	3.6 (1.8)	6.4 (4.8)	<.001
Wearable Data Collection
Time with Data, mean (SD), day	13.3 (5.9)	13.8 (6.1)	.803

The Fitbit data availability is summarized in Table 2, where the proportion of participants with any available HR data is 93.2% (N=96/103) and with any available sleep metrics is 91.3% (N=94/103). Across the 96 participants with any HR data, 74.7% of their recorded days have HR data (1505 days total) and 58.7% of their recorded days have sleep metrics (1185 days total). Step count data are only recorded when the device is recording HR, so the missing data for step counts aligns with that of HR. The distribution of participants with different numbers of days available for analysis is shown in eFigure 6.

OPEN IN VIEWER

Table 2.

Wearable data summary.

​	HR	Daily sleep metrics
Participant: availability rate (%)	93.2% (N=96/103)	91.3% (N=94/103)
Day: availability rate (%)	74.7% (N=1505 days/(96*21 days))	58.7% (N=1185 days/(96*21 days))

Periodicity of the 24-hour activity rhythm

The Periodicity of the 24-hour Activity Rhythm increased over time for both groups following surgery; however, the increase was delayed in the group with complications. The changes in the Periodicity of the 24-hour Activity Rhythm over the 21-day study period for groups with and without complications were shown in Figure 1(a). A significant difference between the groups was observed when comparing the adjusted means of Periodicity on each POD over the 21-day period (p=.008, linear mixed-effect model).OPEN IN VIEWER

For the normative recovery trajectory, Periodicity of the 24-hour Activity Rhythm reached a plateau at a value of 2.6 on POD 11 (95% confidence interval [CI]: POD 9-14, Figure 1(b)). This suggests that during the early postoperative period, RARs are often absent or weaker than in the later recovery period, with ultradian rhythms (<24-hour period) potentially dominating activity patterns. In contrast, the model failed to converge for the 21-day trajectory of the group with complications, suggesting that the trajectory did not plateau within the 21-day study period or could not be predicted based on the available data. These differences in the recovery trajectories between groups are illustrated in Figure 1(c). Across POD 1-20, a lower proportion of patients with complications exhibited clear RARs compared to those without complications (Figure 2).OPEN IN VIEWER

Amplitude of the 24-hour activity rhythm

Like Periodicity, the Amplitude of the 24-hour Activity Rhythm increased over time for both groups, and the group with complications exhibited a delayed increase, not reaching a plateau during the study period. Figure 3(a) illustrates changes in the Amplitude of the 24-hour Activity Rhythm over the 21-day study period for groups with and without complications. The average standard error of the residuals from the cosinor regression was 0.68 in the group without complications and 0.43 in the group with complications. These values are small relative to the overall range of step count measurements in the dataset (0–120 steps/min), supporting the adequacy of the model fit. A significant difference in amplitude was observed between the groups when comparing the adjusted mean on each POD over the 21-day study period (p=.003, linear mixed-effect model). The normative recovery trajectory for the group without complications, shown in Figure 3(b), reached a plateau at the modeled peak amplitude of 6.5 step counts per minute on POD 20 (95% CI: POD 16-24). The trajectory fit for the group with complications did not converge (Figure 3(c)).OPEN IN VIEWER

Circadian quotient of the 24-hour heart rate rhythm

Like the activity rhythm metrics, the Circadian quotient of the 24-hour HR Rhythm over the 21-day study period increased for both groups (Figure 4(a)). However, the group with complications exhibited a slower increase than the group without complications and did not reach a plateau during the study period. The mean of standard errors on residuals of the cosinor regression fit was 0.96 for the group without complications, compared to 2.18 for the group with complications. These values are small relative to the overall range of HR measurements in the dataset (40–150 bpm), supporting the adequacy of the model fit. A significant difference was observed between the groups when comparing the adjusted mean values on each POD over the 21-day study period (p<.001, linear mixed-effect model). The normative recovery trajectory fit, shown in Figure 4(b), reached a plateau at the modeled peak amplitude of 0.16 on POD 12 (95% CI: POD 9-15), while the trajectory fit for the group with complications did not converge (Figure 4(c)).OPEN IN VIEWER

Limitations

The limitations of this study should be considered when interpreting the findings. From a study design perspective, the comparison between patients with and without postoperative complications is constrained by our cohort’s small number of complications. Spline curves could only be generated for the cohort with complications since the recovery trajectory did not converge. It remains unclear whether this is due to the nature of the wearables data related to the complications or our cohort’s size. This study also followed patients through POD 21 only. While we assume that the cohort with complications will eventually reach stability in RAR values, this was not demonstrated within the 21-day period and warrants further study. Moreover, relying on the proposed metrics to detect complications may not capture all cases due to individual variability in postoperative recovery. However, the normative recovery trajectories can serve as a reference for generating threshold-based alerts for clinical check-ins after hospital discharge, during the early postoperative time when complications are likely to develop.

In terms of device and data limitations, the devices and data export used in this study did not provide estimates of sleep onset, midpoint, or wake times, limiting our ability to directly calculate sleep-specific circadian misalignment. ⁹⁰ In the future, next-generation wearables and validated sleep estimation methods^91–93 could be leveraged to further characterize potential circadian or RAR misalignment in pediatric populations. Furthermore, because appendectomy is often an emergency procedure, preoperative data were unavailable for comparisons to a per-participant baseline. Future studies incorporating preoperative baselines can offer greater insights into circadian differences during recovery, and should also explore other circadian metrics such as acrophase. ⁹⁴

Finally, several external and analytic factors should be considered. There are many factors that can influence circadian rhythms, such as postoperative recovery, stress, pain, medication, light exposure, feeding schedules, exercise, and social activities. Due to the practical limitations of unconstrained, real-world data collection in this study, it is nearly impossible to isolate the effect of each factor on the wearables data in this study. Further studies or methods may be needed to evaluate how these and other factors affect circadian rhythms during recovery to develop more personalized approaches to monitoring and intervention. Similarly, we did not attempt to decouple the effect of physical activity from heart rate in the recorded signals. Although other studies have estimated this using predictive models, ⁹⁵ we determined this was not feasible given the amount of missing data in the daily time series recordings and its unknown impact on the generalizability of such an approach. Future work may consider imputation based on the correlated or prior data to address this issue, but additional validation would be needed given the high level of variability in pediatric populations. Future work should also assess the applicability of our findings across additional surgical procedures, at diverse institutions, and within adult surgical populations.