Biorhythms derived from consumer wearables predict postoperative complications in children

Abstract

Postoperative complications pose substantial health risks to children who undergo surgery, yet timely detection of complications after discharge is challenging due to reliance on subjective symptom reports from children and caregivers. Alternatively, wearable devices can provide objective health measurements for continuous recovery monitoring, potentially enabling earlier complication detection in the hospital or community. This study examined biorhythm-based metrics (circadian and ultradian rhythms, derived from the daily activity and heart rate patterns recorded by a consumer wearable) and their relationship to postoperative recovery in children with and without complications. Wearables were given to 103 children for 21 days immediately after appendectomy, and biorhythm metrics were extracted from per-minute data. A machine-learned model using these metrics retrospectively predicted postoperative complications up to 3 days before formal diagnosis with 91% sensitivity and 74% specificity. Our findings suggest that wearable-derived biorhythms offer a promising, unobtrusive method for evaluating postoperative recovery. This approach has broad clinical implications for pediatric health monitoring across various care settings.

Biorhythms (activity and heart rate patterns) from consumer wearables can predict complications in children after appendectomy.

INTRODUCTION

An estimated 14% of children develop a complication after surgery (1). Postoperative complications can have devastating results if not identified and treated promptly, including prolonged hospitalization (2) or even death (3–5). Children often have greater difficulties than adults communicating their symptoms (6–8) and, after hospital discharge, typically rely on their caregiver’s judgements to inform the next steps of care. However, caregivers are often uncertain about the severity and seriousness of their child’s symptoms, typically having only subjective observations or imprecise household equipment (e.g., thermometer) to evaluate them. This can lead to delayed diagnosis of complications (9, 10), especially after discharge from the hospital, or inefficient health care utilization, such as unnecessary visits to the emergency department (11). Thus, continuous, precise, and remote monitoring of a child’s postoperative recovery would have substantial clinical value for identifying complications early and accurately, with the potential to reduce burden on the patient, family, and health care system as a whole (12–14).

Consumer wearables, which record and compute objective health metrics related to heart rate, physical activity, and sleep (15–18), offer a potential solution for remote postoperative monitoring (13). Recent studies have shown that consumer wearables such as the Fitbit can suitably capture these metrics in children (17–24). Furthermore, clinicians feel more confident when making remote treatment decisions when clinical scenarios are supplemented with these objective measures produced by the wearables (25–27). When paired with machine learning algorithms, the metrics computed by these devices have demonstrated promise in detecting postoperative complications early, with our previous work demonstrating 76% sensitivity to detect complications up to 3 days before formal diagnosis (13). However, improvements are needed to develop more sensitive and reliable algorithms for evaluating recovery, especially given the high variance in the health metrics and the intermittent adherence to device wear that are expected in broad pediatric populations (28). Thus, additional previously unknown metrics, being less sensitive to individual variations and missing data, derived from the wearables, could improve the overall performance and robustness of algorithms for postoperative monitoring in children.

We propose that biological rhythms, or biorhythms, are another potential biomarker of postoperative recovery in children. Biorhythms represent a chronobiological concept (29, 30) relating to the recurrence of biological events at relatively regular intervals, including circadian (24 hours) and ultradian (<24 hours) rhythms (Fig. 1A). Previous studies have found links between disrupted circadian rhythms and mental/mood disorders (31–33) or neurodegenerative disease progression and aging (34–36). Notably, biorhythms are disrupted by acute inflammation, such as from surgery or infections (37). As such, recovery of biorhythms [such as the normative, 24-hour fluctuations in heart rate and activity, which increase during the day and decrease at night (38)] could serve as novel, clinically relevant metrics to distinguish children with healthy postoperative recovery from those experiencing complications.

Fig. 1. Biorhythms and example wearables data.

(A) Biorhythms [suprachiasmatic nucleus (SCN) is the central pacemaker of the circadian timing system, and pineal gland (PG) generates melatonin, regulating most biorhythms in the body]. (B) Biorhythm changes after appendectomy shown with per-minute data of heart rate and step counts collected by consumer wearables. Patient #67 is a female, 12 years old when had surgery; patient #53 is a female, 13 years old when had surgery. bpm, beats per minute.

The objective of this study is to extract and examine biorhythm metrics from consumer wearables after surgery and to evaluate biorhythms as predictors of postoperative recovery in children. We tested this approach for pediatric appendectomy, which is the most common surgery in children (39) and has up to 38% rate of complications for laparoscopic appendectomy with complicated appendicitis (40). We hypothesized that biorhythms related to heart rate and physical activity, as derived from wearables, would differentiate children with and without postoperative complications. We believe that this approach will broaden the clinical utility of consumer wearables in children, because prior work with these devices has been primarily oriented toward adults (41). Insights from this study can inform new methods of pediatric monitoring to enhance patient care and outcomes, including evaluation of postoperative recovery and early detection of complications.

RESULTS

Data overview

Consumer wearables (Fitbit) were provided to 103 children (ages 3 to 18 years old) after laparoscopic appendectomy for complicated appendicitis. Per-minute heart rate and step counts data, as well as per-day estimates of sleep quality and duration, were collected from the wearable over a 21-day postoperative monitoring period starting postoperative day 1 (POD1).

To ensure that full-day biorhythms could be calculated, we only included patients in the analysis who had at least 1 day with wearable data recorded during the day and night (i.e., daily heart rate and sleep). Of the 2163 total days of monitoring (21 days from 103 children), 1259 days from 94 children had heart rate, step counts, and sleep metrics available. Demographics and per-day availability of wearable data for these 94 children are provided in Fig. 2. On average, 63.8% (SD, 11.0%) of the wearable data were available per POD. There were 10,872 total episodes of heart rate (duration $\ge$2 min) that were unavailable in this dataset, with 59.8% of these episodes lasting 2 to 10 min, suggesting brief periods of non-wear or improper device positioning (15).

Fig. 2. Data overview.

(A) Availability ratio of wearable data per postoperative day (POD). (B) Number of episodes in duration of missing data. Distributions of (C) age, (D) sex, (E) race/ethnicity, and (F) length of stay (LOS) in the hospital.

Among the 94 children, we identified three recovery scenarios (Fig. 3): 58 (61.7%) had normative recovery with no known issues, 23 (24.5%) recovered with abnormal symptoms (e.g., diarrhea, emesis, and abdominal distension) that did not lead to a complication, and 13 (13.8%) had one or more postoperative complications. Of the 13 children with complications, 10 had intraabdominal abscess, 2 had superficial surgical site infections, and 1 had a small bowel obstruction. The average postoperative length of stay (LOS) in the hospital was 4.0 days (SD 2.6) before the children were discharged home.

Fig. 3. Example recovery scenarios.

(A) Number of patients per recovery scenario. Patients who demonstrated (B) normative recovery, (C) abnormal symptoms not leading to complications, and (D) postoperative complications (with or without additional abnormal symptoms) during the 21-day postoperative monitoring period.

Some patients did not have wearable data available on the day of the diagnosed complications or abnormal symptoms. There were 9 days with data available after a diagnosed complication (diagnosed on POD3, POD6, POD7, POD9, POD10, and POD13) and 18 days with data available during abnormal symptoms that did not lead to a complication (on POD1 to POD8 and POD14).

Interpreting biorhythms

Figure 1B demonstrates the rhythmic fluctuations in wearable data recorded over four continuous PODs for two example patients with similar demographics: one with healthy normative recovery (patient #67) and one with a diagnosed complication on POD6 (patient #53). Per-day biorhythms were computed using 31 different metrics to describe the pattern of activity (step counts) and heart rate within that day (Table 1 and table S1). These metrics were extracted using standard circadian concepts and two additional modeling approaches, namely, periodogram and cosinor regression, which are further described in Materials and Methods. These modeling approaches are illustrated for the two example patients on POD6 (heart rate in Fig. 4 and step counts in fig. S1).

Table 1. Biorhythm metrics.

Extraction method	Feature/metric	Relationship to circadian rhythm
Circadian concept	Mean activity*	Hourly average total step counts in a day; large changes may indicate disruption.
	Mean of 10 most active hours	Phase; large change may indicate disruption.
	Mean of 5 least active hours	Phase; large change may indicate disruption.
	Relative amplitude	Strength of rhythm; reduction in relative amplitude may indicate disruption.
	Intradaily variability	Frequency of hour-to-hour transitions between activity and rest across the day; high variability may indicate disruption (fragmentation).
	Number of episodes*	More episodes of short durations may indicate disruption of all types.
	Average episode duration*
Periodogram	PSD24h/threshold of significance	Strength of the 24-hour rhythm; large values indicate strong regularity of the period; value less than 1.0 indicates lack of 24-hour dominant rhythm.
	PSD of second dominant rhythm/threshold of significance	Strength of the second dominant rhythm; large values indicate strong regularity of the period.
	Period of the dominant rhythm	The dominant rhythm
	Period of the second dominant rhythm	The second dominant period
	Number of periods with significance other than 24 hours	Periods of identified none 24-hour rhythms
Cosinor regression	Cosinor acrophase	Acrophase with negative values may indicate chronodisruption.
	Cosinor amplitude	Reduced amplitude may indicate disruption (fragmentation, arrhythmicity, and increased light phase activity).
	Cosinor mesor	Large change of mesor may indicate disruption.
	Cosinor relative amplitude	Reduction in relative amplitude may indicate disruption.
	Cosinor error (SE on residuals)	Large errors indicate worse fit of the cosinor model.

*Computed with data of step counts only.

Fig. 4. Example heart rate biorhythms.

Biorhythms extracted from heart rate of (A) patient #67 with normative recovery and (B) patient #53 with a diagnosed complication on POD6. Raw data of these two patients are in Fig. 1B.

Periodogram modeling (Fig. 4 and fig. S1, left panels) identifies time periods where the power spectrum density (PSD) values exceed a significance threshold, thereby indicating that a clear rhythmic pattern existed for that period. For example, patient #67, with normative recovery, exhibited a substantial 24-hour rhythm (circadian) in both heart rate and step counts, whereas patient #53, with a diagnosed complication, lacked a substantial rhythm for any period within 24 hours.

Cosinor regression modeling (Fig. 4 and fig. S1, right panels) captures the goodness of fit for a 24-hour periodic curve using an F test and P value, derived from the model sum of squares and sum of square residuals (42). For example, patient #67, with normative recovery, exhibited a significant model fit for both heart rate and step counts (P < 0.01), supporting that these data could be fit with a 24-hour periodic pattern. Conversely, patient #53, with a diagnosed complication, showed poor model fit for both heart rate (P = 0.31) and step counts (P = 0.39), indicating that the data could not be suitably fit to a 24-hour period.

The periodogram and cosinor regression analyses demonstrated that a patient with normative recovery exhibited a dominant circadian (24-hour) rhythm in both heart rate and activity on POD6. In contrast, a patient with a diagnosed complication experienced circadian disruption, with no circadian rhythm in heart rate and step counts on the same POD, which was also the day that their complication was diagnosed.

Early prediction of postoperative complications

A balanced random forest (BRF) classifier was selected to predict days of and leading up to a complication, following the methodologies from our previous studies (13, 43). This early prediction model was trained using 31 biorhythm metrics (Table 1 and table S1) and validated using nested leave-one-subject-out (LOSO) cross-validation. The day of the diagnosed complication and 1, 2, and 3 days prior were combined into the positive class (Fig. 5A), representing 2.72% of the full dataset (Fig. 5B). To mitigate this imbalance and reduce bias, the BRF trained each decision tree on a bootstrap sample with equal proportions of the positive (minority) and negative (majority) class (44). For each patient, the total number of samples was equal to the number of days with available Fitbit data. This number was similar between the group without complications [n = 81; average 13.3 days (SD = 5.9)] and the group with complications [n = 13; average 13.8 days (SD = 6.1)] (two-sample t test, P = 0.80).

Fig. 5. Early prediction of postoperative complications.

(A) Strategy of labeling each day of recovery. (B) Class distribution. (C) Results of early prediction of postoperative complications by BRF machine learning model. (D) Area under the curve of the “receiver operating characteristic” (AUC-ROC) curve. (E) Area under the curve of the precision and recall (AUC-PR) curve. (F) Prediction results for each patient who had complications on each day.

The model achieved 91% sensitivity and 74% specificity (Fig. 5C), with Matthews correlation coefficient (MCC) of 0.23 and area under the curve of the receiver operating characteristic (AUC-ROC) and precision-recall (AUC-PR) values of 0.86 (confidence interval (CI) = [0.81, 0.90]) and 0.14 (CI = [0.08, 0.22]), respectively, after bootstrapping (Fig. 5, D and E). This outperforms a naïve classifier that has an AUC-PR of 0.01. The model predicted 29 days of the positive class correctly, missing 3 days from two patients (Fig. 5F).

The most important biorhythm metrics to monitor post-appendectomy recovery

Table 2 summarizes the model’s ranked importance of the metrics by the data type and extraction method. Activity biorhythms (importance sum, 0.75) were ranked highest, followed by heart rate biorhythms (importance sum, 0.25). Heart rate biorhythms extracted by cosinor regression were more important than those extracted from other methods, whereas activity biorhythms showed comparable importance across all three methods. A full list of the metric rankings from the BRF model is provided in fig. S2.

Table 2. Sum of biorhythm importance from prediction model (grouped by extraction method).

Data	Method	Sum of importance
Heart rate	Circadian concept	0.05
	Periodogram	0.04
	Cosinor regression	0.16
Step counts	Circadian concept	0.29
	Periodogram	0.21
	Cosinor regression	0.25

After removing highly correlated metrics (tables S2 and S3), we identified six key biorhythms from the top 50%, extracted via various data types and methods. Four were computed from step counts data and two were computed from heart rate data. Figure S3 illustrates changes in these metrics over the entire 21-day monitoring period for patients with and without complications. Individuals’ values on the day of a diagnosed complication and up to 3 days prior are shown in figs. S4 to S9.

Among these key biorhythms representing the existence, strength, and robustness of the circadian rhythm in activity and heart rate, patients with complications exhibited significant differences from those without complication across five of them [one-way analysis of variance (ANOVA) with repeated measures: P < 0.05; fig. S3]. Only the PSD24h/threshold of significance computed from heart rate was not significant (P = 0.09). We found that patients with complications have a delayed recovery in reestablishing the circadian rhythm in activity and heart rate after surgery and showed a weaker 24-hour rhythm compared to patients without complications during the 21-day study period.

Evaluation of abnormal symptoms not leading to complications

We assessed the 31 biorhythm metrics over the 21-day monitoring period to differentiate between patients with abnormal symptoms that did not lead to a complication, those with normative recovery, and those with diagnosed complications. Comparing the patients with abnormal symptoms to patients with normative recovery, 76.5% of activity biorhythms from step counts and 7.1% of heart rate biorhythms from heart rate were not significant (P ≥ 0.05, one-way ANOVA with repeated measures; table S4). While comparing patients with abnormal symptoms to those with diagnosed complications, 41.2% activity biorhythms and 50.0% heart rate biorhythms did not show significance (P ≥ 0.05, table S4).

We found that the longitudinal changes in activity biorhythms for patients with abnormal symptoms closely resembled those with normative recovery, whereas the longitudinal changes in heart rate biorhythms for patients with abnormal symptoms fell between normative recovery and recovery with diagnosed complications. This holds true for the six key biorhythm metrics presented for the three recovery scenarios (Fig. 6).

Fig. 6. Comparison of the 21-day monitoring period.

Selected biorhythm metrics shown here were considered most important for a prediction model to distinguish PODs with complications versus those without. (A) Step counts, PSD24h/threshold of significance. (B) Step counts, Cosinor amplitude. (C) Step counts, number of episodes. (D) Step counts, average episode duration. (E) Heart rate, PSD24h/threshold of significance. (F) Heart rate, cosinor relative amplitude.

Although there were limited statistical differences in metrics between patients with normative recovery and patients with abnormal symptoms that did not lead to a complication, the BRF performance when trained with all three classes (normative, abnormal symptoms, and complications) demonstrated a lower, 56% recall of patients with complications (fig. S10), compared to the 91% recalled by a model trained with two classes only (no complications and complications).

Reliability of biorhythm metrics in the presence of missing data

Missing data, which can arise when the wearable device is not worn or has poor contact with the wrist, is a common occurrence in community data acquisition. To mitigate variability in wearable-based outcomes that could be explained by different wear times between patients or between days, many studies (45, 46) might require a minimum “wear time” of the device on any given day to include that day’s data in the analysis. The trade-off in wear time requirement is that higher values (more restrictive) mean that the metrics are more representative of the true, full-day behavior, whereas lower values (less restrictive) often mean that more patients and days with lower adherence can be analyzed. The amount of wearable data available for analysis in this study, based on different wear time requirements, is provided in fig. S11.

To examine the reliability of biorhythm metrics in the presence of missing data, we conducted sensitivity analyses to compute biorhythms based on different minimum wear time requirements, including 0, 3, 6, 8, 10, 12, 14, 16, 18, 20, and 23 hours, and different data sampling windows, including 10, 15, 30, 60, 120, and 180 min. These windows were informed by the summary of missing data for this dataset (Fig. 2B). The impact of minimum wear time requirements and sampling windows to the prediction model was minimal, demonstrated by paired t tests between the LOSO results in tables S5 and S6.

The best prediction model performance was achieved using a minimum wear time requirement of 0 hours and a data sampling window of 120 min (fig. S12 and S13). A minimum wear time of 0 hours means that any amount of heart rate data obtained on a given day with available sleep metrics generated from the Fitbit was acceptable to include that day in the analysis. A window of 120 min was the best size to reduce noise in per-minute data and handle short missing episodes of data. Only the PSD24h/threshold of significance metric from heart rate was significant in distinguishing complications from no complications when using smaller window sizes of 10 and 15 min (P’s < 0.05, one-way ANOVA with repeated measures), and the rest of the key biorhythms did not show a difference (table S7). This did not affect the results of the early prediction model. We also found that the six key biorhythms were robust to different wear time criteria when the window size was 120 min (table S8).

DISCUSSION

Our study is the first to investigate biorhythms in children after surgery and transforming concepts adapted from other domains into practical outcomes for this clinical question. We found that biorhythms were a promising potential digital biomarker to evaluate postoperative recovery after appendectomy in children with complicated appendicitis. Specifically, a machine learning algorithm using wearable-derived biorhythms related to activity and heart rate predicted most complications up to 3 days before diagnosis, likely due to the delayed recovery of typical biorhythms in children who experienced complications after surgery. Most previous work (41, 47–51) has focused on the postoperative recovery of adults rather than the pediatric population. Compared with other studies using Fitbit (41, 47, 48), clinical electric health record (EHR) data (49, 51), or patient characteristics (50) to evaluate recovery in adults, such as after pancreatic surgery, sepsis, and cardiac surgery, our work showed a comparable or improved performance, with AUC-ROC values ranging from 0.71 to 0.85 (48–51) whereas ours was 0.86 [0.81, 0.90]. Using the same dataset in pediatric appendectomy, algorithm sensitivity was boosted by 10 to 15%, compared to those trained on wearable-generated daily summaries and daily statistical features extracted from per-minute wearable data in our previous studies (13, 43).

Our results suggest the value of monitoring biorhythms, specifically, circadian rhythms, in children across the inpatient and outpatient care settings. Although circadian mechanisms, including their regulation by the hypothalamic suprachiasmatic nuclei and their importance in coordinating rhythms of physiology and behaviors, are taught during medical training (29, 52, 53), evaluations of circadian rhythms are rarely encountered in clinical practice, perhaps due to a lack of feasibility or proof of value, until recently (54). Our study addresses these gaps by extracting metrics related to intraday patterns in activity and heart rate from a consumer wearable, thereby characterizing each child’s biorhythms and the strength of the circadian signal over time.

In addition to circadian rhythms, we observed that ultradian rhythms (<24 hours) became dominant in some cases. We suspect that factors such as pain and medication may cause short-term fluctuations in heart rate (55). However, due to the retrospective nature of this study, we lack high-resolution ground truth data to confirm the nature of these ultradian rhythms or determine why they emerge as dominant patterns. The urgency of pediatric appendectomy did not allow us to collect preoperative Fitbit data, thereby preventing comparisons of each patient’s biorhythms before and after surgery. Without preoperative data or longer-term follow up beyond the 21 days of postoperative monitoring, we cannot quantify individuals’ shifts in rhythms or draw conclusions like those in longitudinal studies of neurological diseases (34–36). Additionally, without preoperative data, we cannot investigate the impact of individual chronotypes on postoperative recovery or model predictions. Future work should consider these constraints when designing new investigational studies or interpreting results related to biorhythms.

We found that children who recovered with abnormal symptoms, but no diagnosed complication had activity biorhythms that resembled children with normative recovery, whereas their heart rate biorhythms oscillated between resembling normative recovery and recovery with complications. This differential behavior of activity and heart rate biorhythms may be helpful in distinguishing abnormal symptoms that do not progress to complications from true postoperative complications. Abnormal symptoms are often a source of unnecessary health care utilization following surgery, including unnecessary referrals to the emergency department, additional laboratory and imaging evaluations and may even necessitate inpatient admission for observation and diagnostic evaluation to rule out a complication (11). As such, the ability to differentiate an abnormal recovery event from a true complication would offer the potential for substantial cost savings and improved health care efficiency. With further validation and development, biorhythms could provide data-driven decision support for clinicians to remotely monitor abnormal symptoms with greater confidence and optimize additional care referrals for those more likely to signify true complications. There will be important logistical and operational considerations when implementing wearable data into the clinical workflow (e.g., integrating with EHR, interpreting algorithm output, and clinical alert system integration).

The understanding of surgical recovery has transformed from simple monitoring of physiologic endpoints (i.e., heart rate) to a more complex framework of multimodal data including metrics of physical activities (56), such as total step counts (57–59) and cadence (60), and innovative heart rate metrics (61). Our results suggest that activity and heart rate biorhythms both have values to the predictive model. Based on the result of the feature importance ranking by the ML model, the activity biorhythms are more important than heart rate biorhythms. This is consistent with our previous work (13, 43) using other features of activity, heart rate, and sleep where activity features were found to be more important to the machine learning algorithms.

In contrast to our findings, a previous study (47) applied similar methods for adults to monitor adults after pancreatic surgery and found that biorhythms did not play a strong role predicting readmissions. In our data, 89% of patients who had complications after discharge were readmitted and were all correctly predicted by the biorhythm model. A possible explanation for this seemingly different value of biorhythms is that children may have more structured lifestyles than adults. For example, the complex day-to-day engagements that impact adults’ circadian rhythms are less prevalent for children, who tend to follow an imposed schedule for daily activities, school, mealtimes, bedtime, etc., and chronobiology can shift over time, especially during puberty and adolescence (62). Thus, the typicality of daily biorhythms may be more evident in younger pediatric groups.

Last, our study used real-world high-resolution wearable data that were remotely obtained from children. Community wearable data collection suffers challenges of noncompliance and missing data due to improper device use, which is a principal challenge in pediatric studies (63–65). Existing studies often exclude days in which wear time is determined to be less than 10 hours in 1 day (45, 46). Biorhythm computation is more resilient to low wear times that result from frequent, short episodes of missing data because circadian patterns can still be modeled if the existing data are distributed throughout the 24-hour period. Using the sleep metrics generated by the Fitbit as an inclusion criterion ensures that the data at night are collected. It has been previously suggested to compute circadian rhythms from wearable device data sampled every 15 min (66). However, our study found a window size of 120 min was superior to all other window sizes evaluated, including a 15-min window. We conducted this analysis with the maximum window size of 180 min, because our modeling approach needed at least eight data points to fit the curve or analyze in the frequency domain. This approach can address the challenge of missing data in community data analysis, especially when daily heart rate or activity patterns are expected to be disrupted. Some potentially meaningful ultradian rhythms may be overlooked by the model due to the presence of short periods of missing data (e.g., 2 to 10 min). However, these data gaps are likely to be present in any real-world deployment of wearables, when the device could be removed periodically for a variety of reasons. In this retrospective study, we lack high-resolution ground truth data to further characterize missing data, and, in a remote community monitoring study, obtaining ground truth at this level of detail is impractical. It remains to be determined what a minimum acceptability of missing data is to maintain model performance. Future work could further explore the relationship between ultradian rhythms and postoperative complications in children, as well as the effect of varying data sparsity through synthetic or authentic addition and removal techniques.

For practical deployment of this model in real-world clinical settings, additional challenges must be considered. Now, wearable data and AI models have not been implemented for postoperative recovery monitoring in the real world. Knowledge of how wearable-based predictions may change clinical decision-making is limited; however, clinicians acknowledge the importance of continuous patient monitoring and report that they may be more confident in making treatment decisions with the assistance of a digital tool that offers continuous monitoring on patients (26). Future work should continue to validate this method in additional hospitals and its ability to prospectively identify complications following surgery must be evaluated before clinical decision support integration. For effective knowledge translation and clinical interpretation, clinician feedback must be gathered about how model alerts should be delivered to the patient care team and the relative acceptability of false positives for a model with high sensitivity to predict complications. Additional development may be needed to reduce instances of false positives. Furthermore, missing data are a major challenge in remote patient monitoring for real-world deployment. While a substantial advantage of this approach is that the model can tolerate a level of missing data, the robustness of the model to different levels of missingness should be tested further. Building on this work, additional biorhythm features can be engineered, such as by considering ultradian rhythms caused by pain or medication, or biorhythms can be combined with clinical, demographic, or other Fitbit data features to compare performance.

Our study has some limitations to be considered. First, the number of patients included in this study is relatively small, especially the number of patients with postoperative complications. The site where this study was conducted typically sees 13% readmission rates and 17% postoperative complication rates after appendectomy (67), compared to up to 38% in other hospitals (40). A critical next step will be to validate these findings externally, with patients at multiple institutions, and refine the model as needed. Second, variances of a single metric in patients with normative recovery were observed due to individual differences such as having good, moderate, or marginal normative recoveries. It is challenging to identify one metric that substantially evaluates postoperative complications to all normative recovery. To address this, we used machine learning algorithms to learn from multidimensional features from multimodal data. Future research will focus on improving the machine learning algorithms to reduce the number of false positives, which may occur for patients with the marginal normative recovery that are difficult to differentiate from patients with a postoperative complication. Third, the labeling strategy of defining the positive class as up to 3 days before a diagnosed complication enables early detection of complications while also accounting for potential delays between the complication onset and the formal clinical diagnosis. The question of how early a complication can be detected by wearables data is important to consider when deploying these tools in the real world. Future work will continue investigating the feasibility and accuracy of early detection for each type of complication, because abscesses and surgical site infections may develop across different timelines. Last, some of the clinical information used by the study relied on self-reports from the patients or their caregivers to confirm abnormal symptoms through phone call surveys on POD3, POD7, POD10, POD14, and POD21. Patients labeled as having the normative recovery may have abnormal symptoms that were not reported. Future studies plan to conduct daily questionnaires to minimize recall bias. Despite these limitations, we believe that our study provides valuable insights into these novel recovery metrics that can be extracted from consumer wearables to address key challenges in postoperative monitoring of children.

This study focused on the single clinical use case of pediatric appendectomy. More use cases should be investigated in future studies to evaluate the predictive value of biorhythms for other pediatric monitoring applications or even within different types of appendicitis, i.e., simple versus complicated appendicitis (68). For example, this methodology not only can be tested in other pediatric postoperative care scenarios, such as pectus excavatum (69), tonsillectomy (70), orthopedics (71), oncologic resections (72), and postoperative pain management (73), but also can aid in the diagnosis of other conditions which disrupt physiological and/or activity patterns including psychiatric (74) and neurodevelopmental disorders (32, 75), cardiovascular disease (76), and infectious diseases (77–84). Future work should also consider monitoring biorhythms in infants, who demonstrate regular ultradian rhythms during the development of circadian rhythms (85).

Our work aligns with the increasing interest in remote patient monitoring and digital medicine for the pediatric population. Various consumer wearable devices have been applied to children in and out of hospital (86, 87). However, due to previously described gaps, their clinical use remains limited at present. We expect that our methodologies not only will fill in these gaps but also will be translated to other pediatric and even adult clinical scenarios (41, 74, 77–79, 81–83, 88–94), which may broaden the clinical use of consumer wearable devices.

MATERIALS AND METHODS

Study design

We enrolled children aged 3 to 18 years old over 3 years who underwent laparoscopic appendectomy for complicated appendicitis at a tertiary children’s hospital between 2019 and 2022. This study was approved by the Institutional Review Board at Ann & Robert H. Lurie Children’s Hospital (IRB no. 2018-1836). Complicated appendicitis is defined as acute appendicitis with associated gangrene, perforation, or abscess (95). Eligible patients were identified through the EHRs with the exclusion criteria of any preexisting mobility limitations, comorbidities, or doctor-ordered physical activity limits of more than 48 hours post-surgery that would affect the postoperative recovery. Patients and/or their caregivers (parent or guardian) who did not speak English or Spanish were excluded because the cost and time associated with translation services.

A consumer wearable (Fitbit, the Inspire HR or Inspire 2) was given to the child to be worn on the wrist during a 21-day monitoring period beginning shortly after their surgery. Per-minute heart rate and step counts, as well as daily measures of total activity, sleep, and average resting heart rate were automatically uploaded to a cloud platform, Fitabase (96). These data were later downloaded from Fitabase for post-processing and analysis.

Clinical information was collected from the EHR, including abnormal symptoms or postoperative complications diagnosed by the child’s clinical team. Regular phone call surveys were conducted on POD3, POD7, POD10, POD14, and POD21 to evaluate the child’s recovery, including any new signs/symptoms that may have developed or health care utilization events. After reviewing the EHR and survey information, clinicians helped to determine the POD on which an abnormal symptom or postoperative complication was assigned.

Postoperative complications and recovery scenarios

Complications were defined by National Surgical Quality Improvement Program (97) and included diagnoses such as surgical site infection, intraabdominal abscess, small bowel obstruction, and Clostridium difficile infection. Abnormal symptoms during recovery included decreased urine output, emesis, diarrhea, drainage from a surgical incision, and postoperative ileus. These symptoms could be present for children with or without a confirmed postoperative complication; when there was no complication, these symptoms were resolved on their own or with minimal intervention.

We categorized each patient’s 21-day monitoring period as one of three recovery scenarios: normative recovery, recovery with abnormal symptoms that did not lead to a complication, and with postoperative complications. Patients with normative recovery were those who did not have any symptoms outside expectations for postoperative recovery or any need of additional intervention beyond standard treatment protocols (98). Patients with abnormal symptoms that did not lead to a complication were those who had one or more abnormal symptoms but no diagnosed complication. Patients with postoperative complications were those who had one or more complications, either with or without abnormal symptoms before or after the complication. These three recovery scenarios are illustrated in Fig. 3 (B to D).

Data preprocessing

We used per-minute heart rate data to determine whether the wearable was worn during the day and sleep metrics to determine if the device was worn during sleep. Thresholds for the minimum wear time requirement were established on the basis of the number of hours that the device was worn. Any day with data recorded for more hours than the specified threshold was included in the dataset for analysis. To ensure that the per-minute data span the 24-hour period in a day, we only included patients who had heart rate data and daily sleep metrics available for at least 1 day during the 21-day study period. We sampled the per-minute heart rate and step counts data with a moving average in the window size of 120 min, which resulted in up to 12 data points per day for heart rate (Fig. 4, right) and step counts (fig. S1, right).

Biorhythm metric extraction

Three methods were used to extract biorhythm metrics over each day (Table 1). First, after sampling per-minute data, circadian concepts from chronobiology were used to compute metrics such as the mean activity, relative amplitude, and intradaily variability (29). Second, the periodogram modeling method identifies dominant periods in discrete time-series data by estimating the spectral density of the data (99). Third, expecting 24 hours (circadian) to be the dominant period when a patient is homeostatic (29), the method of cosinor regression fits the data to a cosine curve with an assigned period of 24 hours and evaluates the goodness of the fit (100).

After reviewing literature on chronobiology (29, 62) and other studies using biorhythms for clinical monitoring, such as for neurological disease (31–33) or mental health (34–36), we included any relevant features supported by previous work that could also be computed using the available per-minute Fitbit data. In total, we extracted 31 biorhythm metrics from the consumer wearable, representing the existence, strength, and robustness of biorhythms including circadian and ultradian rhythms, of which 17 were computed with per-minute step counts depicting the activity biorhythm and 14 were derived from per-minute heart rate in beats per minute portraying the heart rate biorhythm. Changes in each metric over time can characterize circadian disruption and recovery, as summarized in Table 1. Equations used to compute the metrics are presented in table S1. Details of extracting each biorhythm metric can be found in the Supplementary Methods.

Predicting postoperative complications early

We combined normative recovery (Fig. 3B) and recovery with abnormal symptoms that did not lead to a complication (Fig. 3C) into one cohort, who recovered from the appendectomy without complications. To predict complications, we trained and tested a machine learning model (BRF) on the biorhythms extracted from each day. Because the treatment and resolution of these complications is a separate topic of investigation, data recorded in the days following a confirmed complication were excluded from both model training and testing (Fig. 5A). Because a goal of the model was to predict complications before their clinical diagnosis, we labeled the 3 days before their diagnosis as the positive class, enabling our model to learn from patterns in data during the onset of the complication (Fig. 5A). This labeling strategy is also consistent with our previous work, demonstrating that including the 3 days before a diagnosis for model training maximized the early detection performance (43).

To evaluate the prediction model, we used nested LOSO cross-validation and evaluated the model performance by comparing the model predictions for each day to the known label, specifically examining the model’s ability to predict a postoperative complication during the 3 days before and on the day of diagnosis. The details of the machine learning framework and its evaluation can be found in fig. S14. Model performance was summarized using the confusion matrix, sensitivity, specificity, MCC, AUC-ROC, and AUC-PR. To illustrate the variance in the model’s prediction capability, 95% CIs are calculated with a bootstrapping approach.

Given the imbalance of the dataset, a baseline dummy classifier, which randomly samples one-hot vectors from a multinomial distribution parametrized by the empirical class prior probabilities, was used to compare to our method. This dummy classifier does not consider any input features or addresses class imbalance.

Identifying important biorhythm metrics to monitor post-appendectomy recovery

We examined the ranked importance of biorhythm metrics provided by the trained BRF, comparing the relative importance of step counts and heart rate biorhythms, as well as the relative importance of the different modeling methods, using sum of importance for each of these metric groups. We computed the correlation between pairs of biorhythm metrics and removed the metrics that were highly correlated. Then, we identified the most important biorhythm metrics produced from BRF and selected the top ranked metrics per methodology for heart rate (physiological) and activity (step counts) biorhythms. Last, the significance of biorhythm metrics in relation to circadian disruption (Table 1) was evaluated for the application of pediatric post-appendectomy recovery.

Examining biorhythm changes

Clinically, not only are the days of diagnosed complications important, but also the days before a complication are meaningful for earlier identification of complications to mitigate their physiological impact for the patient. For patients who experienced complications, we created scatterplots to examine values of the most important biorhythm metrics on the day of and 3 days before the complication in relation to the mean and SD of these same metrics from patients without complications.

It takes time for children to recover from surgery to their baseline physiologic and activity levels (56, 60), and postoperative complications alter the recovery trajectory by prolonging the time a patient spends at abnormal levels (58). Therefore, understanding biorhythm changes over time, between different recovery scenarios is critical for guiding efforts to improve post-acute phase care (101). We compared the most important biorhythm metrics over the 21-day monitoring period between patients with versus without complications. For each biorhythm metric, the mean and SD were calculated by the group of patients with and without complications on each POD. We assessed each biorhythm metric on the difference between the two groups via one-way ANOVA with repeated measures.

Abnormal symptoms with no complications

Patients who experienced abnormal symptoms that did not lead to a complication were expected to exhibit fluctuating biorhythms, because their symptoms could affect activity and heart rate. These symptoms are usually not treated unless a complication is diagnosed, and they typically resolve on their own. Clinically, it is important to differentiate the patients from those with true complications, because these symptoms could lead to unnecessary health care utilization in the postoperative setting (11).

The 31 biorhythm metrics over the 21-day monitoring period were compared between patients with normative recovery, patients with abnormal symptoms that did not lead to a complication, and patients with diagnosed complications, via one-way ANOVA with repeated measures. To investigate the potential of differentiating recovery with abnormal symptoms that did not lead to a complication, we trained and tested a separate model that included abnormal symptoms as a separate class from normative recovery. We used the same machine learning framework to compare its performance to the two-class model predicting patients with diagnosed complications versus those with no complications.

Reliability of biorhythm metrics in the presence of missing data

To systematically assess the reliability in the presence of missing data, we considered different sampling window sizes and wear time requirements to extract the 31 biorhythm metrics. Sensitivity analyses on the sampling window size for preprocessing per-minute data and wear time requirements were performed to identify the optimal window size and wear time requirement. We conducted analyses on the 21-day monitoring period comparison via one-way ANOVA with repeated measures for each metric and then compared LOSO results from the early prediction model via paired t tests. For each metric, we also examined the relationship between values computed from the available data after imposing different wear time requirements and sampling window sizes using Pearson’s correlation coefficient. Correlation values (0.9 ≤ $\mid r\mid$ ≤ 1) were considered very highly correlated (102).

Statistical analysis

Data were processed, analyzed, and modeled using Python 3.10. An F test based on the model sum of squares, and sum of square residuals was used to assess the cosinor regression model fit. Comparisons of differences in types of recovery scenarios were performed with one-way repeated measures ANOVA when data were normally distributed or with Mann-Whitney U test otherwise. A Shapiro-Wilk test was performed to check for normality. Two-sample t test was used to assess whether the difference between the means of the number of days with wearable data available for the groups with and without complications is significantly different. Paired t tests were performed to assess the difference of results of the machine learning model on early prediction of postoperative complications. Pearson’s correlation was used to assess the correlations. The statistical significance level was set to 0.05.

Supplementary Materials

This PDF file includes:

Supplementary Methods

Figs. S1 to S14

Tables S1 to S8