可穿戴传感揭示日常驾驶相关的心脏激活结构
Wearable sensing reveals the structure of cardiac activation associated with everyday driving
可穿戴传感结合多模态移动感知技术,通过持续自然环境下的生理监测,刻画了日常驾驶行为中反复出现的心脏激活模式。驾驶时的心率升高源于认知负荷和环境不确定性而非代谢消耗,为理解重复性日常行为的心血管反应提供了新测量框架。
Introduction
Many everyday behaviors are neither rare nor extreme, but are encountered repeatedly over weeks, months, and years. Even when the physiological response to an individual episode is modest, repeated exposure may give rise to substantial cumulative physiological activation simply because the behavior occurs so frequently. Quantifying these repeated physiological exposures has historically been difficult because it requires continuous physiological monitoring together with synchronized information about behavioral and environmental context under naturalistic conditions1,2.
Recent advances in wearable and multimodal mobile sensing, together with the emergence of digital phenotyping as a framework for continuous naturalistic measurement, have substantially changed this situation34. Long-duration physiological monitoring is now feasible outside the laboratory, making it possible to characterize repeated physiological exposure associated with ordinary daily-life behaviors rather than isolated laboratory tasks5. Among continuously measurable physiological variables, heart rate has emerged as the most widely available marker in naturalistic settings because it is routinely acquired by modern wearable devices.
The key challenge now lies not in measuring heart rate, but in interpreting what it represents. Heart rate measured during everyday life is influenced by both stable individual characteristics and changing behavioral and environmental conditions. Large longitudinal studies have shown substantial between-person variability in resting heart rate while also demonstrating that individuals maintain characteristic physiological baselines despite day-to-day fluctuation6. Personalized wearable models likewise show that cardiovascular responses depend on both the individual and the surrounding context7. Consequently, the same observed heart-rate value may carry different physiological meaning across individuals and situations. This raises a fundamental methodological question: how can repeated context-associated physiological activation be quantified while accounting for stable individual physiology?
Driving provides an informative test case for this measurement problem. It is common, repeated, and often prolonged, particularly in car-dependent regions. As a behavioral context, driving may involve sustained vigilance, risk monitoring, time pressure, frustration, and environmental uncertainty while differing substantially from other low-motion daily-life activities. Unlike physical activity, where elevated heart rate primarily reflects the metabolic demands of bodily exertion and can be evaluated using established laboratory and clinical exercise-testing paradigms8, elevated heart rate during naturalistic driving occurs in a low-motion context and reflects a different combination of behavioral and environmental influences. Cardiac activation associated with everyday driving has no standard clinical analogue. It is not defined by a single maximal or structured episode, but by ordinary observations repeated across days under changing behavioral and environmental conditions. Continuous naturalistic monitoring is therefore well suited to characterizing the cardiac operating regime associated with driving, although it cannot by itself isolate the causal effect of driving from all other contextual differences.
Research on the physiology of driving has primarily examined physiological responses to acute stressors, elevated cognitive demand, and other transient operational states. On-road and simulator studies have shown that heart rate increases with increasing driving-related cognitive demand9,10,11. Other work has demonstrated that different classes of driving stressors produce distinct patterns of physiological arousal and behavioral response12. A systematic review of acute physiological stress responses to driving found that studies had examined a range of autonomic and cardiac responses, while concluding that evidence for a consistent acute stress response remained limited and heterogeneous13. More recent work has continued this emphasis through investigations of stress, workload, hazard anticipation, and related operational states using cardiac and multimodal physiological measurements1415.
More recent work has extended operational-state assessment through multimodal physiological sensing and machine-learning methods. In driving, systems integrating cardiovascular and neurophysiological measurements have been developed for fatigue detection, while related research in other safety-critical domains, such as air traffic control, has explored explainable artificial-intelligence frameworks for physiological workload assessment1617. However, these studies remain primarily concerned with transient-state detection, workload classification, or immediate physiological responses. Considerably less attention has been given to characterizing the cardiac operating regime associated with everyday driving within an individual’s broader daily-life physiology.
A related limitation appears in the broader wearable-analytics literature. Research on wearable cardiovascular prediction and physiological state assessment has expanded rapidly, encompassing personalized prediction, multimodal sensing, and explainable artificial intelligence18,19,20,21. Much of this work, however, emphasizes prediction, classification, or workload detection rather than physiological interpretation across repeated naturalistic behavioral contexts. Although recent work on explainable wearable analytics has highlighted the importance of interpretability21, the primary objective has remained understanding model predictions rather than interpreting the physiological organization of heart-rate responses themselves. For questions involving repeated everyday driving, prediction alone is insufficient. The scientific need is to distinguish stable baseline physiology, context-associated cardiac activation, and residual modulation in a way that supports physiological interpretation across time horizons.
Together, these literatures leave a specific methodological gap. Wearable and digital-phenotyping research increasingly enables continuous physiological measurement in daily life, while driving research has largely focused on acute stress and workload responses. What remains lacking is an interpretable framework that separates stable baseline physiology from the reproducible cardiac activation associated with the driving context itself. Such a decomposition is necessary to determine how much of the observed heart-rate elevation during driving reflects baseline physiology, how much reflects a driving-associated context offset, and how repeated context-associated activation accumulates under recurring driving schedules.
In this study, we address this gap using the validated NUBI II dataset and a baseline-centered decomposition framework tailored to longitudinal naturalistic multimodal sensing. NUBI II extends our previously reported NUBI I dataset4 from naturalistic driving alone to broader waking-life monitoring. Specifically, whereas NUBI I contained only driving data, NUBI II includes multimodal observations from driving, non-driving sedentary behavior, and physically active periods, along with a larger cohort and richer covariate coverage. This broader daily-life coverage allows driving to be interpreted not as an isolated task episode, but as a recurring behavioral exposure embedded within each participant’s observed physiological week.
Within this framework, instantaneous heart rate is decomposed into a participant-specific baseline, a context-specific offset, and additional modulation by observed covariates. We refer to the empirical elevation above participant baseline statistically associated with the driving context as the driving offset. The corresponding quantity in the non-driving sedentary stratum is termed the non-driving sedentary offset. These offsets describe statistical associations rather than causal effects. Physical-activity periods are retained to characterize the broader daily-life physiological ecology but are excluded from the primary comparison because they represent a fundamentally different physiological context. We further examine whether the relationship remains stable under weekday and weekend partitions, which substantially alter the composition of non-driving sedentary behavior.
Because cardiac activation associated with driving can be represented as a context-specific elevation above baseline, this framework can be used to illustrate how modest minute-level differences accumulate under repeated exposure schedules. Such projections are intended as illustrative scenario analyses rather than estimates of long-term physiological burden. Real-world changes often alter the duration or frequency of an existing physiological demand rather than introducing a fundamentally different one; for example, a short daily commute may become hours of professional driving. Thus, beyond explaining instantaneous heart-rate variation, the framework also provides a way to explore cumulative baseline-referenced activation under repeated exposure schedules. Accordingly, this study asks three questions:
-
(1)
Is everyday driving associated with a distinct baseline-referenced cardiac operating regime relative to non-driving sedentary behavior within each participant’s broader daily-life physiology?
-
(2)
How much of observed heart-rate variation is explained by participant-specific baseline, a context-specific offset, and additional modulation by weather, traffic-related conditions, workload, and individual traits?
-
(3)
Can this decomposition support illustrative long-horizon scenario analyses?
Answering these questions matters beyond the immediate comparison of driving and non-driving sedentary behavior. More broadly, the contribution of this paper is not merely a model for predicting heart rate, but an interpretable framework for separating participant-specific baseline physiology, context-associated cardiac activation, and residual modulation. By grounding long-horizon scenario analyses in observed naturalistic physiology, the framework provides a bridge from short-timescale wearable measurements to the characterization of repeated physiological exposure while preserving the distinction between statistical association and causal effect.
Results
Sample and data overview
We begin by characterizing the analyzed NUBI II sample, the distribution of observations across behavioral contexts, and the empirical support for the main covariates used in the later modeling analyses. Participants recruited in Texas were monitored continuously for one week as part of an IRB-approved naturalistic study. The dataset combined wearable, smartphone, and vehicle data with contextual information such as weather and traffic density, together with psychometric and workload measures including the Big Five and NASA-TLX. Observations were organized into three validated contexts—DRIVING, NONDRIVING_SEDENTARY, and PHYSICAL_ACTIVITY—with the primary comparison focused on driving versus non-driving sedentary behavior.
At the 60-s resolution, the analyzed dataset comprised 57 participants, 393 participant-days, and 225 calendar days, for a total of 177,318 rows and 2,955.3 hours of observation (Table 1). The sample was approximately balanced by sex (30 male, 27 female) and consisted predominantly of young adults (age: 26.2 ± 5.0 years, mean ± SD). Observations were strongly imbalanced across behavioral contexts. DRIVING accounted for 18,487 rows (10.4%) spanning all participants, 379 participant-days, and 222 calendar days, whereas NONDRIVING_SEDENTARY accounted for 153,120 rows (86.4%) spanning all participants, 392 participant-days, and 225 calendar days. PHYSICAL_ACTIVITY was much sparser, contributing 5,711 rows (3.2%) from 44 participants across 102 participant-days and 88 calendar days. This imbalance is also visible in Fig. 1c, where most participants contributed predominantly NONDRIVING_SEDENTARY data, with substantial individual variation in the fraction of driving exposure. Figure 1a, b further shows that the cohort was heterogeneous rather than tightly clustered around a narrow behavioral profile. Participant-level distributions of the Big Five traits were broad, although conscientiousness tended to be shifted toward the higher end of its range. NASA-TLX subscales also showed broad spread, particularly for temporal demand, mental demand, effort, and frustration. By contrast, self-rated performance was shifted toward higher values, which is consistent with the uneventful nature of the monitored trips: participants generally completed the drives without crashes or traffic citations. Thus, the TLX results suggest that participants commonly viewed their driving performance as successful, while still differing substantially in the demands and effort they experienced during the trips.
a Participant-level distributions of the Big Five personality traits. b. Participant-level distributions of NASA-TLX subscales. c Participant-level exposure mix across DRIVING, NONDRIVING_SEDENTARY, and PHYSICAL_ACTIVITY. d Participant-level distributions of median raw heart rate, \({\widetilde{HR}}_{{{{\rm{raw}}}},i}\), and participant-level baseline heart rate, HRbase,i. e Distribution of observations across early-day (AM) and late-day (PM) periods. f Distribution of weather conditions during DRIVING intervals. g Distributions of selected driving-related variables. h Distribution of hand-energy rotation during DRIVING intervals.
The physiological overview in Figure 1d motivates a baseline-referenced representation of cardiac activation. Participant-level median raw heart rate values, denoted \({\widetilde{HR}}_{{{{\rm{raw}}}},i}\), were generally higher than participant-level baseline heart rate values HRbase,i, indicating that observed daily life often involved sustained elevations above baseline. Here, HRbase,id denotes the participant-day baseline heart rate available in the dataset, and HRbase,i denotes the participant-level average of those daily baseline values. Accordingly, the analyses that follow treat heart rate relative to participant-specific baseline rather than in raw form alone.
The remaining panels of Figure 1 summarize the empirical support of the contextual covariates used later in the analysis. For Figure 1e, we define Early-Day as midnight to noon and Late-Day as noon to midnight. Most observations fell in the late-day period, and the same pattern was also present within DRIVING, indicating that monitored driving activity was concentrated in the second half of the day. Figure 1f then summarizes weather specifically during DRIVING intervals. Driving weather was dominated by cloudy and clear conditions, with smaller but still meaningful support for rain and an aggregated adverse-weather category. Thus, favorable driving weather was most common, but less frequent conditions remained sufficiently represented for analysis, including 603 DRIVING rows in the adverse-weather grouping. Figure 1g shows substantial spread in driving dynamics, including ATP, FF, and speed. ATP denotes Absolute Throttle Position and tracks vehicle acceleration demand, whereas FF denotes free flow and can be interpreted as the road’s speed rating. Hand energy rotation (Figure 1h) was strongly right-skewed, consistent with many low-to-moderate steering episodes and fewer high-rotation events. Together, these distributions indicate that the dataset captures a broad and ecologically realistic range of participant profiles, daily timing, environmental conditions, and driving dynamics.
Data completeness was generally favorable for key predictors, although missingness varied across variable classes and contexts (Table 2). Among DRIVING rows, most kinematic and vehicle variables had negligible missingness, typically below 1%. The collapsed trip-level variables also showed complete coverage: trip ID, trip distance, and trip duration were all observed for the 973 identified driving trips. Baseline heart rate missingness was modest at the participant-day level (7.9% overall), and NASA-TLX subscales, computed at the participant-day-period level restricted to periods with driving, each showed 8.9% missingness. The main exception was HRraw during DRIVING, where missingness reached 13.6%, compared with 3.5% in NONDRIVING_SEDENTARY and 2.3% in PHYSICAL_ACTIVITY. This somewhat higher missingness during DRIVING is unsurprising, as the mobile platform was acquiring not only heart rate but also concurrent vehicle and steering-related signals, making data collection in that context technically more challenging.
To evaluate whether the higher heart-rate missingness observed during driving could bias the physiological analyses, we performed the missingness analyses described in the Methods. A mixed-effects logistic regression confirmed that heart-rate measurements were more likely to be missing during driving than during non-driving sedentary behavior after adjustment for day period, day type, weather, and participant-level clustering. However, when the analysis was restricted to driving observations, heart-rate missingness was not significantly associated with vehicle speed, traffic congestion (Jam Factor), wrist linear acceleration energy, or wrist rotational acceleration energy. Thus, although missingness was more frequent during driving, the available data provide no evidence that it preferentially affected objectively more demanding driving segments. Detailed mixed-effects logistic regression estimates for both analyses are provided in Supplementary Table S1.
Finally, the structure of the driving data reflects repeated naturalistic episodes rather than isolated events. Across the dataset, we identified 973 driving trips, with a median trip duration of 15.0 minutes. This supports the view that the observed cardiovascular signal reflects repeated everyday driving exposures embedded within much longer periods of sedentary and non-driving behavior.
Overall, the exploratory analysis yields three conclusions that motivate the modeling results that follow. First, participants differed substantially in trait profiles, workload reports, and exposure composition. Second, the dataset was strongly dominated by NONDRIVING_SEDENTARY periods, making context-aware modeling essential. Third, the physiology summaries support a baseline-referenced treatment of heart rate, while the contextual panels show that the main temporal, weather, and driving-dynamics covariates are empirically supported in the observed data.
Baseline-referenced cardiac activation
To illustrate the baseline-referenced interpretation at the individual level, Fig. 2 shows a representative week-long trace for one participant (participant-level traces for all subjects are provided in the Supplementary Information). The figure displays the available participant-day baseline values, HRbase,id, together with the observed heart-rate signal, HRraw,itd. Physical-activity segments produce the largest departures from the participant-day baseline. During most driving periods, heart rate remains well above this baseline, whereas during non-driving sedentary behavior it more often fluctuates around it. Thus, the physiological interpretation of a raw heart-rate value depends on both its elevation above baseline and the behavioral context in which it occurs. Although Figure 2 displays the participant-day baseline, HRbase,id, all subsequent baseline-referenced analyses use the participant-level baseline, HRbase,i, obtained by averaging the available daily baseline values for each participant. Accordingly, baseline-referenced cardiac activation for each observation is defined by equation (1):
$${{{{\rm{NHR}}}}}_{itd}=H{R}_{{{{\rm{raw}}}},itd}-H{R}_{{{{\rm{base}}}},i}.$$
(1)
where i, t, and d index participant, time point, and day, respectively, and HRbase,i denotes the participant-specific baseline heart rate.
Raw heart rate for participant P62 across study days 1–7 (Monday–Sunday) at 60-s resolution, shown together with the participant-day baseline, HRbase,id, and behavioral context labels. Baseline-referenced analyses use the participant-level baseline HRbase,i, defined in the Methods.
Figure 3 summarizes the empirical organization of baseline-referenced cardiac activation during driving and non-driving sedentary behavior. The three columns show the results for all observations combined and separately for weekdays and weekends. The weekday/weekend stratification was motivated by participant enrollment interviews, which indicated that participants were predominantly office workers who generally worked during weekdays and had weekends off. Although neither calendar category represents a homogeneous behavioral context, this partition provides a useful robustness analysis because it changes the mixture of sedentary activities against which driving is compared. Figure 3a shows the distributions of the underlying 60-s NHR observations. In all three calendar strata, the driving distribution was visibly narrower and more concentrated above zero than the corresponding non-driving sedentary distribution. However, the participant-level comparisons in Table 3 show that this pattern did not arise from a higher mean level of baseline-referenced activation. Mean NHR was similar during driving and non-driving sedentary behavior overall (11.2 versus 11.5 bpm, p = 0.613), on weekdays (11.1 versus 11.6 bpm, p = 0.467), and on weekends (12.5 versus 11.3 bpm, p = 0.313).
Normalized heart rate (NHR) was calculated for each 60-s observation as the observed heart rate minus the participant-specific physiological baseline. Columns show results for all observations combined (Overall) and separately for Weekdays and Weekends. a Distributions of the underlying 60-s NHR observations during driving and non-driving sedentary behavior. b Participant-level percentage of observed time with NHR > 0, indicating time spent above the participant-specific baseline. c Participant-level standard deviation of NHR, representing within-participant variability in baseline-referenced cardiac activation. d Participant-level frequency of returns to baseline per observed hour. A return to baseline was defined as a transition from NHR > 0 to NHR≤0 between consecutive available 60-s observations within the same participant, behavioral context, and calendar stratum. In Panels b–d, each metric was first calculated separately for each participant from the corresponding 60-s observations and was then summarized across participants, preventing participants with more observations from receiving greater weight. Bars show cohort means of the participant-level values, and error bars indicate the corresponding 5th–95th percentile ranges. Gold denotes driving and gray denotes non-driving sedentary behavior.
The principal distinction instead concerned the persistence and variability of activation. Overall, participants remained above their physiological baseline for a mean of 93.7% of observed driving time, compared with 81.3% during non-driving sedentary behavior (p < 0.001; Figure 3b). Within-participant NHR variability was also substantially lower during driving (7.98 versus 12.40 bpm; p < 0.001; Figure 3c), and returns to baseline occurred less frequently (1.31 versus 2.46 events per observed hour; p < 0.001; Figure 3d). Thus, driving was not characterized by a significantly higher participant-level mean NHR, but by activation that remained above baseline more consistently, varied within a narrower range, and returned to baseline less often.
The same operating-regime pattern was evident in both calendar partitions. On weekdays, driving was associated with more time above baseline (93.8% versus 82.1%), lower within-participant variability (7.88 versus 12.11 bpm), and fewer returns to baseline (1.34 versus 2.51 events per observed hour); all paired comparisons yielded p < 0.001. On weekends, the corresponding differences were 94.7% versus 78.4% for time above baseline, 5.99 versus 11.93 bpm for within-participant variability, and 1.01 versus 2.28 events per observed hour for returns to baseline; again, all paired comparisons yielded p < 0.001.
This calendar robustness is important because weekday and weekend sedentary periods likely contain different mixtures of occupational, domestic, social, and leisure activities. If the observed driving pattern were primarily an artifact of one particular sedentary comparison group, the relationship would be expected to change appreciably between the two partitions. Instead, driving retained the same defining characteristics in both: a greater proportion of time above participant-specific baseline, reduced within-participant variability, and fewer returns to baseline, despite no significant difference in mean NHR.
Taken together, Figure 3 and Table 3 show that driving occupies a distinct and persistent baseline-referenced cardiac operating regime. The principal difference from non-driving sedentary behavior lies not in average activation magnitude, but in its temporal organization: driving-associated activation remains above baseline more continuously, fluctuates within a narrower range, and returns to baseline less often.
This distinction also defines the scope of the subsequent analyses. We focus on the two low-motion behavioral contexts, DRIVING and NONDRIVING_SEDENTARY, because baseline-referenced cardiac activation in these contexts can be interpreted primarily in relation to cognitive, emotional, contextual, and other everyday influences. Observations annotated as PHYSICAL_ACTIVITY were excluded from the primary comparison because heart rate during physical activity reflects both exercise-related and non-exercise-related activation, making the contribution of interest difficult to isolate.
Predictive decomposition of heart rate
The preceding analyses established that everyday driving occupies a distinct baseline-referenced cardiac operating regime. Although participant-level mean NHR did not differ significantly from non-driving sedentary behavior, driving was characterized by greater persistence above physiological baseline, lower within-participant variability, and fewer returns to baseline, a pattern that remained consistent across weekday and weekend partitions (Figure 3; Table 3). These operating-regime characteristics describe how cardiac activation is organized during driving but do not explain why activation varies from one driving episode to another. We therefore next decompose baseline-referenced heart rate into a constant driving-related component and additional modulation associated with behavioral, environmental, and individual covariates.
Specifically, we represent instantaneous heart rate as the sum of a participant-specific baseline, a context-specific offset, additional covariate-based modulation, and residual variation (Fig. 4). Thus, the predictive target in this section is instantaneous raw heart rate HRraw,itd, while the baseline-referenced formulation serves to clarify how the predictive signal is organized. Because each participant’s baseline HRbase,i is known, predictions of raw heart rate can be converted directly into predicted baseline-referenced cardiac activation according to equation (2):
$${\widehat{{{{\rm{NHR}}}}}}_{itd}={\widehat{HR}}_{{{{\rm{raw}}}},itd}-H{R}_{{{{\rm{base}}}},i}.$$
(2)
which is the quantity of primary physiological interest for the operating-regime analyses and the subsequent long-horizon scenario projections. Equivalently, after subtracting the participant-specific baseline, the remaining signal can be written in baseline-referenced form as equation (3):
$${{{{\rm{NHR}}}}}_{itd}={{{{\rm{HR}}}}}_{{{{\rm{raw}}}},itd}-{{{{\rm{HR}}}}}_{{{{\rm{base}}}},i}={\tau }_{c}+\phi ({X}_{itd})+{\epsilon }_{itd},$$
(3)
where i, t, and d index participant, time point, and day; τc denotes the mean empirical offset statistically associated with context c; ϕ(Xitd) captures additional modulation by observed covariates; and ϵitd denotes residual variation. Thus, once participant-specific baseline differences are removed, the remaining predictive structure consists of a context-specific offset, finer-grained covariate-based modulation, and residual error.
Instantaneous heart rate is represented as the sum of a participant-specific baseline, a context-specific empirical offset, additional covariate-based modulation, and residual variation.
To determine how much each component contributes to prediction, we decomposed heart-rate prediction into three nested stages within each behavioral stratum and evaluated each stage out of sample using a nested grouped cross-validation framework, with performance estimated from the outer cross-validation folds: (i) a baseline-only model, (ii) a baseline-plus-offset model, and (iii) a full elastic-net (ENet) model. The baseline-plus-offset model corresponds to the participant-specific baseline and context-specific offset components shown in Figure 4. In the DRIVING stratum, we refer to this quantity as the driving offset. In the NONDRIVING_SEDENTARY stratum, we refer to the corresponding quantity as the non-driving sedentary offset. Each offset represents the mean elevation above participant-specific baseline statistically associated with observations in that stratum; neither is interpreted as a causal effect of the behavioral context.
Figure 5 and Table 4 show a clear hierarchy of predictive gains under the nested grouped cross-validation framework. Here, ΔRMSE denotes the change in root-mean-square prediction error relative to the baseline-only model, and Δr2 denotes the corresponding change in the squared Pearson correlation. In both DRIVING and NONDRIVING_SEDENTARY, the largest reduction in RMSE occurs when moving from the baseline-only model to the baseline-plus-offset model, as shown in Figure 5a. In DRIVING, RMSE decreases from 14.24 to 8.88 bpm, whereas the full ENet model yields only a further reduction to 8.77 bpm. In NONDRIVING_SEDENTARY, RMSE decreases from 17.53 to 13.35 bpm with the offset term, followed by a smaller further reduction to 13.04 bpm with ENet. Thus, most of the improvement in absolute prediction error is achieved by incorporating the context-specific offset within each behavioral stratum. An RMSE of about 8.8–8.9 bpm in DRIVING indicates moderate minute-level prediction error in a naturalistic wearable setting: the model does not recover every short-timescale fluctuation, but it does recover the dominant context-dependent operating level of heart rate substantially better than baseline alone. Quantitatively, the offset term accounts for approximately 97.9% of the total RMSE improvement in DRIVING and 93.2% in NONDRIVING_SEDENTARY, with the remaining gain attributable to the richer covariate set.
Bars show mean performance across the five outer cross-validation folds; error bars represent standard errors across folds. a Root-mean-square error (RMSE) for the baseline-only, baseline-plus-offset, and ENet models in each behavioral stratum. Most of the reduction in absolute prediction error comes from adding the mean context-specific elevation above participant baseline, termed the driving offset in DRIVING and the non-driving sedentary offset in NONDRIVING_SEDENTARY. b Squared Pearson correlation (r2) for the same models. Gains in r2 arise primarily from the full ENet model, indicating that the covariates mainly explain residual within-stratum variation rather than shifting the overall predicted level. The corresponding conventional out-of-sample \({R}_{{{{\rm{OOS}}}}}^{2}\) values for the final ENet models are reported in Table 4.
The decomposition of r2 leads to a complementary conclusion. As shown in Figure 5b and summarized numerically in Table 4, the baseline-plus-offset model does not increase the correlation-based r2 relative to the baseline-only model, whereas the full ENet model produces a modest but consistent gain in predictive performance. In DRIVING, the correlation-based r2 rises from 0.486 to 0.511, while the corresponding conventional out-of-sample coefficient of determination is \({R}_{{{{\rm{OOS}}}}}^{2}=0.480\). In NONDRIVING_SEDENTARY, the correlation-based r2 rises from 0.231 to 0.259, while the corresponding conventional out-of-sample coefficient of determination is \({R}_{{{{\rm{OOS}}}}}^{2}=0.246\). This pattern indicates that the offset term primarily corrects the overall level of heart rate, thereby reducing absolute error, while the covariates captured by ENet explain additional within-stratum variation.
Taken together, these results suggest that heart rate is organized by two components. The first is a dominant context-specific offset, reflecting the mean elevation statistically associated with observations in the corresponding behavioral stratum. The second is a smaller modulation component, reflecting additional structured variation around that context-dependent level. This structure is present in both strata, but predictive performance is consistently stronger in DRIVING, consistent with driving being a more physiologically constrained and behaviorally structured state than heterogeneous NONDRIVING_SEDENTARY periods.
Non-physiological modulators beyond baseline and context-specific offset
We next asked what explains residual variation in heart rate after accounting for the two largest sources of coarse predictive structure in the data: person-specific baseline and the context-specific offset. At this stage, the candidate modulators were non-physiological observed variables, including participant-level traits and demographics, temporal and contextual factors, and, in the driving stratum, movement- and vehicle-related features, including engineered short-term summaries of steering and driving dynamics. Thus, the analysis asks how non-physiological characteristics of the person, situation, and task explain residual cardiovascular variation beyond participant-specific baseline and the mean offset associated with the behavioral stratum.
As shown in Fig. 6, this remaining signal is not homogeneous across strata. Instead, the predictors that matter beyond baseline and offset differ substantially between driving and non-driving sedentary behavior. In the driving stratum, grouped importance was dominated by day period (AM vs. PM) and recent hand rotational movement, followed by sex, other movement-derived features, study day, and driving weather. This pattern indicates that residual heart-rate variation during driving is shaped primarily by proximal, state-dependent factors tied to the immediate demands of the task. In the non-driving sedentary stratum, by contrast, the leading grouped predictors were openness, age, extraversion, neuroticism, conscientiousness, time of day, sex, day number, and agreeableness, pointing to a broader residual structure in which dispositional and temporal factors play a larger role.
a Grouped feature importance within each behavioral stratum, showing that residual variation during driving is dominated by immediate task-linked factors, particularly day period (AM vs. PM), recent hand rotational movement, and driving weather, whereas residual variation during non-driving sedentary behavior is more strongly associated with broader dispositional and temporal variables, including openness, age, extraversion, neuroticism, conscientiousness, day period, sex, study day, and agreeableness. b Signed standardized coefficients for selected interpretable predictors, showing comparatively modest directional effects in driving, including cloudy and adverse-weather indicators, and clearer directional structure in the non-driving sedentary stratum. c Partial effect curves for the top continuous modulator in each stratum: recent hand rotational movement (1-min mean) during driving and openness during non-driving sedentary behavior. The y-axis shows the change in predicted heart rate relative to the model reference level. d Incremental predictive gain of the ENet relative to the baseline-plus-offset model, expressed as improvement in root-mean-square error (RMSE) and correlation-based r2. Gains are positive in both strata but larger in non-driving sedentary behavior, indicating that residual physiological variation outside driving is more readily captured by additional covariates. ENet, elastic net; AM, ante meridiem; PM, post meridiem; RMSE, root-mean-square error.
The signed coefficients reinforce this contrast. In driving, interpretable effects were comparatively sparse and modest, though study day, trait anxiety, and weather-related indicators such as cloudy and adverse-weather conditions showed directional structure. This is consistent with a regime in which cardiovascular demand is already elevated and more tightly constrained by the task itself. In non-driving sedentary behavior, several interpretable predictors showed clearer directional structure, especially openness, agreeableness, age, sex, and study-day indicators, suggesting that outside the driving task, baseline-referenced heart rate remains more sensitive to a wider range of person-linked modulators. This contrast is also visible in Figure 6c, where the dominant continuous modulator in driving was recent hand rotational movement (1-min mean), whereas in the non-driving sedentary behavior stratum it was openness.
Importantly, these additional predictors improved prediction beyond the baseline-plus-offset model in both strata, but the gain was substantially larger in non-driving sedentary behavior (Figure 6d). This pattern is theoretically coherent. Driving is a relatively constrained behavioral state associated with a persistently elevated and comparatively narrow cardiac operating level, so once baseline and the context-specific offset are accounted for, less structured variance remains to be explained. Non-driving sedentary behavior, in contrast, aggregates a more heterogeneous set of low-movement situations, leaving more room for person-level and temporal covariates to improve prediction. Thus, the value of the ENet is not only that it adds predictive accuracy, but that it reveals two distinct patterns of residual structure: a more task-proximal one in driving, and a broader trait- and day-structured one in non-driving sedentary behavior.
Taken together, these results suggest that baseline-referenced heart rate is organized hierarchically rather than governed by a single generic process. The dominant structure consists of a participant-specific baseline together with the context-specific offset, reflecting the mean elevation statistically associated with observations in that behavioral context. Once these two components are accounted for, the remaining explainable variation depends strongly on behavioral stratum: residual variation within the driving stratum is dominated by immediate task-state modulation, whereas non-driving sedentary residuals retain broader person- and temporal structure.
Long-horizon scenario projections
The preceding analyses established that baseline-referenced cardiac activation is organized hierarchically. Participant-specific baseline and the context-specific offset account for most of the predictable structure in heart rate, while observed participant-, task-, and environment-related characteristics explain additional residual variation around that operating level. We next ask what these minute-level differences imply when the associated behavioral context is encountered repeatedly over long periods.
Because the fitted ENet model predicts baseline-referenced cardiac activation at the minute level, it can be used to construct illustrative long-horizon scenario projections under specified repeated driving schedules. These projections are not intended as forecasts of cardiovascular outcomes or estimates of the causal physiological effect of driving. Rather, they illustrate how modest differences in predicted minute-level activation accumulate mathematically under repeated behavioral exposure.
To illustrate this capability, we considered trait anxiety, which emerged as one of the interpretable participant-level modulators in the preceding analysis (Figure 6a, b). Using the final ENet model fit to the complete driving dataset with hyperparameters selected by nested cross-validation, we generated two scenario-specific prediction sets by applying the model to all observed driving epochs while fixing the trait-anxiety variable to the lowest (STAI = 20) and highest (STAI = 61) observed participant values. Each observed driving epoch retained its participant-specific baseline together with all other observed task-, environmental-, and participant-related predictors. Thus, trait anxiety was the only model input changed between the paired counterfactual predictions. The STAI trait-anxiety scale ranges from 20 to 80, with higher scores indicating greater trait anxiety. Minute-level predicted baseline-referenced cardiac activation was then accumulated under three illustrative daily driving schedules: a 30 min/day commute, a 2 hr/day commute, and an 8 hr/day professional-driving schedule (Fig. 7a; Table 5).
a Illustrative long-horizon scenario projections of baseline-referenced cardiac activation. Annual cumulative activation was projected by applying the final ENet model to all observed driving epochs while fixing trait anxiety to the lowest and highest observed participant values (STAI = 20 and STAI = 61, respectively). Each observed driving epoch retained its participant-specific baseline together with all other observed task-, environmental-, and participant-related predictors, so trait anxiety was the only model input changed between the paired counterfactual predictions. Projections are shown for three illustrative daily driving schedules (30 min/day, 2 hr/day, and 8 hr/day). Error bars show the empirical 5th–95th percentile participant-bootstrap intervals. b Normalized RMSE as a function of temporal aggregation. Prediction error was recomputed after averaging the original 1-min observations and their corresponding 1-min predictions over progressively longer contiguous windows. RMSE values were normalized to the corresponding 1-min RMSE within each behavioral stratum. In the DRIVING stratum, normalized RMSE remained approximately constant across aggregation horizons, whereas in the NONDRIVING_SEDENTARY stratum it decreased as short-timescale residual variation averaged out.
In the full-sample fitted model, the higher trait-anxiety profile yielded greater projected baseline-referenced activation than the lower trait-anxiety profile. The predicted minute-level difference was modest (0.99 bpm). In a leave-one-participant-out (LOPO) sensitivity analysis, the contrast remained positive in all 55 non-boundary refits, with an empirical central 95% interval of 0.61–1.56 bpm. Two additional refits yielded boundary solutions in which regularization reduced the trait-anxiety coefficient to zero and were reported separately. Scaling the full-sample point estimate under repeated exposure yielded projected annual differences of approximately 180 NHR-hours/year for a 30 min/day commuting schedule, 721 NHR-hours/year for a 2 hr/day schedule, and 2,886 NHR-hours/year for an 8 hr/day professional-driving schedule (Table 5); the corresponding empirical 95% LOPO intervals were 112–284, 447–1,136, and 1,788–4,543 NHR-hours/year.
The long-horizon projections illustrate how a modest modeled minute-level contrast can scale into progressively larger cumulative differences under repeated exposure, but should be interpreted as illustrative scenarios rather than evidence of a trait-anxiety effect. More broadly, this illustration is not specific to trait anxiety. Trait anxiety serves only as one example of a participant-level characteristic identified by the predictive decomposition. The same framework can be applied to any observed modulator or combination of modulators represented in the predictive model, allowing investigators to explore how different behavioral, environmental, or participant profiles translate into different cumulative physiological exposure under specified repeated driving scenarios. In this way, the decomposition framework extends naturally from explaining minute-level cardiac activation to exploring its long-horizon implications.
Long-horizon stability of prediction error
The illustrative long-horizon scenario projections presented above are obtained by averaging the model’s minute-level predictions and scaling the resulting mean according to assumed daily driving duration. This procedure does not require prediction at horizons longer than one minute. Rather, it assumes that the average of the minute-level predictions remains a reliable estimate of average cardiac activation when accumulated over repeated driving exposure. That assumption could be compromised if the prediction residuals contained persistent temporal structure arising from unmeasured participant-, task-, or environment-related influences, because such structure would not necessarily average out over longer periods. We therefore examined how prediction error behaves when the original 1-min observations and their corresponding 1-min predictions are averaged over progressively longer contiguous windows extracted directly from the naturalistic recordings, ranging from 1 to 60 min (Figure 7b). Horizons beyond 60 min were not considered because the available recordings contained too few complete continuous segments to support stable estimation.
The resulting horizon profile indicates that prediction error remained stable over the examined aggregation windows. In the DRIVING stratum, normalized RMSE showed only minor fluctuations and no systematic increase as the averaging horizon increased. Thus, averaging minute-level predictions over progressively longer contiguous windows did not reveal evidence of persistent residual effects that became more pronounced with increasing temporal aggregation. In the NONDRIVING_SEDENTARY stratum, normalized RMSE decreased steadily with increasing horizon, indicating that a substantial portion of the remaining residual variation reflects short-timescale fluctuations that average out over longer observations. Together, these findings indicate that temporal aggregation does not amplify residual prediction error, supporting the use of averaged minute-level predictions as the basis for the long-horizon scenario projections described in the preceding subsection.
Discussion
This study examined whether everyday driving is associated with elevated heart rate above participant-specific baseline in naturalistic life and whether heart rate can be decomposed into interpretable components that support longer-horizon exposure projections. Three main findings emerged. First, driving was associated with a more persistent baseline-referenced cardiac operating regime, spending substantially more time above participant-specific baseline than non-driving sedentary behavior. Second, across both the driving and non-driving sedentary strata, most of the predictable structure in heart rate was captured by two coarse components: participant-specific baseline and a context-specific offset. Additional situational and person-level covariates improved prediction but primarily explained residual within-context variation. Third, in the illustrative trait-anxiety scenario, the full-sample model showed that a modest minute-level difference in predicted baseline-referenced activation could accumulate into sizeable differences in projected annual physiological exposure; the contrast remained positive across all non-boundary leave-one-participant-out refits.
The first major contribution of the study is empirical. Objective characterization of sustained physiological activation during everyday life has remained challenging because it requires continuous physiological measurements together with behavioral context under naturalistic conditions. Our results help address that challenge by showing that driving, relative to non-driving sedentary behavior, was associated not merely with transient increases in heart rate but with a more persistent baseline-referenced cardiac operating regime. Importantly, this association was observed under naturalistic conditions rather than in a simulator or short laboratory task. In that sense, the study extends prior driving research, much of which has focused on acute workload, hazard anticipation, or state classification, by demonstrating that everyday driving was associated with a sustained baseline-referenced cardiac operating regime embedded in daily life. Importantly, this qualitative relationship remained consistent when the analyses were repeated separately for weekdays and weekends, despite differences in the mixtures of sedentary activities represented in those calendar partitions. This robustness supports the interpretation that the observed driving-associated operating regime is not dependent on a particular composition of the sedentary comparison group.
At the same time, the findings sharpen the interpretation of cardiac activation in wearable data. Raw heart rate alone is difficult to interpret across people because the same absolute value can reflect different physiological states depending on the individual. Participant-specific baseline therefore plays a foundational role in interpretation. Using resting heart rate as the physiological reference aligns with the objective of quantifying sustained cardiac activation relative to each participant’s resting state rather than another behavioral context. The exploratory analyses showed pronounced between-person heterogeneity in both baseline and raw heart-rate distributions, and the predictive analyses confirmed that baseline alone explains a substantial share of heart-rate structure. This is consistent with prior work showing substantial and stable between-person differences in resting heart rate.6 More broadly, it reinforces a key point for digital phenotyping and wearable health analytics: physiologically meaningful inference in everyday life often requires centering signals relative to the individual rather than relying on raw values alone.
A second major contribution is conceptual. The results suggest that heart rate in naturalistic daily life is usefully understood as layered rather than unitary. In the present data, the dominant organization was not fine-grained moment-to-moment fluctuation, but the combination of a stable individual baseline and a context-specific offset above that baseline. This structure matters because it distinguishes two scientific questions that are often blended together: what determines the operating level of physiology in a given context, and what determines shorter-timescale deviations around that level. In our analyses, the context-specific offset accounted for most of the reduction in absolute prediction error, whereas additional covariates mainly improved explained variance. In substantive terms, the model was better at capturing where cardiac activity tends to sit under a given everyday context than at reproducing every short-term fluctuation. For the present scientific aim, that is a strength rather than a weakness, because the central question concerns sustained activation rather than second-to-second state tracking.
The present framework intentionally estimates a common context-specific offset rather than participant-specific offsets. A fully personalized offset would absorb much of the stable between-participant variation in context-specific operating level, leaving substantially less variation to be explained by demographic, psychometric, and contextual covariates. Such individualized calibration addresses a different objective: maximizing within-person prediction after calibration. In contrast, the goal of the present study is to characterize the average driving-associated operating level across previously unseen participants and explain systematic deviations from that population-level relationship.
This layered view also helps explain the contrast between driving and non-driving sedentary behavior. Driving showed stronger predictive performance overall and a residual structure dominated by immediate task-proximal factors such as day period, recent rotational movement, and weather. Non-driving sedentary behavior, by contrast, showed weaker overall predictability and a broader residual structure involving age, personality traits, sex, and temporal factors. One plausible interpretation is that driving represents a more constrained behavioral regime, involving a comparatively coherent set of attentional, perceptual, and motor requirements and, in the present observational data, a more stable cardiac operating level once baseline was taken into account. Non-driving sedentary behavior, in contrast, aggregates a wider range of low-motion situations and therefore leaves more room for person-level and temporal factors to shape residual variation.
More broadly, these results illustrate why interpretability matters. A purely predictive model could have been judged only by accuracy, but that would have obscured the scientific meaning of the signal. By separating baseline, offset, and residual modulation, the present framework makes it possible to ask not only whether heart rate can be predicted, but what kind of structure is doing the predictive work. In that sense, the framework may be useful beyond driving, as a general strategy for distinguishing stable physiology from average contextual loading and finer-grained modulation in other naturalistic domains.
The long-horizon simulations demonstrate how relatively small minute-level differences can accumulate into substantial differences in projected physiological exposure under repeated daily schedules. This illustrates a broader principle in wearable physiology: persistent modest activation may be more consequential over time than isolated short-lived responses.
Cumulative NHR-hours should therefore be interpreted as a physiological exposure metric rather than as a validated clinical endpoint. Epidemiological studies have established prognostic associations for elevated resting heart rate and reference ranges for real-world heart rate, but these frameworks do not provide clinically validated thresholds for baseline-referenced cardiac activation integrated over time22,23,24,25. The purpose of NHR-hours is to summarize cumulative activation under repeated exposure and enable quantitative comparisons across behavioral scenarios and individual profiles. Whether this quantity predicts subsequent cardiovascular outcomes remains an open empirical question requiring prospective longitudinal validation; establishing such exposure–outcome relationships would be necessary before clinically meaningful reference ranges or thresholds could be defined.
Equally important, the horizon-scaling analysis indicates that residual prediction error does not increase when observations are aggregated over longer time windows. In driving, normalized residual error remained broadly stable as the averaging horizon increased, whereas in non-driving sedentary behavior it declined. Thus, over the observational horizons represented in the data, temporal aggregation did not degrade agreement between window-averaged observations and predictions; if anything, it improved robustness in the more heterogeneous non-driving sedentary context. This is encouraging for digital-health applications because many practical questions concern prolonged exposure patterns rather than minute-to-minute fluctuations. Accordingly, the framework supports interpretable long-horizon scenario analyses, while the horizon-scaling analysis should be understood as characterizing residual behavior under temporal aggregation rather than validating the scenario projections themselves.
Some limitations should be noted. First, the cohort consisted of a moderate number of young adult drivers observed over a one-week period. Although the data were dense and naturalistic, the sample does not represent all age groups, driving populations, or health profiles. The results therefore should not be generalized automatically to older adults, professional drivers as an occupational population, or individuals with cardiovascular disease. Second, the study is observational. The driving offset and the simulation contrasts should not be interpreted causally. Driving differs from non-driving sedentary behavior along numerous measured and unmeasured dimensions, including posture, time of day, social context, anticipation, destination urgency, caffeine intake, trip purpose, preceding activities, passenger presence, and the reasons for entering each context. Although the analyses account for several measured factors—including temporal context, weather, traffic-related conditions, driving dynamics, workload, personality traits, and anxiety—residual confounding and context-selection effects remain possible. The driving offset therefore represents an empirical above-baseline association observed in this dataset, not the physiological cost caused by driving. Third, the long-horizon projections necessarily depend on the assumed exposure schedules and on the assumption that the observed statistical associations remain stable beyond the one-week observation period. They should therefore be interpreted as illustrative scenario-based activation estimates rather than as forecasts of any specific individual’s future cardiovascular outcome.
These limitations point directly to future work. An important next step is to evaluate the framework in larger and more diverse cohorts, including broader age ranges, different geographic settings, and populations with clinically relevant cardiovascular risk. Larger and more heterogeneous datasets would also provide denser coverage of participant characteristics and behavioral contexts, enabling the model to learn population-level physiological structure more effectively and potentially improving predictive performance and generalizability. Future studies could also enrich the framework by incorporating additional wearable markers of autonomic and cardiovascular regulation, such as heart-rate variability, sleep-related recovery indices, and multimodal physiological burden measures, together with richer contextual information about the social and behavioral environment, including passenger presence and other currently unobserved aspects of driving. Another important direction is longitudinal extension. Future studies spanning months or years of passive monitoring could determine whether participant-specific calibration provides additional benefit for individualized forecasting relative to the population-level framework developed here. Such data would also make it possible to examine how repeated contextual loading evolves with life transitions, seasonal conditions, occupational demands, aging, or behavioral interventions. More broadly, the framework may have relevance beyond driving by providing a general approach for converting short-timescale wearable measurements into interpretable estimates of recurring physiological activation under real-world conditions.
Methods
Study design and analytic sample
This study builds on the NUBI-Drive naturalistic field protocol previously reported in4. In that protocol, young adult drivers were monitored continuously in everyday life for one week using their own Apple iPhones and Apple Watches, together with an on-board diagnostic (OBD2) device installed in their personal vehicles. The protocol was designed to capture routine behavior under ecologically valid conditions rather than under scripted laboratory tasks. The study was conducted by the Texas A&M Transportation Institute (TTI), approved by the Texas A&M University Institutional Review Board (IRB ID: IRB2021-0799D), and all participants provided written informed consent prior to participation. All procedures were conducted in accordance with the ethical principles of the Declaration of Helsinki and applicable U.S. federal regulations governing research involving human participants.
The present paper analyzes the complete NUBI II dataset, which extends the previously reported NUBI I dataset in two important respects. Whereas the earlier publication reported the initial analyzable cohort of 21 participants and focused exclusively on driving-related cardiac activation, the present study analyzes the complete cohort of 57 participants and considers the broader daily-life context by jointly analyzing driving and non-driving behavior.
Participants were observed across a full week so that both weekday and weekend behavior were represented within a common monitoring window. The protocol used continuous multimodal monitoring in naturalistic daily life, together with psychometric and workload-related instruments, to support analysis of cardiac activation in relation to both situational and participant-level factors. Participants typically began recording in the morning as they started their normal daily routines and continued until early evening. Because the protocol required simultaneous operation of multiple sensing devices, battery capacity was typically exhausted by early evening, after which the devices were recharged overnight before recording resumed the following day. Consequently, the analyzed dataset represents waking daytime physiology and does not include overnight recordings.
For the present analyses, activity was organized using the final three-level context annotation: driving, non-driving sedentary, and non-driving physical activity. Because the primary objective of this study was to characterize non-exercise-related baseline-referenced cardiac activation, the principal comparative analyses focused on driving and non-driving sedentary behavior. Physical-activity periods were identified as a separate behavioral context but excluded from the primary comparisons because heart-rate elevations during physical activity reflect metabolic demands that are fundamentally different from those associated with low-motion everyday behaviors. This design permits comparison of cardiac activation during driving against a low-motion daily-life reference condition within the same naturalistic cohort.
Sensing platform and variables collected
The multimodal sensing framework combined wearable, smartphone, vehicle, and survey-derived measures to characterize cardiac activation together with its behavioral, environmental, and person-level correlates across the participants’ waking week. The collected variables can be grouped into five categories.
Physiological variables
These included continuous heart-rate measurements acquired from the wearable devices. In addition, participant-day baseline heart-rate estimates were derived from these recordings. Heart rate served as the primary outcome, while the estimated baseline provided the physiological reference for quantifying baseline-referenced cardiac activation. As with all consumer wearable devices, physiological measurements may be affected by motion, device fit, skin contact, and optical sensing conditions26,27. These limitations are particularly important for beat-to-beat or waveform-derived biomarkers. In contrast, the heart-rate estimate analyzed here is a temporally averaged quantity, and Apple Watch heart-rate measurements have shown good agreement with reference measurements in independent validation studies28,29,30.
Behavioral and motion variables
These included hand-motion features derived from the wrist-worn smartwatch, together with location- and movement-derived features from the paired smartphone. They were collected to characterize bodily movement and task-related activity, helping distinguish low-motion sedentary states from physically active periods and, during driving, capture behaviorally relevant motion such as steering-related hand activity.
Vehicular and trip-related variables
These were available during driving through the OBD2 interface and related processing streams, including measures such as speed and other vehicle telemetry, as well as trip-level descriptors. They were collected to characterize operational aspects of the driving task that could modulate cardiac activation.
Environmental and contextual variables
These included GPS-derived location information, traffic-related indicators, weather conditions, and temporal context such as study day and period of day. They were collected because cardiac activation in daily life depends not only on the participant and the task, but also on surrounding situational conditions.
Participant-level variables
These included demographic, psychometric, and workload-related measures, such as age, sex coding, personality traits, trait anxiety, morning anxiety, and NASA-TLX subscales. Trait anxiety was assessed using STAI Form Y-231. It was included because anxiety predisposition is closely linked to sympathetic activation32 and may also shape driving-related behavior3334. NASA-TLX subscales were included because they provide a multidimensional measure of perceived workload35 and have been reported to covary with arousal during driving12. Together, these variables captured relatively stable individual differences and subjective workload factors that could help explain between-participant variation in cardiac activation.
Preprocessing, data cleaning, and context labeling
The original NUBI data streams were available at 1-s resolution. For the present study, these streams were aggregated to fixed temporal resolutions of 10, 30, or 60 s, since 1-s observations were considered too fine-grained relative to the noise level of the naturalistic measurements. Quantitative variables were aggregated within each window using the mean, whereas categorical variables were aggregated using the mode. The primary analyses reported here were conducted at the 60-s level, which we considered an appropriate temporal resolution for near-continuous daily-life monitoring.
Timestamps were interpreted as local clock time in the Chicago time zone. Study-day labels followed a Monday–Sunday weekly cycle, with day 1 corresponding to Monday and day 7 corresponding to Sunday. Context labels were assigned using a multimodal procedure that combined vehicle-state information with activity information from the smartphone–smartwatch ecosystem. Driving intervals were identified from the on-board diagnostic (OBD2) stream, which indicated when the participant was operating the vehicle. Outside those intervals, activity information from the Apple HealthKit framework was used to distinguish non-driving physical activity from non-driving sedentary behavior. This yielded the final three-level context annotation used in the present study: driving, non-driving sedentary, and non-driving physical activity.
The primary analyses were restricted to driving and non-driving sedentary segments. Physical-activity periods were retained for descriptive context but excluded from the primary analyses.
Data cleaning retained observations with valid timestamps, valid participant identifiers, valid study-day assignments, and valid context labels. Missingness was assessed at the natural granularity of each variable class, and quantitative summaries of missingness are reported in the Results section. Variable screening for modeling suitability is described below.
Heart-rate missingness analysis
To assess whether missing heart-rate measurements could bias the comparison between behavioral contexts, we conducted two mixed-effects logistic regression analyses. In both analyses, the binary outcome indicated whether the raw heart-rate measurement was missing for a given 60-s observation, and participant was included as a random intercept.
The first analysis included observations from the DRIVING and NONDRIVING_SEDENTARY strata. Fixed effects were behavioral context (DRIVING versus NONDRIVING_SEDENTARY), day period (AM versus PM), day type (weekday versus weekend), and weather condition.
The second analysis was restricted to DRIVING observations and examined whether heart-rate missingness was associated with objectively measured driving conditions. Fixed effects were vehicle speed, traffic congestion (Jam Factor), wrist linear-acceleration energy, and wrist rotational-acceleration energy. Complete regression results are presented in Supplementary Table S1.
Baseline definition and feature engineering
The dataset contained participant-day baseline heart-rate values, denoted HRbase,id, where i indexes participant and d indexes study day. The daily baseline values were obtained from the Apple HealthKit restingHeartRate metric36. According to Apple’s documentation, this quantity estimates the user’s daily resting heart rate from heart-rate measurements collected during sedentary periods, using motion information to distinguish resting from active periods. Although Apple does not publicly disclose the complete implementation of this proprietary algorithm, the reported quantity is intended to estimate resting heart rate rather than the daily minimum heart rate. Because the scientific objective of the present study is to quantify cardiac activation relative to each participant’s resting physiological state, resting heart rate provides the appropriate physiological reference. Adaptive stress regulation requires activation in response to behavioral demands followed by recovery toward baseline once those demands subside, whereas persistent or frequently repeated activation contributes to the development of allostatic load37. The present study does not measure allostatic load itself. Rather, it characterizes one physiological component relevant to that broader framework by quantifying the magnitude and persistence of baseline-referenced cardiac activation, together with the frequency of returns to baseline. Accordingly, the participant-specific baseline was defined as the average of the available daily resting-heart-rate estimates:
$$H{R}_{{{{\rm{base}}}},i}=\dfrac{1}{{D}_{i}}{\sum}_{d=1}^{{D}_{i}}H{R}_{{{{\rm{base}}}},id},$$
(4)
where Di is the number of study days with available baseline values for participant i.
Let HRraw,itd denote the momentary heart rate for participant i at time point t on day d. Using the participant-specific baseline defined in equation (4), baseline-referenced cardiac activation was defined as
$$NH{R}_{itd}=H{R}_{{{{\rm{raw}}}},itd}-H{R}_{{{{\rm{base}}}},i}.$$
(5)
Thus, NHRitd represents the deviation of momentary heart rate from the participant-specific baseline. Using a participant-specific resting baseline places all observations on a common physiological reference scale, allowing cardiac activation to be compared across participants and behavioral contexts. As a robustness analysis, all descriptive baseline-referenced analyses were repeated separately for weekday and weekend observations. Weekdays were defined as Monday through Friday, and weekends as Saturday and Sunday. This partition was motivated by post-study debriefings indicating that participants were predominantly office workers who generally worked during weekdays and had weekends off. The robustness analysis was intended to assess whether the observed relationship between driving and non-driving sedentary behavior remained qualitatively consistent across weekday and weekend observations rather than to compare work and leisure per se.
After defining the physiological reference quantities, feature engineering was performed separately for the two primary behavioral strata, driving and non-driving sedentary, with the aim of representing both stable participant characteristics and short-term contextual variation. For the driving stratum, additional features were engineered from the underlying vehicle and wrist-sensor streams. In particular, hand-acceleration energy and hand-rotation energy were derived from the Apple Watch accelerometer and gyroscope signals, respectively, to summarize the intensity of wrist movement associated with driving behavior. Driving-specific dynamic features were also constructed from the underlying sensor signals using lagged values, first differences, rolling means, rolling standard deviations, and rolling slopes over 1-, 3-, and 5-min windows. These features were intended to capture short-term fluctuations in driving behavior and environment rather than only static conditions.
Temporal and contextual variables were also engineered. Study day and period of day were represented explicitly, and weather categories were consolidated when necessary to ensure adequate support for analysis. In particular, rare unfavorable weather conditions were grouped into a common adverse weather category, allowing more stable comparisons against the more prevalent clear and cloudy conditions. Trip-level descriptors such as trip distance and trip duration were retained for the driving analyses, whereas the non-driving sedentary stratum relied more heavily on participant-level and temporal/contextual variables.
Although baseline-referenced heart rate (NHRitd, equation (5)) was used for descriptive analyses, the prediction target for the machine-learning models was the instantaneous raw heart rate, HRraw,itd. Accordingly, variables representing that quantity or closely related derived measures were excluded to reduce the risk of target leakage.
Prediction models
Guided by the baseline-referenced decomposition framework, instantaneous heart rate was modeled as
$$H{R}_{{{{\rm{raw}}}},itd}=H{R}_{{{{\rm{base}}}},i}+{\tau }_{c}+\phi ({X}_{itd})+{\epsilon }_{itd},$$
(6)
where HRbase,i is the participant-specific baseline, τc is the context-specific offset associated with behavioral context c (DRIVING or NONDRIVING_SEDENTARY), ϕ(Xitd) represents additional modulation by observed covariates, and ϵitd is residual variation. Using the definition of baseline-referenced heart rate in equation (5), this decomposition becomes
$$NH{R}_{itd}=H{R}_{{{{\rm{raw}}}},itd}-H{R}_{{{{\rm{base}}}},i}={\tau }_{c}+\phi ({X}_{itd})+{\epsilon }_{itd},$$
(7)
Thus, heart rate was decomposed into a participant-specific baseline, a context-specific offset, structured variation explained by observed covariates, and residual variation. Based on this framework, we compared three nested models within each behavioral stratum.
The baseline-only model predicted heart rate using the participant-specific baseline alone:
$${\widehat{HR}}_{{{{\rm{raw}}}},itd}^{(0)}=H{R}_{{{{\rm{base}}}},i},$$
(8)
The baseline-plus-offset model added a context-specific offset:
$${\widehat{HR}}_{{{{\rm{raw}}}},itd}^{({{{\rm{offset}}}})}=H{R}_{{{{\rm{base}}}},i}+{\tau }_{c},$$
(9)
where τc is the mean empirical offset statistically associated with observations in context c.
The third model used Elastic Net (ENet) to represent the modulator term ϕ(Xitd). Elastic Net is a penalized linear regression model that combines L1 and L2 regularization and is well suited to predictor sets that are moderately high-dimensional and potentially correlated. In the present study, ENet was used to capture structured variation beyond the participant-specific baseline and the context-specific offset while retaining interpretability at the feature level. All three models were fit separately within the DRIVING and NONDRIVING_SEDENTARY strata.
Cross-validation, tuning, and evaluation metrics
Model evaluation was based on a fully nested grouped cross-validation framework. The outer loop consisted of grouped 5-fold cross-validation, with participant as the grouping unit so that all observations from a given participant were assigned to the same outer fold. To support direct comparison between behavioral strata, the analysis was restricted to participants represented in both the driving and non-driving sedentary strata.
For the baseline-plus-offset model, the context-specific offset τc was defined as the mean baseline-referenced heart rate within each behavioral stratum,
$${\tau }_{c}={{{\rm{mean}}}}\left(H{R}_{{{{\rm{raw}}}},itd}-H{R}_{{{{\rm{base}}}},i}\right).$$
(10)
During cross-validation, this quantity was estimated using only the observations in the corresponding outer-training partition and was then applied to the held-out outer test fold. This ensured that the offset was estimated without using observations from the held-out participants.
Candidate predictors were defined separately for the driving and non-driving sedentary strata so that only variables conceptually applicable to each behavioral context were considered. For the ENet model, all data-dependent preprocessing was performed within the relevant training partition to avoid information leakage. The preprocessing workflow accommodated previously unseen nominal levels, one-hot encoded nominal predictors, imputed occasional missing numeric values using training-partition medians, removed zero-variance predictors, and z-score normalized numeric predictors.
ENet hyperparameters were tuned in two stages, using an initial coarse grid over the penalty and mixture parameters followed by a refined grid centered on the best first-stage setting. Within each outer-training partition, grouped 4-fold cross-validation with participant as the grouping unit was used for hyperparameter optimization. The penalty and mixture parameters were selected by minimizing inner-cross-validation RMSE. The selected hyperparameters were then used to fit the model on the complete outer-training partition and generate predictions for the corresponding held-out outer test fold.
Predictive performance was evaluated exclusively from predictions generated for the held-out outer test folds. Model performance was assessed using root mean squared error (RMSE), the squared Pearson correlation between observed and predicted heart-rate values (r2), and the conventional out-of-sample coefficient of determination (\({R}_{{{{\rm{OOS}}}}}^{2}\)). These metrics were computed separately for each outer test fold and summarized as the mean ± SE across the five outer folds. RMSE quantifies the typical magnitude of prediction error in the original units of beats per minute. The correlation-based metric (r2) summarizes how well the predictions reproduce the variation in observed heart rate, independent of the overall error magnitude. Because a high correlation alone does not necessarily imply accurate prediction, we additionally report the out-of-sample coefficient of determination,
$${R}_{{{{\rm{OOS}}}}}^{2}=1-\dfrac{{\sum}_{i}{(\,{y}_{i}-{\hat{y}}_{i})}^{2}}{{\sum }_{i}{(\,{y}_{i}-{\bar{y}}_{{{{\rm{test}}}}})}^{2}},$$
(11)
where yi and \({\hat{y}}_{i}\) denote the observed and predicted heart rate, respectively, for observation i in the corresponding held-out outer test fold, and \({\bar{y}}_{{{{\rm{test}}}}}\) denotes the mean observed heart rate over all observations in that test fold. This metric quantifies the improvement in prediction relative to simply predicting the mean observed heart rate of the held-out test fold.
Long-horizon projection and uncertainty analysis
To translate short-timescale model predictions into interpretable long-horizon quantities, we performed three related analyses. First, we used the driving ENet model to project baseline-referenced cardiac activation under repeated daily driving scenarios. Second, we propagated uncertainty around these projected annual activation estimates. Third, we examined how prediction error changed when model outputs were averaged over progressively longer time horizons.
Scenario projections
We used the driving ENet model, refitted on the full driving stratum using the selected hyperparameters, to generate projected baseline-referenced cardiac activation under repeated daily driving scenarios. Predicted instantaneous heart rate values were converted to baseline-referenced predicted activation according to the prediction analogue of equation (5),
$${\widehat{NHR}}_{itd}={\widehat{HR}}_{{{{\rm{raw}}}},itd}-H{R}_{{{{\rm{base}}}},i}.$$
(12)
A single illustrative scenario projection was performed using trait anxiety as an example of a participant-level characteristic. The projection was based on all observed driving epochs in the dataset. For each driving observation, the trait-anxiety value was replaced by either the lowest or highest observed participant-level trait-anxiety value while all other observed predictors, including the participant-specific baseline heart rate, were retained unchanged. Thus, each observed driving epoch was evaluated twice under two otherwise identical counterfactual conditions, with trait anxiety as the only model input changed between the paired predictions.
For each simulated profile, the model generated predicted baseline-referenced activation for every observed driving epoch using equation (12), and these predictions were averaged to obtain the corresponding mean activation during driving. The resulting mean activation was then projected to annual cumulative activation under three illustrative daily driving schedules: a 30 min/day commute, a 2 hr/day commute, and an 8 hr/day professional-driving schedule. Annual cumulative activation was computed by converting daily driving duration into annual exposure hours and multiplying by the mean predicted baseline-referenced activation. Results were expressed as cumulative NHR-hours per year.
Uncertainty estimation
For the long-horizon projections, uncertainty intervals around the individual low- and high-trait-anxiety scenario means were generated by nonparametric bootstrap resampling of participants while holding the fitted ENet fixed. For each simulated profile, participants were resampled with replacement, the mean predicted baseline-referenced activation during driving was recomputed from the corresponding bootstrap sample, and annual cumulative activation was calculated for each driving schedule. Each trait-anxiety profile under each exposure schedule was simulated 20,000 times. Uncertainty was summarized using the empirical 5th–95th percentile range of the simulated annual activation values. These intervals characterize participant-sampling variability in the projected scenario means conditional on the fitted model; they do not incorporate uncertainty in the estimated trait-anxiety contrast.
To assess the stability of the high-minus-low trait-anxiety contrast to individual participants, we performed a leave-one-participant-out sensitivity analysis. The driving ENet was refit 57 times, each time excluding one participant while retaining the same fixed model specification and hyperparameters as the scenario analysis. Each refitted model was then applied to the original observed driving epochs under the paired STAI = 20 and STAI = 61 scenarios, and the difference in mean predicted baseline-referenced activation was calculated. Two refits yielded boundary solutions in which regularization reduced the trait-anxiety coefficient to zero; these were reported separately and excluded from the empirical interval. For the remaining 55 refits, stability was summarized by the empirical central 95% interval of the leave-one-participant-out contrasts. Annual contrast intervals were obtained by applying the same deterministic exposure scaling used for the point projections.
Horizon-scaling analysis
Finally, to assess how prediction error changed under progressively longer temporal aggregation, we analyzed the out-of-fold ENet predictions from the outer folds of the nested grouped cross-validation procedure. For each prediction, the raw heart-rate residual was computed as the observed raw heart rate minus the ENet-predicted raw heart rate. Residuals were then ordered within participant and stratum by time and divided into contiguous temporal segments. A new segment was started when consecutive observations were separated by a gap larger than 1.5 times the expected sampling interval or occurred on different study days.
Within each contiguous segment, residuals were aggregated into complete non-overlapping blocks corresponding to 1, 2, 5, 10, 15, 30, and 60 minutes. For each block, the mean residual was computed. Horizon-specific error was then summarized as the root mean squared error of these block-mean residuals. To compare scaling across horizons, RMSE values were normalized to the 1-minute horizon within each stratum.
Together, these analyses move from short-timescale physiological predictions to interpretable projections of repeated baseline-referenced cardiac activation, while also quantifying uncertainty and evaluating how prediction error behaves under temporal aggregation.
来源:Nature 健康护理主题 · nature.com