Long COVID, or postacute sequelae of SARS-CoV-2 infection, affects many individuals following COVID-19 infection, resulting in persistent and often debilitating symptoms lasting months to years beyond the acute infection phase [1-4]. Among the most commonly reported symptoms are profound fatigue, postexertional malaise (PEM), and sleep disturbances, which collectively impair quality of life and functional capacity [5,6]. Understanding the physiological mechanisms underlying these symptoms and their temporal dynamics is critical for developing effective management strategies.

Sleep disturbances are prevalent in long COVID, with studies in post-COVID cohorts reporting sleep disruption in 40% to 75% of affected individuals [1,6]. Concurrently, disruptions in autonomic nervous system function have been documented, including reduced heart rate variability (HRV, a marker of parasympathetic activity and cardiovascular adaptability) [7,8]. Lower HRV has been associated with worse long COVID symptoms, suggesting impaired autonomic regulation may contribute to symptom persistence [9,10].

HRV, particularly during sleep when parasympathetic activity normally predominates, provides a window into autonomic recovery processes [11]. The high-frequency component of HRV (HRV-HF, 0.15‐0.4 Hz) reflects vagal/parasympathetic modulation, while the low-frequency component (HRV-LF, 0.04‐0.15 Hz) reflects mixed sympathetic and parasympathetic influences [12]. The LF/HF ratio is often interpreted as an index of sympathovagal balance, though its physiological specificity is debated [13]. In healthy individuals, sleep is characterized by stage-dependent autonomic modulation, with parasympathetic predominance during non–rapid eye movement (REM) sleep and relative sympathetic activation during REM [14].

Previous research has focused on unidirectional relationships, investigating how sleep quality or HRV predicts subsequent symptom severity [15]. A critical gap exists in understanding the reverse direction: whether daily symptom severity influences physiological recovery during subsequent sleep. This bidirectional perspective is particularly relevant in long COVID, where symptom-driven behavioral adaptations and psychological states may impact sleep quality and autonomic recovery [16].

PEM, a hallmark of long COVID and myalgic encephalomyelitis/chronic fatigue syndrome, involves symptom worsening following physical or cognitive exertion [17-19]. COVID-era physical inactivity has been documented as an additional concern for individuals recovering from SARS-CoV-2 infection [20]. The relationships between daily activity levels, sleep quality, and symptoms remain poorly understood. Evidence from accelerometer-based studies suggests that sleep duration, sedentary behavior, and physical activity should be considered as codependent daily behaviors rather than isolated exposures [21]. Activity may act as a moderator, with effects on symptoms depending on sleep quality.

To address these gaps, we conducted an exploratory longitudinal study using continuous wearable device monitoring in individuals reporting post-COVID fatigue. We employed a bidirectional analytical approach examining (1) prospective associations: how does the previous night’s sleep/HRV relate to next-day symptoms?; (2) reverse-direction associations: how do evening symptoms relate to that night’s sleep/HRV?; and (3) activity moderation: does physical activity interact with sleep quality in predicting symptoms? We hypothesized that (H1) sleep duration would show within-person day-to-day associations with symptoms; (H2) evening symptom severity would predict subsequent nocturnal autonomic function; and (H3) physical activity would moderate the sleep-symptom relationship.

This was a longitudinal observational study conducted between May 2025 and September 2025. Participants were recruited from individuals with post-COVID fatigue who had previously participated in a National Institute for Health and Care Research funded study at Newcastle University (Pausing-PCF [22]) and were engaged in patient and public involvement and engagement activities, as well as through online patient communities and social media. Inclusion was based on self-reported post-COVID fatigue symptoms and willingness to wear the device continuously. No formal clinician-adjudicated diagnosis was required, and no healthy control group was recruited; this represents an acknowledged limitation (see Limitations). This study uses a different cohort, device, and analytical framework from the authors’ prior feasibility study in this population [15], which involved 68 different participants, a different wearable device (Axivity AX6 accelerometer with VitalPatch ECG patch), and did not include reverse-direction associations or within-person effect disaggregation.

The study protocol was approved by the Newcastle University Ethics Committee (reference 47832‐2023). All procedures were conducted in accordance with relevant guidelines and regulations, and all participants provided written informed consent prior to enrollment. Participant data were deidentified prior to analysis. No individual identifying information appears in any figures or tables. Participants were volunteers who received no financial compensation. The ethics approval permitted secondary analysis of wearable data collected under this protocol without additional consent.

Participants wore a Fitbit Inspire 3 wristband continuously (24 h/d) for the study duration, charging for approximately 1 to 2 hours per day during inactive periods (eg, while showering). The Inspire 3 uses photoplethysmography to estimate interbeat intervals, from which HRV metrics are derived using proprietary algorithms. It is important to note that validation studies of consumer wearable photoplethysmography sensors have primarily examined heart rate accuracy and sleep staging [23-25]. Where spectral HRV has been formally evaluated in ring- and wrist-worn photoplethysmography devices against electrocardiogram (ECG) during nocturnal monitoring, agreement has been inconsistent: nightly-average HF power shows moderate-to-good correspondence, but 5-minute segment estimates of LF power (r≈0.42) and LF/HF ratio (r≈0.36) show poor agreement with ECG-derived values [26,27]. No independent peer-reviewed study has specifically validated the Fitbit Inspire 3’s proprietary spectral HRV algorithm against ECG; these metrics should therefore be treated as indicative rather than definitive measures of autonomic function. The device’s internal HRV sampling algorithm processes photoplethysmography segments during sleep, enabling frequency-domain HRV estimation [28]; the “1 Hz” sampling figure sometimes quoted for this device refers to display resolution rather than the raw optical sampling rate, which is substantially higher. Participants were instructed that step data would be collected automatically during waking hours.

Sleep metrics extracted included sleep duration (min), sleep efficiency (proportion of time asleep while in bed, %), deep sleep duration (min), and REM sleep duration (min). The device automatically detected sleep onset, awakening, and sleep stages using movement patterns and heart rate data [29].

HRV metrics included HRV-RMSSD (root mean square of successive differences), a time-domain measure reflecting parasympathetic activity [30]; HRV-HF (high-frequency power, 0.15‐0.4 Hz), a frequency-domain measure of vagal modulation [12]; HRV-LF (low-frequency power, 0.04‐0.15 Hz); and the LF/HF ratio, included as an exploratory autonomic index but interpreted cautiously given its debated physiological specificity [13]. HRV metrics were measured exclusively during sleep, providing an index of nocturnal autonomic recovery. Sleep periods with <50% HRV coverage were excluded, as signal quality is lowest during restless sleep. Physical activity was quantified as daily step count accumulated between sleep offset and symptom report time (prospective analysis) or from sleep offset to evening report (reverse-direction analysis).

Participants completed brief symptom reports via the FatigueSense smartphone app [31] at their convenience throughout each day. Reports were automatically timestamped. Participants rated fatigue severity on a 4-level categorical scale (none, mild, moderate, severe); and energy level on a 5-level ordinal scale (0, 25, 50, 75, 100). Reports were categorized by time of day: morning (5 AM-11:59 AM); afternoon (noon-4:59 PM); evening (5 PM-9:59 PM); and night (10 PM-4:59 AM). Participants were encouraged to report at least once daily but were not assigned fixed reporting times, introducing potential variability in the gap between waking and morning report completion and between evening report and subsequent sleep onset.

A key methodological feature of this study was the explicit use of symptom report timestamps to enable bidirectional analysis. Two analytically distinct datasets were created.

Data were extracted via the Fitbit Web API. HRV metrics (RMSSD, HF power, and LF power) were obtained from the intraday HRV end point (fields hrvRmssd, hrvHf, and hrvLf), which returns values at a native 5-minute resolution during device-detected sleep; the device does not expose beat-to-beat interbeat intervals through this endpoint. For each sleep period, HRV metrics were averaged across all available 5-minute segments, and periods with <50% segment coverage were excluded. Across the cohort, 36,521 five-minute HRV segments were recorded over 471 sleep periods (mean ≈78 segments, ≈6.5 h per period); thus, the “HRV segments” we report denote counts of 5-minute windows rather than individual minute-level readings. Sleep periods were identified using device-detected onset and offset times. Daily aggregates were computed for each 24-hour period (noon to 11:59 AM the following day). Symptom reports were matched to the appropriate sleep period based on timestamps.

All analyses were conducted in Python 3.13 and R (version 4.5.2; R Foundation for Statistical Computing) using the ordinal, lme4, lmerTest, repeated-measures correlation (rmcorr), and car packages.

Fourteen individuals reporting post-COVID fatigue symptoms (12 female [85.7%], 2 male [14.3%]; age range 35‐64 y) enrolled and completed the study protocol. Age distribution: 35 to 44 years (n=4, 28.6%), 45‐54 years (n=3, 21.4%), 55‐64 years (n=7, 50%). Participant characteristics and monitoring details are summarized in Table 1. Descriptive statistics for all variables are presented in Table 2.

We examined how the previous night’s sleep and HRV were associated with next-day fatigue and energy (n=260 prospective observations, 14 participants). Figure 1 illustrates the observation flow.

We examined whether evening symptom severity predicted that night’s sleep and HRV (n=105 reverse-direction observations, 71 with complete HRV, 12 participants). These results are presented in Table 5.

This exploratory longitudinal wearable study examined bidirectional associations between sleep, nocturnal autonomic function, and daily symptoms in individuals with long COVID. The main findings were as follows: (1) sleep duration and REM sleep showed significant within-person (day-to-day) associations with next-day fatigue, and sleep efficiency showed a counterintuitive negative association with next-day energy that is most plausibly a chance finding; (2) HRV-RMSSD was a significant between-person predictor; individuals with chronically higher HRV reported consistently lower fatigue and higher energy but did not show significant within-person day-to-day associations; (3) physical activity interacted with sleep duration, with the protective effect of sleep on fatigue attenuated on more active days; and (4) in an exploratory reverse-direction analysis, evening energy (but not fatigue) showed a preliminary association with that night’s nocturnal HRV that did not survive Bonferroni correction and requires replication.

The distinction between within-person and between-person effects is conceptually critical for this study. The Mundlak models and rmcorr analyses converge on a consistent picture: day-to-day fluctuations in sleep duration predict day-to-day fluctuations in fatigue within the same individual. This is the most directly actionable finding: on nights when a person with long COVID sleeps longer than usual, they tend to feel less fatigued the next day.

The HRV-RMSSD finding is different in nature. Individuals with chronically higher HRV tend to report chronically better symptoms, but a given night's HRV-RMSSD does not predict whether that same individual will have a better or worse day of symptoms than their own average. This between-person difference may reflect stable individual characteristics, including underlying autonomic regulation capacity, severity of long COVID–related dysautonomia [34,35], baseline cardiovascular fitness, or other constitutional factors. This finding does not support HRV monitoring as a real-time, day-to-day symptom predictor in this population but does position HRV as a potentially useful biomarker of individual physiological reserve and overall symptom burden.

The pooled Spearman correlations (Multimedia Appendix 1) appeared to show much stronger HRV-symptom associations (ρ=−0.317 to −0.359 for fatigue, ρ=0.416 to 0.498 for energy). These were inflated by between-person variance: participants with higher average HRV also had lower average symptom burden, and, by treating 260 clustered observations as independent, this inflated both correlation coefficients and P values. The rmcorr and Mundlak analyses are methodologically appropriate for addressing the within-person scientific question and should be considered the primary analysis.

In an exploratory analysis of the reverse-direction pathway, evening energy but not fatigue showed a within-person association with that night’s HRV-RMSSD (rmcorr r=0.327, P=.01; log-linear model β=0.0069, P=.02). This association was nominally significant but did not survive Bonferroni correction (α=.0050) and is therefore preliminary. If replicated, it would be consistent with the hypothesis that energy and fatigue reflect partially distinct processes: energy, assessed on a 0 to 100 scale, may capture perceived physiological vitality more sensitively than the 4-level fatigue scale, which was dominated by “moderate” responses (388/597, 65% of reports) and whose restricted variance may have limited its sensitivity (see Limitations). Energy levels on this scale cannot be equated with physiological capacity, and the association may partly reflect shared diurnal patterning, the 24-hour movement behavior framework [21], or other unmeasured factors. Replication in larger, adequately powered samples is required before any inference is drawn.

Neither activity level alone nor its direct association with symptoms was significant. However, the significant Sleep × Steps interaction (P=.02) suggests that the relationship between sleep and fatigue is context-dependent. On more active days, longer sleep conferred less symptomatic benefit than on less active days. This is consistent with PEM mechanisms [36], wherein greater exertion may place demands on physiological recovery that partially counteract the restorative function of sleep, and with behavioral evidence that perceived effort shapes activity decisions in individuals with fatigue [37]. These findings align with emerging evidence that sleep, sedentary time, and physical activity should be considered as co-dependent daily behaviors influencing symptom outcomes [21]. The findings do not support direct exercise prescription, particularly given PEM concerns in this population. Low-burden, symptom-contingent movement approaches, including mind-body modalities such as Tai Chi, Qigong, or yoga, may offer more suitable future intervention directions [38-40], though the evidence specifically in long COVID requires dedicated investigation.

Activity guidance in long COVID may need to be personalized according to recent sleep quality and daily energy levels, in line with adaptive pacing principles [41].

A key contribution of this study is the explicit use of symptom report timestamps to construct distinct prospective and reverse-direction analytical datasets, enabling temporally grounded bidirectional analysis. The use of rmcorr and Mundlak CLMMs represents an advance over standard pooled analyses and random-intercept-only models in repeated-measures ecological momentary assessment research.

Future studies in this space should preregister the distinction between within-person and between-person hypotheses, use rmcorr or mixed-model–based estimates as primary bivariate descriptors, and ideally include Mundlak specifications in all primary analyses.

Several limitations should be considered when interpreting these findings. First, the small sample size (N=14) limits statistical power and the generalizability of the results. The cohort was predominantly female (n=12, 85.7%) and largely aged 45 to 64 years, reducing applicability to men and younger populations. Furthermore, the age range overlaps substantially with the perimenopausal and menopausal period, during which hormonal changes can independently influence sleep, autonomic function, and fatigue, representing a potential uncontrolled confounder. Participants were also recruited through support groups and patient and public involvement and engagement activities rather than clinician-confirmed diagnosis, which may have introduced heterogeneity in the sample. In addition, the absence of a healthy control group limits conclusions about whether the observed relationships are specific to long COVID–related fatigue.

Second, several methodological limitations should be acknowledged. HRV was derived from a consumer-grade Fitbit Inspire 3 using proprietary photoplethysmography algorithms, and spectral HRV measures, particularly LF power and the LF/HF ratio, have shown limited agreement with ECG-based measurements in previous validation studies. Consequently, these measures should be interpreted as indicative rather than definitive markers of autonomic function. The study also did not control for important factors known to influence HRV, including caffeine, alcohol, medications, or undiagnosed sleep disorders such as obstructive sleep apnea. Symptom reports were completed at participants’ convenience rather than at fixed times, introducing potential recall bias and variability in the temporal alignment between symptoms, activity, and sleep. Additionally, excluding nights with poor HRV coverage (<50%) may have disproportionately removed periods of disturbed sleep, potentially introducing selection bias.

Finally, limitations related to outcome measurement and statistical modeling should be considered. Fatigue ratings were heavily clustered around moderate severity, reducing variability and potentially limiting the ability to detect day-to-day associations; future studies would benefit from validated multidimensional fatigue measures. Although the reverse-direction model showed acceptable improvement following log transformation, residuals remained imperfectly normal, indicating that model assumptions were only approximately satisfied.

Fatigue was modeled as a numeric outcome in the Sleep × Steps interaction model, whereas it is treated as ordinal in the primary cumulative link mixed models; because the observed fatigue scale has only three levels (mild, moderate, severe) with no assumed interval spacing, this continuous treatment is a simplification, and the interaction estimate should be interpreted with corresponding caution. In addition, because step counts were accumulated between sleep offset and a variable symptom report time, total steps are correlated with time awake at the point of reporting, and the Sleep × Steps interaction could therefore partly reflect a Sleep × Time Awake interaction.

These exploratory findings generate several hypotheses for future investigation. (1) Sleep duration and REM sleep are within-person day-to-day predictors of fatigue (the third association surviving correction, between sleep efficiency and next-day energy, is not carried forward here as it most plausibly reflects chance); sleep hygiene interventions targeting these metrics may offer symptom benefit. (2) HRV-RMSSD is a between-person correlate of average symptom burden, positioning it as a potential baseline biomarker rather than a real-time tracking tool. Whether autonomic-targeted interventions (HRV biofeedback, vagus nerve stimulation, slow breathing, mind-body practices) can improve between-person HRV levels and thereby reduce symptom burden in long COVID warrants experimental investigation [42,43]. (3) The preliminary, replication-dependent evening energy → nocturnal HRV association raises the hypothesis not yet supported by corrected evidence that evening relaxation strategies targeting subjective energy may support nocturnal autonomic recovery; testing this requires randomized experimental designs. (4) Activity pacing should incorporate recent sleep quality, with higher-exertion activities potentially better tolerated following nights of adequate sleep.

Randomized controlled trials manipulating sleep opportunity, HRV via biofeedback, or relaxation before sleep with crossover designs within the same individuals to leverage within-person variation would be the appropriate next step to test these hypotheses.

This exploratory longitudinal study identified bidirectional, temporally specific associations between sleep, nocturnal autonomic function, and daily symptoms in individuals with long COVID. Within-person analyses revealed that sleep duration and REM sleep predict day-to-day fatigue fluctuations, while HRV-RMSSD differentiated individuals with better versus worse average symptom burden (a between-person rather than day-to-day predictor). A third association, between sleep efficiency and next-day energy, also survived correction but ran counter to expectation and most plausibly reflects chance. Evening energy, but not fatigue, showed a within-person association with subsequent nocturnal autonomic function. These findings underscore the importance of disaggregating within-person from between-person effects in ecological momentary assessment research and highlight the potential complementary roles of sleep tracking and HRV monitoring: the former for real-time daily prediction, the latter for characterizing physiological reserve in long COVID management. Experimental work is needed to clarify causal mechanisms.