The adoption of consumer-grade wearable activity trackers into routine use has been increasing rapidly in recent years, with approximately 1 in 5 adults in the United States reported to regularly use wrist-worn smartwatches and fitness trackers in 2019 [1]. This phenomenon has generated an unprecedented scale of consumer health data and led to many studies on the wider health uses of such data. These studies are increasingly generating evidence to reveal relationships between recordings from wearable activity trackers and the risk for conditions ranging from mental health and infectious diseases [2,3] to cardiovascular and metabolic (collectively referred to as cardiometabolic) diseases [4-7]. Among these, owing to the apparent links between activity levels and cardiometabolic health, the evidence for broader health uses of wearables is most established in the cardiometabolic domain [4,8-11].

Previous studies in the cardiometabolic domain have focused on the utility of wearable-derived summary statistics, and fall into 1 of 2 categories. First, electrocardiogram signals from wearables have been studied in relation to the development of cardiometabolic conditions, such as atrial fibrillation [12-14], hyperkalemia [15,16], and heart failure [17-19]. As many of these conditions are amenable to early intervention via dietary changes or increased physical activity, there is also an interest in using wearables to promote self-awareness and regulation [20] and to enhance screening [11]. Second, wearable-derived measures, such as circadian measures, sleep patterns and quality [11,21], step counts [4], wearable-derived resting heart rate [4,8,10,21,22] and heart rate variability [23-27] have been found to correlate with outcomes in cardiometabolic disease. As such, there is increasing recognition in the clinical community to incorporate wearable-derived measures into practical cardiometabolic disease management [6,28].

Rapid and ongoing developments in consumer wearable technology are enabling ever-richer measurements with finer temporal resolution for heart rate, activity, and sleep dynamics in free-living states [6,29,30]. Principled analyses of such data streams could generate new insights beyond summary statistical measures for cardiometabolic health and disease management. However, the analysis of time series data recorded in free-living states is challenging, as these data tend to exhibit real-world noise and fluctuations and typically lack important physical and physiological contexts. A few recent studies have used black-box deep neural networks to relate high-resolution heart rate and step count time series recorded using wearables to the risk of developing atrial fibrillation, sleep apnea, and hypertension [31,32]. As their primary goal focused on risk target classification, the nature of the intermediate predictive time series features and their connection with known clinical and biological markers of cardiometabolic disease remains unresolved.

In this study, we aimed to derive high-resolution digital phenotypes from consumer wearable heart rate recordings and to examine their associations with diverse risk markers for cardiometabolic disease. Specifically, we sought to develop a time series feature extraction approach, contextualized by activity state, to meaningfully represent heart rate dynamics recorded by consumer wearables in free-living conditions. We then applied our approach to multidimensional data from normal volunteers in the SingHEART study [33] to assess the extent to which the derived high-resolution wearable features could predict expressed clinical risk markers for cardiometabolic disease. Furthermore, we assessed whether these high-resolution features also represent more subtle physiological changes associated with an inherent genetic predisposition to cardiometabolic disease. Finally, we qualitatively characterized these wearable phenotypes in volunteers who had actualized clinical events to assess connections beyond risk markers to manifest cardiometabolic diseases.

We sourced data from the SingHEART study (NCT02791152) as of October 8, 2019. Enrollment targeted healthy volunteers who provided written informed consent to use the data (including electronic health records) for research. Participants were required to fulfill the inclusion criteria presented in Textbox 1.

At enrollment, each participant was profiled using a range of health assessment modalities. The resulting data set included (1) heart rate and step count time series recordings over 3 to 5 days from consumer wearable devices (Fitbit Charge HR), together with the associated sleep logs generated by Fitbit, (2) self-reported answers to a lifestyle and quality-of-life questionnaire [4], (3) genotypic data from whole genome sequencing using the Illumina HiSeq X platform, and (4) laboratory measurements for 9 clinically relevant markers (systolic and diastolic blood pressure; blood levels of triglycerides, total cholesterol, high-density lipoprotein, and low-density lipoprotein; fasting blood glucose level; waist circumference and BMI). As of October 8, 2019, the full study cohort contained 1101 participants, of whom 692 (62.8%) participants had wearable recordings. We focused on this subset of participants for subsequent analysis: a detailed breakdown of the data is provided in Table 1.

Furthermore, we also tracked each participant for the occurrence of any actual clinical event. We extracted all clinical codes (based on the International Classification of Diseases, 10th Revision) pertaining to any acute care use events in the regional health system associated with the National Heart Centre Singapore until January 2021 to characterize the links among data features, risk markers, and actual clinical events.

The SingHEART study (NCT02791152) was established at the National Heart Centre Singapore, a tertiary specialty hospital in Singapore, and was approved by the SingHealth Centralized Institutional Review Board (ref: 2015/2601 and 2018/3081) [33,34].

Given a time series segment, it is possible to define a set of high-resolution features using approaches such as the highly comparative time series analysis [35,36] and time series feature extraction on the basis of scalable hypothesis [37,38]. However, such approaches can generate many redundant features, and the process of selecting a concise but effective representation is often not straightforward. A recent study [39] introduced a minimally redundant and interpretable set of 22 features, termed as Canonical Time-series Characteristics 22 (Catch22) features, which have high predictive value across 93 diverse time series classification data sets. As this Catch22 feature set was designed to reduce interfeature redundancy, it provides a compendious representation of the different dynamic properties of the time series.

The Catch22 features fall into seven main categories, namely (1) distribution, (2) extreme events, (3) symbolic, (4) linear autocorrelation and periodicity, (5) nonlinear autocorrelation, (6) successive differences, and (7) fluctuation analysis. The distribution-based features represent summary statistics of the distribution of the measured values in the series (while ignoring the chronological order of these values). The extreme event features represent intervals between successive outlier events in the time series. The symbolic features represent statistics summarizing the outputs of symbolic transformations of the actual time series values. The linear autocorrelation and periodicity features comprise summary statistics on inherent periodicities in the time series. The nonlinear autocorrelation features involve summary statistics on periodicities based on nonlinear transformations of the time series. The successive difference features represent statistics based on the time series of the incremental differences. Finally, the fluctuation analysis features quantify the statistical self-affinity of the time series. Detailed descriptions of each of the 22 features are provided in Table S1 in Multimedia Appendix 1.

We now describe the steps to derive resting heart rate, summary statistics on activity and sleep patterns, and high-resolution features from the wearable heart rate and step count time series recordings. As all these physiological features are derived from the same recordings, they are internally consistent and can be meaningfully used for downstream comparative analyses.

We examined how high-resolution wearable-derived heart rate features from sleep, active, and sedentary segments were distributed across study participants. Figure 2 illustrates the empirical distributions of exemplar features drawn from segments corresponding to each of the 3 physical activity states. To examine the variability across participants, we also visualized representative time series at the 2.5th, 25th, 50th, 75th, and 97.5th percentile of the density.

The first example comprises a nonlinear autocorrelation feature (CO_trev1_num, quantifying the time-reversibility statistic <(x_t+1-x_t)³>_t) that relates to the degree of spikiness or regularity in the wearable-based heart rate time series (Figures 2A-2C). The second example comprises a distribution feature (DN_HistogramMode_5, corresponding to the mode of the z-transformed values) that quantifies the degree of nonnormality of the time series values by representing the difference between the most probable values (mode) and the mean of the series (Figures 2D-2F).

To better understand the nature of wearable-derived time series features, we investigated their associations with genomic risk markers for cardiometabolic disease. As probing these associations requires handling diverse multidimensional data types with potentially complex nonlinear relationships, we used a machine learning framework (similar to the one described earlier) to model these relationships. We then used model performance measures to infer the degree of information overlap between wearable features and genomic risk targets. As genomic risk is independent of age, we did not include age in any of the models considered.

We categorized the genetic susceptibility to cardiometabolic diseases using polygenic scores (PGSs). To define the genomic risk for lipid abnormalities, blood pressure abnormalities, and obesity, we used the PGS Catalog [49] to identify relevant polygenic risk scores corresponding to the 3 targets. Specifically, we identified 14 PGS for lipid abnormalities (PGS000060, PGS000061, PGS000062, PGS000063, PGS000065, PGS000115, PGS000192, PGS000309, PGS000310, PGS000311, PGS000340, PGS000677, PGS000688, and PGS000699), 2 for blood pressure abnormalities (PGS000301 and PGS000302), and 1 for obesity (PGS000298). Additional details of the selection process are provided in Section SI-3 (Multimedia Appendix 1).

For each of the 3 targets, we labeled a participant as having high genomic risk if their scores for any of the relevant PGS were in the top or bottom decile (refer to Section SI-3, Multimedia Appendix 1 for how the direction of a PGS is determined), which we term as the 90/10 cut-off. For instance, the high genomic risk group for lipid abnormalities would include members with high-risk scores for at least 1 of the 14 lipid-related PGS. The modeling of these targets and statistical comparison of the performance of different model types were identical to the earlier process described for the clinical risk targets.

To evaluate the sensitivity of the chosen percentile cut-offs for genomic risk scores, we repeated the above analyses for 2 additional sets of cut-offs, namely the 80/20 and 85/15 cut-offs.

Finally, we examined the connections between high-resolution wearable-derived features and actualized cardiometabolic disease events for participants not in our training set of 321. Among these participants, we considered those who actualized cardiometabolic disease events indicated by a primary diagnosis of cardiovascular disease, dyslipidemia, and hypertension (as per International Classification of Diseases, 10^th Revision codes listed in Table S3 in Multimedia Appendix 1). As this set of events spans a broad range of cardiometabolic conditions, anyRISKoutof9 is the closest surrogate marker. Hence, we chose to focus our profiling on the wearable-derived feature set that was most strongly associated with anyRISKoutof9.

For participants selected per the abovementioned criteria, we examined demographic information, physical measurements, genomic risk of disease, and clinical risk markers alongside the wearable-derived features. To interpret how the different wearable-derived features contribute to the model predictions at the individual participant level, we computed the Shapley values (Φ) [50] of each feature using the iml package [51] in R and selected the 5 features with the highest absolute magnitude of Φ for each participant. We illustrate the profiles of the participants, the predictions made by the best-performing model for anyRISKoutof9, and the features that contribute most to these predictions for each selected participant.

All statistical analyses and modeling were performed using R Statistical Software (version 4.0.3; R Core Team 2020). Computation of resting heart rate was performed using R, but all other feature engineering efforts such as annotation of wearable time series recordings and derivation of summary features, as well as the generation of high-resolution features, were performed using Python (version 3.8.6).

All Python and R codes used in feature generation are available in Multimedia Appendix 2.

Unlike summary statistics such as resting heart rate, which averages heart rate measurements across multiple days, our high-resolution feature sets provide more granularity on the heart rate time series dynamics during different physical activity states (sleep, active, and sedentary). Figure 3A illustrates the distributions of the high-resolution wearable feature values across the 321 participants (colored according to their respective activity states). Although the Catch22 algorithm was identically applied to each of the 3 activity segments, we observed that each feature exhibited distinct distributions across the 3 different activity states.

To study whether this difference holds at the participant level, we characterized the correlations among the high-resolution feature sets obtained during the 3 different activity states. For any given feature (eg, CO_trev1_num), we considered vectors of feature values for each physical activity state across the population (eg, CO_trev1_num.active, CO_trev1_num.sedentary, and CO_trev1_num.sleep). We then calculated the Pearson correlation between these feature vectors for each pair of the activity states. This analysis revealed that the feature values from the different activity states were poorly correlated (Figure 3B). In fact, the largest absolute correlation coefficient among any of the pairs was 0.15. Taken together, these findings indicate that heart rate dynamics from different activity states contain distinct information.

Having gained some intuition about the information contained within the wearable-derived feature sets, we considered their predictive value for the clinical markers of cardiometabolic disease risk. Specifically, we trained random forest models to use the different wearable-derived feature sets to classify each of the 4 cardiometabolic disease risk targets. We performed comparative analyses to evaluate the predictive value of the different wearable-derived feature sets for classification of the 4 cardiometabolic disease risk targets.

First, we compared the OOB performance of the models trained using different feature sets for each clinical risk marker target (Table 4). For each target, the best-performing model was based on one of the high-resolution wearable feature sets (HighRes.ActiveSeg, HighRes.SedenSeg, or HighRes.SleepSeg). Specifically, for anyRISKoutof9, the HighRes.SedenSeg model was the best-performing model, with 17.9% and 7.36% lower Brier scores than baselines based on age and gender and resting heart rate, respectively (P<.001 in each case). This finding highlights the predictive value of high-resolution information within wearable heart rate time series recordings.

Second, we observed that heart rate dynamics extracted from different activity level segments have differential predictive potential for the various targets, as evidenced by the statistically significant differences between Brier scores (P<.001) of the HighRes.ActiveSeg, HighRes.SedenSeg, and HighRes.SleepSeg models (Table 4). Of the 3 model types, HighRes.SedenSeg performs best for lipid abnormalities, obesity, and anyRISKoutof9, whereas HighRes.ActiveSeg performs best for blood pressure abnormalities.

Third, to comparatively evaluate contributions from individual wearable-derived features, we trained models that used all features available to predict each cardiometabolic disease risk target and ranked the variable importance in each case. Figure 4 shows the variable importance plots. It is clear that different features affect the performance of the models for each of the 4 targets. For instance, age and gender are the top 2 drivers of model performance for the anyRISKoutof9 target but are not among the top 10 features for both lipids_abnormal and obesity targets. Furthermore, we found that heart rate dynamics from different activity states contained distinct information on cardiometabolic disease risk. For example, the DN_HistogramMode_5 feature from the sedentary and active segments was important for predicting cardiometabolic disease risk markers but the DN_HistogramMode_5 feature from the sleep segment was not (Figure 4).

Fourth, we observed that the top 10 features for each of the 4 targets included features from all 6 feature types (age and gender, wearable-derived resting heart rate, wearable summary statistics, and the 3 sets of high-resolution features from Table 1). This suggests that risk prediction models using wearable-derived features may not exclusively rely on only one of the different feature sets or any one feature drawn from these feature sets. Rather, a collection of different wearable-derived high-resolution heart rate features from distinct activity states is essential for accurately predicting the multiplicity of cardiometabolic disease risk targets.

To further interpret the information contained within the wearable-derived features, we sought to understand how they relate to the genetic predispositions for cardiometabolic diseases. Specifically, we examined the degree of information overlap between the different wearable-derived features (Table 1) and the genomic risk of cardiometabolic conditions. For each pairing between the different wearable-derived feature sets and the 3 genomic risk targets, we trained random forest models and used their Brier scores as indirect measures of the strength of the associations.

The results are presented in Table 5. For each of the 3 abnormality types, we observed that the high-resolution wearable features were more strongly associated with genomic risk levels than sex and resting heart rate (improvement of 11.9%-22.0% in Brier scores; P<.001). We highlight that the trends against baseline and resting heart rate were relatively insensitive to the polygenic risk score threshold used to define high versus low genomic risk (Section SI-4, Multimedia Appendix 1). These results suggest that, in comparison with standard measures, high-resolution features from wearables are better able to represent subtle physiological dynamics related to the genomic risk for cardiometabolic disease.

Finally, we examined the relationship between the wearable-derived feature set most predictive for anyRISKoutof9 and actualized cardiometabolic events. We focused on participants not in our training set and filtered participants with data for the feature set most predictive for anyRISKoutof9 (ie, Catch22 [Sedentary] feature set, based on the abovementioned results). This yielded 197 candidate participants for illustrative profiling. Among these participants, only 5 participants actualized events with primary diagnoses for cardiometabolic conditions (as specified in Table S3 in Multimedia Appendix 1).

Table 6 provides demographic, genetic, and clinical risk profiles along with physical measurements and important wearable features for these 5 participants (A-E). All the participants were aged 54 to 61 years. Of the 5 participants, 4 (80%) were male. Only 1 (20%) participant was obese. We now present the findings on the predictive value of high-resolution wearable-derived features for these participants.

First, we describe participants with abnormalities in both genetic and clinical risk markers, namely participants A and B. Participant A had high genomic risk for all 3 conditions, presented abnormal values for most of the 9 clinical risk markers, and was also diagnosed with all 3 types of cardiometabolic conditions considered (cardiovascular disease, dyslipidemia, and hypertension). Participant B had a genomic risk for lipid and blood pressure abnormalities, abnormal lipid panel values, and a clinical diagnosis of dyslipidemia. While participant A had a wearable-derived resting heart rate slightly above the population average, participant B had a wearable-derived resting heart rate lower than the population average. However, in both cases, our HighRes.SedenSeg model predicted a positive anyRISKoutof9 outcome.

Second, we considered participants with no genomic risk but who presented with abnormal clinical risk markers, namely participant C. This participant had high blood pressure, abnormal cholesterol and blood glucose levels, a clinical diagnosis of dyslipidemia, and wearable-derived resting heart rate slightly above the population average value. However, we noted that our HighRes.SedenSeg model predicted a negative anyRISKoutof9 outcome. This could be due to modeling error or possibly be attributed to the absence of severe changes in heart rate dynamics given the normal genetic background and moderate wearable-derived resting heart rate value.

Third, we highlighted participants who did not exhibit any abnormalities in clinical risk markers and were borderline for cardiometabolic disease risk, namely participants D and E. Participant D only had a genomic risk for blood pressure. Participant E, on the other hand, appeared to have the most benign profile with low genomic risk for all 3 target conditions and normal values for all 9 clinical risk markers (with only the BMI being borderline high). Both participants had wearable-derived resting heart rate values that were lower than the population average. Although participants D and E had a seemingly low-risk profile by standard measures, they had clinical diagnoses of dyslipidemia and cardiovascular disease, respectively. Indeed, our HighRes.SedenSeg model predicted a positive anyRISKoutof9 outcome in each case.

Finally, inspecting the most important features (top 5 Shapley values) contributing to model predictions for anyRISKoutof9 in Table 6 reveals interesting patterns. While age and gender were (expectedly) consistent contributors to prediction scores for most participants, many Catch22 (Sedentary) features also contributed at comparable levels. For instance, DN_Histogram_Mode_5 was important for all 5 participants, whereas CO_Embed2_Dist_tau_d_expfit_meandiff and DN_OutlierInclude_p_001_mdrmd were important for 3 and 2 participants, respectively. In particular, DN_Histogram_Mode_5 was an important feature for most participants in this study. This feature takes on large values when the participant’s heart rate time series exhibits substantial deviations from the mean, which could occur when there are sustained or frequent oscillations with high amplitude. Although such deviations may be common in active states, their presence in sedentary states could forebode cardiovascular abnormalities, as was the case for these 5 participants. Beyond the consistent features noted above, there are other diverse high-resolution features among the top 5 most important contributors for different participants. This suggests that our high-resolution feature extraction approach offers a compact but sufficiently diverse set of predictive heart rate patterns, including those that are consistent across individual participants and those that can cater to participant-to-participant variability. Detailed Shapley Additive Explanations (SHAP) feature importance plots for each participant are provided in Section SI-5 in Multimedia Appendix 1.

Consumer wearables enable the recording of rich high-resolution physiological dynamics in free-living conditions, but how these data relate to health and disease is not fully understood. We introduced a principled framework to derive high-resolution heart rate features from consumer wearable recordings, and applied our approach to a data set containing multidimensional cardiometabolic health parameters from healthy volunteers. Our results show that, in comparison with typical summary statistics, high-resolution features resolving temporal dynamics and activity-dependent patterns in heart rate have stronger associations with modifiable risk markers and inherent genetic predispositions for cardiometabolic disease alike. Our findings imply that these high-resolution digital phenotypes from consumer wearables can provide a more granular picture of cardiometabolic health and disease states, which could have potential use in cardiometabolic health screening and disease management.

Our framework addresses key challenges in mining wearable data recorded in free-living conditions. Unlike clean data from controlled experimental settings, real-world wearable recordings tend to be irregular, contain missing stretches [29], lack clean context annotations, and have variable lengths. As such, analyses based on the naive application of general purpose time series feature extraction methods [36,39,52] may not have ecological validity [53]. To address this gap and derive meaningful physiological dynamics from wearable time series recordings, our feature extraction framework standardizes handling of data irregularities and encodes contextual information about the underlying activity level and physiological state (Figures 1-3). This conceptual framework, although demonstrated here with the Catch22 method [39], is agnostic to the choice of the feature representation method for time series data [36,37]. Furthermore, in contrast to black-box feature learning methods based on large labeled data sets [31], our approach yields more interpretable time series features with smaller unlabeled data sets.

Our framework provides many possibilities for gaining new insights from wearable recordings. Our analyses, using multimodal wearable, genomic, and clinical data from healthy volunteers, highlight 2 possibilities.

First, our results revealed new relationships between high-resolution heart rate dynamics from wearables and the risk of cardiometabolic disease. Most previous studies correlated clinically obtained measures of heart rate dynamics, such as heart rate variability, exercise capacity, and heart rate recovery, with disease risk or outcomes [54-56]. In contrast, our results revealed that heart rate dynamics recorded by consumer wearables, when processed appropriately, are also predictive of cardiometabolic disease risk (Tables 4-6; Figure 4). Furthermore, we found that heart rate dynamics from different activity states contain distinct information about specific cardiometabolic conditions (Table 4; Figures 3 and 4). For example, heart rate patterns in sedentary states are more related to abnormalities in lipid levels and obesity, whereas those in active states may be more related to abnormalities in blood pressure readings (Table 4). These findings highlight the value addition of assessing physiology in free-living activity states (beyond controlled clinical settings) for disease risk monitoring and management [57].

Second, our study provides new perspectives on the interrelations between wearable recordings and genetic predispositions in cardiometabolic diseases. Although there has been a longstanding interest in probing gene-lifestyle interactions and their additive effects on cardiovascular disease [58-60], such studies have had limited visibility on physiology in free-living conditions. We found surprising connections (Table 5) between high-resolution wearable-derived feature sets and genetic predispositions for cardiometabolic disease. As these associations did not appear to depend on the presence or absence of manifest clinical risk markers, we posit that high-resolution phenotypes from wearables may capture subtle subclinical physiological changes stemming from latent predispositions to disease.

Although the uniquely multimodal nature of our data enables us to uncover many novel insights on high-resolution wearable phenotypes, limitations of data set size and cohort design present some challenges. First, it was infeasible to conduct full-scale gene-environment interaction studies [61-63]; or train state-of-the-art machine learning models with large feature sets. Second, as the risk of cardiometabolic disease is highly multifactorial, the limited visibility on relevant physical and lifestyle factors constrains the absolute predictive accuracy of all models presented. For instance, we had limited input on regular exercise habits as the observation span was less than a week, as well as limited overlap between key lifestyle indicators and wearable recordings (eg, only 9 participants who smoked had valid wearable records). Finally, as our study included only a small number of participants with actualized cardiometabolic events, we could not perform quantitative analyses to relate wearable phenotypes with clinical events. Future work based on larger cohorts [64] with more targeted study designs could address some of these limitations and enable cross-cohort validation of our current findings.

In conclusion, we demonstrated that high-resolution digital phenotypes based on heart rate patterns in wearable recordings provide important insights into physiology in free-living conditions. Our results revealed that these measures are associated with both genetic and clinical risk markers of cardiometabolic disease and have additional predictive value beyond wearable-derived summary statistics and clinical measures of cardiometabolic health. Hence, our work expands possibilities to use digital phenotypes from consumer wearables as readily accessible indicators of cardiometabolic health and disease and motivates new approaches for quantitative scoring of cardiometabolic disease risk. Future studies could expand our findings to even higher resolution digital phenotypes that can be extracted from recordings with newer generations of wearable devices [65,66] and target evaluations for precision screening, health monitoring, and disease management applications.