The COVID-19 pandemic continues to pose a significant global challenge, with the World Health Organization reporting >3.5 million new cases and nearly 69,000 deaths in 2024 alone [1,2]. Although multiple testing strategies are available for detecting acute infection, such as reverse transcription polymerase chain reaction (RT-PCR), antigen tests, and serological assays, each presents limitations in cost, sensitivity, availability, or applicability at different stages of the disease [3-5]. Mass availability of these tests is not straightforward as it involves complex production, logistics, and cost-related challenges. Furthermore, these tests are not useful for diagnosing post–COVID-19 condition.

Even after recovering from acute COVID-19, a significant proportion of patients continue to experience physical, psychological, and cognitive symptoms, which is known as post–COVID-19 condition [6,7]. This syndrome is characterized by persistent manifestations beyond microbiological recovery [8]. According to the World Health Organization, post–COVID-19 condition occurs in individuals with a confirmed history of SARS-CoV-2 infection, typically 3 months after onset, with symptoms lasting at least 2 months and not explained by alternative diagnoses [9].

Estimates suggest that at least 65 million people worldwide have post–COVID-19 condition, with the number increasing daily, whereas current diagnostic and treatment options remain insufficient [10]. Meta-analyses have shown that a substantial proportion of survivors of COVID-19 experience residual symptoms that persist for weeks, months, or even a year after recovery from the disease [11-13]. A systematic review [14] showed that post–COVID-19 condition can develop even after mild or asymptomatic SARS-CoV-2 infection.

However, no precise or reliable diagnostic test exists for post–COVID-19 condition, and the assessment of its prevalence remains controversial. Differences in methodologies and case definitions have led to wide variations in reported prevalence, ranging from <10% to approximately 60% of COVID-19 cases [15].

Given these challenges, wearable devices have emerged as a promising tool for continuous health monitoring. The adoption of personal devices that collect physiological data has been increasing annually [16]. These devices enable the noninvasive collection of real-time data that, when appropriately processed, can be used to assess physiological changes, including overall health status [17].

Several studies have investigated the use of physiological data to detect COVID-19, with some reporting successful prediction of infection even before symptom onset [18-24]. These studies analyzed various physiological parameters, including skin temperature, respiratory rate, and heart rate. Among the various biomarkers, heart rate variability (HRV) has gained attention for its potential as an indicator of autonomic nervous system (ANS) activity [25-27].

HRV is a noninvasive, objective, and validated measure of ANS dysfunction, providing insights into the balance between the sympathetic and parasympathetic nervous systems through the variation in time intervals between consecutive heartbeats [28]. While high HRV is associated with good health, a reduction in HRV is linked to various health issues [29,30].

The relationship between HRV and inflammatory markers has been extensively studied. A meta-analysis of 51 studies demonstrated an inverse correlation between HRV indexes and inflammatory markers, including interleukin-6 and C-reactive protein (CRP) [31]. Similarly, other studies have indicated HRV as a sensitive biomarker of inflammatory activity and immunomodulation, with variations in certain HRV parameters associated with elevated levels of proinflammatory cytokines, such as interleukin-6 and CRP [32-34]. Furthermore, there is evidence suggesting that lower HRV is correlated with higher CRP levels, reinforcing the connection between autonomic regulation and inflammatory control [35].

These findings highlight the importance of HRV as a physiological marker of the inflammatory response, with potential applications in monitoring inflammatory and infectious conditions. In addition, HRV reduction has been observed to precede clinical deterioration, suggesting its potential as an early physiological marker of disease progression [36].

Recent evidence indicates that COVID-19 significantly impacts HRV, reflecting autonomic and inflammatory dysregulation during the course of infection. A study of 143 individuals, including 93 patients with COVID-19 and 50 healthy controls, demonstrated that SARS-CoV-2 infection is associated with HRV alterations, indicating potential autonomic and inflammatory impairment [37].

These findings are consistent with a growing body of evidence linking SARS-CoV-2 infection to impaired autonomic function and reduced HRV [38-42]. Moreover, some studies report that HRV changes can precede symptom onset, reinforcing its value for early detection [43,44]. A recent systematic review and meta-analysis further supports the feasibility of using wearable-derived HRV data to detect and monitor COVID-19 [45].

Beyond its utility as a biomarker of SARS-CoV-2 infection, there is evidence suggesting that persistent HRV alterations following SARS-CoV-2 infection may serve as a valuable marker for what is commonly known as post–COVID-19 condition. Several studies analyzing HRV in patients clinically diagnosed with post–COVID-19 condition have consistently shown reduced HRV compared to healthy controls [46-52].

These findings suggest enduring autonomic dysfunction and inflammatory dysregulation in individuals with this condition, indicating that HRV may serve as a marker for identifying and monitoring post–COVID-19 condition. However, most studies identifying these conditions have relied on statistical methods to associate HRV variations with the disease. Furthermore, most existing research has focused on retrospective analyses where physiological data were collected; processed; and subsequently compared to self-reported symptoms, laboratory-confirmed COVID-19 cases, or clinical diagnoses of post–COVID-19 condition.

Building on these efforts, this study focused on identifying distinct HRV patterns among healthy individuals, patients with COVID-19, and those with post–COVID-19 condition. These patterns may reflect autonomic and inflammatory regulation, providing the basis for a proposed machine learning–based predictive model for COVID-19 and post–COVID-19 condition.

The primary contribution of this study lies in delineating these HRV patterns, which capture meaningful physiological changes and enable predictive capabilities. By leveraging wearable devices, this approach facilitates noninvasive health monitoring, representing a significant advancement in the early detection and long-term management of COVID-19 and post–COVID-19 condition with near–real-time prediction capabilities. Unlike traditional diagnostic approaches, this method offers a pioneering advantage by providing, for the first time, predictions of post–COVID-19 condition based solely on physiological data rather than relying exclusively on clinical assessments.

This study aimed to identify distinct HRV patterns capable of differentiating among individuals with active COVID-19 and post–COVID-19 condition and healthy controls using data obtained from wearable devices. By applying supervised machine learning to HRV indexes, we sought to classify physiological states based on autonomic and inflammatory signatures. A secondary objective was to evaluate the technical feasibility of implementing a near–real-time health monitoring system that integrates HRV signal acquisition, preprocessing, and classification into an automated workflow for continuous, noninvasive assessment of health status.

This section outlines the data sources, preprocessing methods, HRV indexes, and machine learning techniques used to delineate distinctive HRV patterns in healthy individuals, patients with COVID-19, and those with post–COVID-19 condition. By integrating clinical and public datasets, we sought to establish physiological signatures of SARS-CoV-2 infection and its persistent sequelae applicable to both existing records and prospective data collection. We further demonstrate a custom-built system for near–real-time HRV monitoring, highlighting its potential as an innovative application of these findings.

This study analyzed HRV data from 3 distinct groups: healthy individuals (20/61, 33%), patients with active COVID-19 (21/61, 34%), and those with post–COVID-19 condition (20/61, 33%). In total, 2 datasets were integrated: a clinical dataset from Professora Lydia Storópoli Hospital, São Paulo, Brazil, and a public dataset from the Fit-COVID Study [65], a cross-sectional observational investigation. The demographic and clinical profiles of these groups are summarized in Table 1, with post–COVID-19 condition symptoms reported by the participants and comorbidities detailed in Table 2. The Fit-COVID dataset provided HRV data from 20 healthy individuals who tested negative for COVID-19 via immunoglobulin M or immunoglobulin G antibody tests, as well as 20 individuals with post–COVID-19 condition who had a confirmed polymerase chain reaction positive COVID-19 diagnosis, exhibited mild to moderate symptoms, and were assessed 15 to 180 days after infection. All participants were aged 20 to 40 years [65].

Exclusion criteria included chronic diseases, smoking, recent vaccination, or medications affecting autonomic function. A general anamnesis, BMI, and physical activity levels were evaluated. All HRV recordings in the Fit-COVID Study were conducted in the morning to minimize the effects of circadian variation. Participants were instructed to abstain from physical activity, caffeine, and alcohol for 24 hours before data collection to reduce external influences on autonomic function. Measurements were taken in a quiet, controlled environment, with participants seated and breathing spontaneously. After a 20-minute rest period, RR intervals were recorded for 25 minutes using a Polar RS800CX device with a 1-kHz sampling rate. HRV indexes were subsequently extracted using the Kubios HRV software (Kubios Oy). The clinical dataset comprised 45 hospitalized patients with active COVID-19 (RT-PCR confirmed, moderate to severe symptoms, and onset of 1-10 days) collected in October 2022. Common comorbidities among these patients included hypertension (21/45, 47%), diabetes (12/45, 27%), obesity (19/45, 42%), chronic obstructive pulmonary disease (5/45, 11%), and smoking (1/45, 2%), with 66% (30/45) vaccinated (predominantly with CoronaVac; 22/45, 49% first dose and 16/45, 36% second dose). During their hospital stay, treatments frequently included antibiotics (37/45, 83%), corticosteroids (42/45, 94%), anticoagulants (40/45, 88%), and antihypertensives (41/45, 90%). To ensure consistency and reduce measurement bias, all HRV recordings for the COVID-19 group were performed under standardized conditions during hospitalization. Each patient remained in a quiet, temperature-controlled room and was instructed to rest in the supine position for 20 minutes before data collection. HRV data were then recorded for 15 minutes using a Polar V800 device at a sampling rate of 1 kHz.

Patients were asked to remain still and silent throughout the session to minimize artifacts. Data were collected during daytime hours, typically in the morning, to reduce the influence of circadian variations on autonomic function. This protocol followed the established recommendations for short-term HRV measurement [66]. Single HRV recording was performed per participant under these controlled conditions as repeat measurements were not feasible in the clinical setting. Even under these controlled conditions, the collected RR interval signals exhibited discrepancies, potentially caused by voluntary or involuntary movements, acquisition failures, or conditions such as excessive sweating.

To ensure data quality, preprocessing steps were applied using the method by Tukey [67] for outlier detection and filtering, with thresholds determined by the IQR, defined as the difference between the third quartile and the first quartile. The following thresholds were used for identifying outliers: first quartile − 2.0 × IQR for the lower limit and third quartile + 2.0 × IQR for the upper limit.

The sensitivity factor K=2.0 was chosen to enhance precision in detecting artifacts and outliers. Despite these efforts, 53% (24/45) of the participants still exhibited significant artifacts or unstable signals and were excluded from the analysis. In this hospitalized population, repeating the HRV measurement was not always feasible due to logistical and clinical constraints. The remaining signals were reassessed using the Kubios software to confirm stability, resulting in 47% (21/45) of the participants providing at least 10 minutes of stable signals, enabling the extraction of HRV indexes.

Unlike several previous studies that relied on physiological data collected under minimally controlled conditions, this study used data collected in accordance with established recommendations for short-term HRV analysis [66], applying standardized protocols across 2 distinct datasets—hospitalized patients in the clinical cohort and outpatient volunteers in the observational one. Both adhered to core methodological practices, including standardized posture, prerecording rest, spontaneous breathing, daytime data acquisition, and quiet environments. This uniformity in data collection helped minimize procedural bias and enhanced the reliability and comparability of the extracted HRV indexes across groups.

The data in Tables 1 and 2, together with the comorbidity information of individuals who were infected described throughout this section, characterize the study cohorts, with the selection of and rationale for the HRV indexes used to identify physiological markers of SARS-CoV-2 infection detailed in the following section.

To identify physiological markers of SARS-CoV-2 infection and its sequelae, we selected HRV indexes from the Fit-COVID and clinical datasets guided by their physiological relevance and established associations with COVID-19 pathology.

We focused on linear methods using time-domain indexes, specifically SDNN and RMSSD as they measure overall HRV and short-term parasympathetic activity, respectively, with SDNN reflecting global variability that decreases with age [66,68]. RMSSD is a reliable index of HRV modulation by vagal activity, as demonstrated in previous studies [66,69,70]. In the frequency domain, we used relative power in LF and HF bands, representing sympathovagal balance, which is less influenced by baseline heart rate and aging compared to absolute power, increasing comparability and generalizability across populations.

Previous studies involving COVID-19 and post–COVID-19 condition have explored absolute power variations in LF and HF measured in milliseconds squared, identifying reductions in these indexes. However, this approach may introduce bias as absolute power is influenced by the individual’s basal heart rate [66] and significantly decreases with aging [71].

Frequency-domain analyses based on absolute power exhibit higher sensitivity than those using relative power [72,73], but relative power mitigates these biases, enhancing cross-study comparability and generalizability, particularly for short-term recordings in populations with COVID-19. Nonlinear methods such as Poincaré plot SD perpendicular to the line of identity 'SD1 and SD2 along the line of identity were excluded due to the lack of established associations with COVID-19 in the literature.

The selection was guided by scientific evidence linking SDNN to overall HRV, RMSSD to parasympathetic modulation, and LF and HF relative power to autonomic dysregulation, all of which are implicated in the inflammatory and cardiovascular responses of SARS-CoV-2 infection and post–COVID-19 condition. We selected LF and HF relative power over absolute power or normalized units to mitigate biases associated with basal heart rate, aging, and interindividual variability, as well as to reduce susceptibility to artifacts in short-term HRV assessments. This choice addresses a gap in the literature, where studies on COVID-19 and post–COVID-19 condition predominantly report absolute or normalized power but rarely relative power despite its advantages for detecting autonomic dysregulation in inflammatory conditions and other ANS anomalies.

The use of LF and HF relative power enhances comparability across diverse populations—including healthy controls, patients with COVID-19, and survivors of post–COVID-19 condition—by being independent of basal heart rate, thereby improving generalizability [66,72-75]. In addition, LF and HF relative power facilitates interpretation and comparison between studies and individuals.

The selected variables (SDNN, RMSSD, and LF and HF relative power) were organized into a CSV file for input into machine learning tools, as detailed in the next section. Statistical analyses, including the Kolmogorov-Smirnov test to assess normality, were conducted. For nonnormally distributed data, the nonparametric Kruskal-Wallis and Mann-Whitney U tests (P<.05) were applied to validate the distribution of these indexes, with results visualized via box plots to assess variability and outliers. These analyses, presented in the Results section, support the validity of our variable selection for distinguishing physiological states across these groups.

To ensure that the observed differences between the groups were not confounded by demographic factors such as age and BMI, we applied statistical analysis before implementing machine learning algorithms.

Analysis of covariance (ANCOVA) was used to control for potentially confounding effects related to the higher age and BMI observed in the infected group compared to the other groups, as well as their influence on HRV variables. ANCOVA is a statistical technique that combines ANOVA and linear regression to adjust for the effects of continuous covariates while examining the influence of the categorical independent variable on the dependent variable [76].

In this study, the COVID-19 group showed significant differences in age and BMI compared to the other groups. Studies indicate that advanced age is associated with a reduction in HRV both in short-term monitoring and in 24-hour recordings [68,77]. In particular, age negatively influences HRV after high-intensity exercise, especially in older individuals [78]. Furthermore, the relationship between age and HRV has also been observed in patients with chronic headache [79].

Similarly, BMI is often associated with reduced HRV, especially in individuals with obesity or metabolic syndrome. However, studies indicate that BMI alone may not be a robust predictor of HRV when compared to more specific measures of adiposity. One study [80] found that, while BMI alone did not show a significant correlation with HRV indexes, central adiposity, measured through waist circumference, was more strongly correlated with reduced HRV. Another study demonstrated that central adiposity has a more direct impact on HRV than overall BMI [81].

The ANCOVA model included HRV indexes (SDNN, RMSSD, and LF and HF relative power) as dependent variables, the infection group as a categorical factor, and age and BMI as continuous covariates. The assumption of homogeneity of slopes was assessed to ensure that the relationship between the covariates (age and BMI) and the dependent variable did not vary between groups.

Before the main modeling, we conducted several tests to verify the assumptions of the ANCOVA. The Pearson correlation was used to assess the relationship between age and HRV, as well as between BMI and HRV. An interaction test (group × covariate) was conducted to confirm the homogeneity of slopes, whereas the Shapiro-Wilk test ensured the normality of residuals. To verify the homogeneity of variances between groups, the Levene test was applied. This methodological approach allowed us to assess both the direct impact of COVID-19 on HRV and the potential interaction between the covariates (age and BMI) and the analyzed groups.

To optimize the classification algorithm for the dataset, the machine learning and deep learning application in MATLAB (version R2023b, Student Edition; The MathWorks, Inc) was used. This platform tested multiple classifiers—decision trees, SVMs, k-nearest neighbor (k-NN), artificial neural networks, and ensemble methods—with automated hyperparameter tuning and cross-validation. The HRV indexes SDNN, RMSSD, and LF and HF relative power, sourced from clinical and public datasets and extracted using the Kubios software, were used to evaluate algorithm performance.

In total, 2 training rounds were conducted: the first used only the 4 HRV indexes, and the second included an additional contextual predictor, covidrecent. This temporal variable was coded as 1 only for participants in the post–COVID-19 condition group and 0 for all others (those with acute infection and healthy) based on patient-reported SARS-CoV-2 infection within the previous 3 months. This criterion mimics real-world triage scenarios where recent infection history is typically available during intake interviews and may aid diagnostic refinement. It was not derived from the output label but rather represents a realistic feature used to enhance model performance in triage scenarios.

The full set of input values used in this phase, including HRV indexes and the covidrecent variable, is available in Multimedia Appendix 1. A 5-fold stratified cross-validation approach ensured robust validation across the 61-participant dataset (n=21, 34% infected; n=20, 33% healthy; and n=20, 33% with post–COVID-19 condition).

The classification accuracy served as the primary evaluation metric during this phase; however, precision, recall, and F₁-score were also considered to gain a comprehensive understanding of algorithm performance. Hyperparameter tuning was performed automatically using MATLAB’s built-in optimization functions to tailor the algorithms to the datasets’ characteristics and enhance predictive performance. Specific hyperparameters adjusted for each model family are detailed in Multimedia Appendix 1, along with accuracy values before and after optimization. The results were ranked based on classification accuracy obtained during validation tests, providing an initial indication of the most suitable models for further refinement.

Following the identification of the most promising models in MATLAB, these algorithms were reimplemented in Python (Python Software Foundation) using the scikit-learn library. During the transition to Python, validation tests were conducted to ensure consistency between the models’ performance in MATLAB and Python. This transition enabled greater flexibility for additional experimentation, facilitated integration with other commonly used machine learning frameworks, and supported the development of a near–real-time prediction solution.

In MATLAB, 5-fold stratified cross-validation was initially applied. For the Python implementation, a 15-fold k-fold stratified strategy was adopted to enhance the robustness of performance estimates. While increasing the number of folds reduces the training set per iteration, it provides more stable aggregated metrics. A fixed random seed (random state=42) was used to ensure full reproducibility across all model training and validation steps.

To assess the feasibility of near–real-time physiological monitoring based on HRV, we developed a custom-designed fingertip pulse oximeter capable of collecting photoplethysmography (PPG) signals and transmitting them to a server using the MQTT protocol. The raw PPG data were stored in a MySQL database and processed using a Python-based pipeline to extract HRV indexes, SDNN, RMSSD, and LF and HF relative power in both the time and frequency domains.

HRV computation was conducted using the HeartPy library [82], which applies preprocessing routines, including outlier rejection, peak detection, and filtering. Before HRV extraction, PPG signals were normalized using minimum to maximum scaling and processed using a second-order Butterworth band-pass filter (0.5-3.5 Hz) to minimize baseline wander and HF noise. A minimum recording duration of 2 minutes was adopted for each session in accordance with established recommendations for LF HRV assessment.

The system was designed to automatically transmit HRV features at the end of each recording cycle, enabling continuous physiological monitoring without user intervention. A web-based interface displays the classified physiological state—COVID-19, post–COVID-19 condition, or healthy—based on previously trained machine learning models.

A feasibility demonstration was conducted with 4 independent participants: 2 (50%) healthy individuals and 2 (50%) individuals with RT-PCR–confirmed active COVID-19. Each participant underwent a 2-minute recording session, after which the system completed signal preprocessing, HRV computation, and classification in approximately 1 second. Figure 1 illustrates the overall workflow of data acquisition, processing, and classification.

This application demonstrates that a fully functional, autonomous system can perform near–real-time HRV-based classification of physiological states. While the system itself is operational, the robustness and generalizability of the physiological signatures used for classification would benefit from validation in larger and more diverse populations. The implications and limitations of these findings are further discussed in the Limitations section.

Additional technical details regarding the hardware architecture, signal processing, and transmission protocols are available in Multimedia Appendix 2.

This study used data from 3 ethically approved sources, all in accordance with Brazilian Resolution 466/12 of the National Health Council and the Declaration of Helsinki.

The first dataset was obtained from a previously approved clinical study involving patients with COVID-19 conducted under protocol 46699521.5.0000.5511 and approved by the research ethics committee of Nove de Julho University.

The second dataset was sourced from the publicly available Fit-COVID Study, a cross-sectional observational investigation conducted under independent ethics approval. This dataset includes anonymized HRV recordings from healthy individuals and patients with post–COVID-19 condition, and its use in this study complied with ethical reuse and secondary data use guidelines.

The third dataset was collected by the authors under protocol 68909623.0.0000.5511, also approved by the research ethics committee of Nove de Julho University. It involved near–real-time HRV index acquisition across different clinical stages.

All participants were aged >18 years and provided written informed consent.

All data were anonymized before analysis, and HRV recordings were acquired following standardized conditions for short-term HRV analysis. This study complied with all applicable regulations concerning human participant research and data protection.

This study demonstrates that HRV can be effectively used to distinguish among healthy individuals, those with active COVID-19, and those with post–COVID-19 condition, in alignment with previous findings on HRV reduction during and after infection. The observed differences in SDNN, RMSSD, and LF and HF relative power among these groups provide a robust basis for developing predictive models that advance beyond existing approaches [37-41,43-53,56-62].

The selection of HRV indexes varies across studies and influences their interpretation. Previous research has often relied on absolute LF and HF power [37,47,62,86], an approach known to depend on the individual’s basal heart rate, as highlighted by the Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology [66] and Kamath et al [72], introducing variability when comparing populations. Normalized units such as LFnu and HFnu [74] face similar limitations. Other studies have used only time-domain indexes such as SDNN or RMSSD [46,50]. In contrast, our study examined SDNN, RMSSD, and LF and HF expressed as relative power (LF% and HF%), consistent with recommendations from [66,72,75] that suggest these metrics enhance comparability across diverse groups.

A time frame for physiological data acquisition is a key consideration in HRV analysis. Data can be collected in short- or long-term settings, each with distinct methodological implications. Short-term HRV studies are conducted under controlled conditions, where confounding factors such as body position, physical activity, respiration, and environmental temperature are monitored and minimized [70]. In contrast, long-term HRV data are typically acquired using wearable sensors (eg, Holter monitors), allowing participants to move freely during daily activities. These methodological differences significantly influence HRV indexes in both the time and frequency domains [66], and their physiological significance may differ as short-term HRV measures correlate only weakly with long-term values [72]. Such variation in measurement protocols can lead to discrepancies in findings across studies. Of the studies cited in this work that assessed HRV in COVID-19 contexts, approximately 59% used short-term protocols.

Aging is a well-established factor associated with the reduction in HRV. Several studies have documented that advancing age is linked to a progressive decline in specific HRV indexes, particularly RMSSD, SDNN, and the HF and LF components [68,78,79]. This reduction primarily reflects a decrease in parasympathetic activity, although global variability is also affected. However, the magnitude of HRV decline associated with aging occurs gradually and progressively, with modest annual changes [77] particularly after the fifth decade of life.

Beyond aging, pathological conditions such as SARS-CoV-2 infection have also been associated with significant reductions in HRV.

Time- and frequency-domain HRV indexes can be influenced by multiple factors, including age-related autonomic changes [87,88], which may complicate comparisons across groups with differing age distributions. Nevertheless, several studies have reported that both acute COVID-19 and post–COVID-19 condition can independently impact HRV [37,49,62,89], reinforcing its potential relevance for identifying autonomic dysfunction in these populations.

In this study, although the COVID-19 group exhibited a higher mean age and BMI compared to controls, the pronounced reductions in HRV indexes substantially exceeded the variations typically attributed to differences in age or BMI alone. Moreover, the BMI distribution within the COVID-19 group was relatively homogeneous without extreme outliers, minimizing the potential for confounding effects related to body composition. These findings suggest that acute SARS-CoV-2 infection likely represented the primary driver of the autonomic dysfunction observed in this cohort rather than demographic or anthropometric factors.

The ANCOVA with PSM confirmed that the reductions in RMSSD and LF relative power in the infected group were primarily driven by SARS-CoV-2 infection independent of age and BMI. By balancing age and BMI across groups, the analysis mitigated the confounding effects of these demographic factors. These findings align with those of previous studies reporting HRV reductions in acute COVID-19, suggesting autonomic dysfunction potentially due to inflammation or direct viral effects on the ANS [37,89].

Although SDNN is widely recognized as a reliable marker of autonomic dysfunction, especially in acute COVID-19, its importance in classifying features in our decision tree model was minimal. This likely reflects its high collinearity with RMSSD in short-duration HRV recordings [70,85,90]. This outcome does not diminish the physiological value of SDNN but rather illustrates how decision tree algorithms may underrepresent physiologically relevant variables when they are statistically correlated with others, particularly in short-term contexts.

Algorithm selection in our study balanced practical and methodological factors. Decision trees were chosen for their simplicity and interpretability. This choice aligns with approaches in prior work, such as the study by Vellido [84], which emphasizes interpretability in medical contexts, and contrasts with studies such as the one by Aliani et al [37] that may use more complex techniques to detect subtle autonomic changes. Meanwhile, the study by Teng et al [83] highlights how decision trees provide rule-based clarity, offering a lens to examine HRV patterns across COVID-19 states, although this approach reveals a trade-off between model simplicity and the broader applicability of findings when compared to other studies.

In summary, factors such as data collection duration, participant demographics, and the selection of HRV indexes and algorithms shape the interpretation of HRV findings and their comparability with those of prior work. Our study builds on existing research [37-42,55-62] by integrating short-term HRV measures and relative power indexes (LF and HF) to delineate COVID-19–related patterns in a cohort of 61 participants, yet it reveals methodological differences that complicate direct comparisons.

For instance, while the study by Aliani et al [37] used random forest and SVM to detect subtle autonomic shifts in COVID-19 severity using HRV data derived from PPG, and other studies [46,50,51] report autonomic dysfunction in post–COVID-19 condition through extended 24-hour Holter monitoring, our short-term approach captures acute differences among healthy, infected, and post–COVID-19 condition groups. These contrasts suggest that standardizing these factors could refine future investigations into HRV as a biomarker for COVID-19 and beyond.

Building on these foundations, this study advances the field through 2 key contributions. First, it delineates a potential physiological signature derived from specific HRV indexes capable of distinguishing healthy individuals; patients with active COVID-19; and those with post–COVID-19 condition, a condition currently diagnosed solely through clinical criteria due to the lack of objective biomarkers. Second, it introduces a fully automated, noninvasive classification system leveraging near–real-time HRV data to identify both acute COVID-19 and post–COVID-19 condition.

The integration of the temporal variable covidrecent into the classification framework provided an important insight for clinical triage and decision support. While traditional statistical analyses (eg, Mann-Whitney U tests) did not show significant differences in HRV indexes between individuals with post–COVID-19 condition and healthy individuals, the decision tree model identified multivariate patterns that separated these groups. The model achieved 90% recall and 92% F₁-score for the post–COVID-19 condition class, demonstrating that combining contextual clinical information with physiological data can improve diagnostic accuracy.

Although HF relative power did not retain statistical significance after adjusting for age and BMI in the ANCOVA, it was selected by the decision tree model as an early-split feature. This apparent discrepancy reflects the fundamental difference between statistical inference and predictive modeling. Machine learning algorithms prioritize features based on their contribution to classification accuracy and reduction in node impurity rather than isolated group comparisons. In this context, HF relative power contributed to a decision threshold that helped differentiate post–COVID-19 condition from other classes even if its adjusted group means were not statistically distinct. This underscores the strength of multivariate pattern recognition in revealing complex physiological dynamics, particularly in heterogeneous syndromes such as post–COVID-19 condition.

These insights are especially relevant when considering the practical deployment of HRV-based classification systems. In demonstrating these contributions, this study also established the feasibility of an autonomous near–real-time HRV-based health monitoring system that is fully operational. The system’s performance depends primarily on the robustness and generalizability of the physiological signatures identified rather than on its technical architecture. Further validation with larger and more diverse populations would refine the HRV patterns used, enhancing the clinical applicability of the system.

This study is subject to several limitations that should be considered when interpreting its findings. The sample size of 61 participants (n=21, 34% with active COVID-19; n=20, 33% healthy individuals; and n=20, 33% with post–COVID-19 condition) may limit the generalizability of the identified HRV patterns to larger and more diverse populations. Although age, sex, and comorbidities were recorded, no subgroup analysis was conducted to evaluate their specific influence on HRV, limiting insights into potential effect modifiers.

The HRV indexes for healthy participants and those with post–COVID-19 condition were derived from preprocessed data, preventing verification of whether the original RR interval measurements were affected by artifacts such as body movements, improper finger oximeter placement, or environmental interference. Furthermore, HRV data for participants with active COVID-19 were collected at a single research center, which may introduce geographic or institutional bias and reduce the representativeness of the findings.

The reliance on short-term HRV recordings, which were collected under controlled conditions using a custom finger oximeter, differs from long-term monitoring approaches such as 24-hour Holter recordings in both methodology and physiological interpretation. While this short-term approach effectively captured distinct HRV patterns for immediate classification, the lack of longitudinal data restricts the ability to assess changes in autonomic function over time in COVID-19 or post–COVID-19 condition progression.

The near–real-time health monitoring system demonstrated technical feasibility by accurately classifying 4 independent participants (n=2, 50% healthy and n=2, 50% with active COVID-19) based on HRV signatures derived from the main study cohort (N=61). While this supports the system’s operability, the underlying physiological patterns have not yet been validated across larger and more diverse populations. Therefore, the limitation lies not in the system itself but in the need for broader validation of the HRV-based signatures used for classification.

The classification model’s strong reliance on the temporal variable covidrecent, as shown by the drop in accuracy from 87% to 76.4% when it was removed, suggests that HRV patterns may require complementary clinical context to support robust classification, particularly for distinguishing post–COVID-19 condition. This highlights the need for further refinement and validation of physiological signatures across different clinical scenarios.

Finally, the validation strategy used in this study relied on stratified 15-fold cross-validation within a combined dataset of 61 participants, a decision driven by the need to preserve class balance and optimize statistical power across the 3 diagnostic groups. While this approach improves reliability over simpler methods, it cannot entirely eliminate the risk of performance inflation due to possible information leakage across folds, particularly in modest samples. Future studies with larger and more diverse populations are needed to support external validation and further assess the generalizability of the proposed models.

This study advances the understanding of HRV as a biomarker by identifying multivariate physiological signatures—combinations of SDNN, RMSSD, and LF and HF relative power—that differentiate healthy individuals, those with active COVID-19, and those with post–COVID-19 condition. Rather than focusing on isolated HRV reductions, this approach captured integrated patterns across time- and frequency-domain metrics.

Machine learning models trained solely on HRV indexes achieved 76.4% accuracy, with high discriminative power for active COVID-19 (F₁-score of 88% and AUC of 0.8536). However, the detection of post–COVID-19 condition cases was limited, with an F₁-score of 56%, reflecting the greater subtlety and variability in the autonomic changes associated with this condition. When the contextual variable covidrecent, indicating a confirmed SARS-CoV-2 infection in the previous 3 months, was incorporated, the model’s performance improved substantially. Overall accuracy increased to 87%, and the classification of post–COVID-19 condition cases improved markedly, with the F₁-score increasing from 56% to 92% and the AUC increasing from 0.7543 to 0.8421 while preserving the model’s high ability to distinguish individuals with active COVID-19 from healthy individuals.

Finally, this study demonstrated the feasibility of applying a trained model to new short-term HRV data under near–real-time conditions. The ability to acquire, process, and classify signals on demand underscores the value of adaptive health monitoring in both clinical and remote environments. These physiological signatures derived from time- and frequency-domain HRV features help bridge the current gap in objective diagnostic frameworks for post–COVID-19 condition and contribute to early detection in acute COVID-19. With broader validation, they may enable scalable, data-driven solutions for individualized care, aligned with emerging trends in digital and predictive health care.