Mental health disorders constitute a major global public health concern, with an escalating number of patients requiring mental health services with considerable social cost (1). Depression and anxiety are the two most disabling mental disorders, ranked among the top 25 leading causes of healthcare burden worldwide in 2019 (2). In South Korea, the problem is severe, as the suicide rate has consistently ranked first in the OECD for more than 10 years, with around 25 people per 100,000 resorting to suicide each year (3). The disease burden of mental and behavioral disorders was estimated to account for 6.4% of the total disease burden in South Korea (4). The importance of mental health services therefore cannot be overstated.

Traditionally, patients receive diagnoses through face-to-face consultations with psychiatrists, which pave the way for various treatments such as counseling, medication, and hospitalization. However, this approach can render mental health services less accessible in certain areas or under specific circumstances (5). For instance, during the recent COVID-19 pandemic, the number of patients with depression surged; however, providing appropriate services proved challenging due to social distancing measures and other factors (6). Even outside of pandemic conditions, it is consistently reported that current mental health services are unable to cope with the rapid increase in the number of psychiatric patients (7). Furthermore, face-to-face consultations have inherent limitations; they rely on individuals’ ability to recollect their symptoms, which can introduce significant bias (8).

Given these limitations in the provision and access to adequate mental health services under the current system, applications of digital technologies are increasing (9–11). Digital technologies can help overcome the issue of accessibility in providing mental health services and can also alleviate recollection bias by offering real-time physiological digital markers to both physicians and patients. Commercialized services for some conditions, such as insomnia, already exist. For instance, the Sleep Healthy Using the Internet (SHUTi) service has been effective for insomnia and significantly reduced depression and anxiety symptoms (12). Moreover, studies on mobile intervention platforms for insomnia in South Korea have emphasized the potential of wearable devices (13). The mobile and wearable devices enable scalable sampling of the experiences and feelings of a patient (14), thus facilitating well-being reports collection systematically and objectively at scale (15). Wearable devices not only address accessibility concerns in traditional healthcare, but also enable healthcare providers to achieve more precise diagnoses through continuous collection of real-time patient biometrics, allowing physicians to analyze a patient’s condition over a broader spectrum (16).

Nevertheless, the information that wearable devices can collect is somewhat restricted in both quality and quantity compared with medical devices. Therefore, clinically relevant mental health data of the user is a priority for wearable devices. Research has subsequently expanded to heart rate variability (HRV) (17), referring to the small variations between heartbeat cycles. In a healthy human heart, a dynamic relationship exists between the parasympathetic nervous system (PNS) and the sympathetic nervous system, often referred to as autonomic nervous system balance. Consequently, HRV is associated with numerous psychiatric symptoms. It has been suggested that patients with depression exhibit lower HF power, which indicate a diminished regulatory ability of the parasympathetic nervous system and short-term flexibility of the autonomic nervous system, respectively (18). In addition, a meta-analysis has revealed that anxiety disorders are associated with significant reductions in both high-frequency and time–domain HRV metrics. This reduction may signify a failure of inhibition, characterized by a diminished capacity to inhibit responses, leading to decreased vagal outflow and lower HRV (19). Furthermore, it has been suggested that the LF/HF ratio, generally indicative of autonomic balance, may reflect aspects of an individual’s resilience profile (20).

However, HRV is traditionally measured using an electrocardiogram, which is time-consuming and resource-intensive. Therefore, various methods to measure HRV using scalable devices, such as wearables, have been developed (21, 22). Among them, the photoplethysmography (PPG) method is favored owing to its reliability compared with the gold standard electrocardiogram method (23–25). Therefore, in this study, we aimed to advance this exploration and investigate whether HRV measured using wearable devices could be applied to depression and anxiety. This study will elucidate whether real-time signals measured through wearable devices correlate with patient mental health. We therefore examined the association between HRV metrics collected in real-time while wearing a smartwatch and self-reported depression and anxiety in healthy adults.

The association between HRV and psychiatric symptoms is acknowledged in the literature (18–20, 36). From a physiological perspective, the abnormal serotonergic system observed in various psychiatric conditions may contribute to cardiovascular dysregulation through alterations in endocrine and autonomic functions (37). This relationship is thought to be partly modulated by the hypothalamic-pituitary-adrenal (HPA) axis. Furthermore, because the serotonin transporter is predominantly found in platelets – where serotonin acts as a vasoconstrictor – a potential pathogenic link between psychiatric conditions and cardiovascular dysregulations may exist (38).

In line with previous literature, we found that some HRV metrics were significantly correlated with clinical measures with self-report questionnaires. What distinctly sets our findings apart from previous studies is that we obtained HRV data from PPG signals collected via a highly accessible wearable device, suggesting its potential to monitor mental health across a wide demographic range. Furthermore, we observed somewhat different patterns of associations between HRV metrics and clinical measures. Previous studies have suggested that HF is significantly associated with depression or anxiety, as lower HF power is linked to stress, panic, or anxiety (39). The relationships between time–domain measures and clinical measures are complex, yet multiple correlations have been continuously reported (18, 19). In contrast, our study results demonstrated correlations between LF and depression or anxiety, rather than HF. Furthermore, the LF/HF ratio showed only one significant correlation coefficient with clinical measures. Moreover, only few correlations were observed between time–domain measures, except for SDNN, and depression or anxiety.

Nevertheless, our results are partially consistent with cardiac electrophysiology. The LF band (0.04–0.15Hz), which predominantly reflects baroreceptor activity under resting conditions (40), may plausibly be associated with mental health conditions. For the LF/HF ratio, while it is often thought to reflect the balance of sympathetic and parasympathetic nervous systems, it has also been suggested that this ratio does not always reflect autonomic balance (41). Regarding time–domain measures, SDNN, generally considered the gold standard for clinical HRV metrics and medical stratification of cardiac risk (42), showed more than one correlations with clinical measures. A Similar HRV metric, SDSD, also showed two significant correlations with clinical measures. On the other hand, the RMSSD, a measure of beat-to-beat variance in heart rate representing vagal-mediated changes reflected in HRV and related to the parasympathetic activity (43), showed only one significant correlation. Similarly, PNN50, although recognized as a less sensitive measure of the PNS, was not significantly correlated with clinical measures in this study. These results suggested that HRV is related to the balance of sympathetic and parasympathetic nervous systems; however, metrics biased toward the parasympathetic region are less likely to be associated with depression and anxiety.

Among the significant associations between HRV metrics and clinical measures, all metrics from the first two weeks were associated only with clinical measures at baseline or week 2. Similarly, HRV metrics from the following two weeks were associated with clinical measures at weeks 2 or 4. This suggests that HRV metrics may reflect temporal changes in psychiatric symptoms. However, the inconsistency in statistical significance across items, the relatively small number of overall participants and their measured PPG signals, and the division of the entire observation period into two halves create difficulty in comprehensively interpreting the temporal association between HRV and clinical symptoms based solely on the results of this study.

This study had several limitations. First, we recruited participants from specific contexts, limiting the generalizability of our results. Our study population consisted of well-educated young adults, not a diverse demographic. This limits the scalability of our findings, as groups such as the elderly – who have less access to digital technologies – might display different characteristics. Moreover, extending our results to children or adolescents should be done with caution, as their cardiac electrophysiology differs from that of adults (44). Second, most participants were asymptomatic or exhibited mild symptoms, as indicated by mean scores around three on the PHQ-9 and GAD-7. This necessitates cautious statistical interpretation of the correlations due to potential bias. A separate study with a different population is required to confirm clinical utility. Third, all symptom assessments were self-reported, which risks social desirability response bias (45) and may not accurately reflect true symptoms. Furthermore, since the survey was web-based, we cannot confirm that the actual respondents were the intended participants. Meta-analyses have shown that clinician-rated depressive symptoms have a significantly larger effect size than self-reported symptoms (46). Incorporating clinician interviews or clinician-rated scales could yield more comprehensive results. Fourth, although we excluded participants diagnosed with any psychiatric comorbidity, including alcohol and other substance use disorders, we could not rule out the use of substances like caffeine or alcohol during the actual experiment period. Nevertheless, considering our study’s aim to evaluate the utility of a smartwatch as a real-time monitor for HRV as a digital biomarker of psychiatric symptoms, the consumption of certain substances within daily intake ranges should be considered part of our research. Lastly, the quality of biomedical signals collected by smartwatches has been a subject of ongoing debate. Specifically, PPG signals are argued to not accurately represent HRV, especially under free-living conditions (47) or without controls for breathing (48). Nevertheless, we did filter out noise from the PPG signals throughout the processing pipeline (31). This study has demonstrated that estimated HRV from PPG signals is significantly correlated with ECG-measured HRV metrics, indicating the reliability of our estimated HRV.

Despite these limitations, our study is significant because it demonstrated that biomedical signals obtained through simple methods, such as common smartwatches, are associated with psychiatric symptoms such as depression and anxiety. Further research should aim to highlight the existing shortage of mental health services. Additionally, with the integration of digital technology, these methods can provide more advanced mental health services in conjunction with the digital platforms that have gained substantial attention. We look forward to future research on a broader range of mental health symptoms based on various biomedical signals, not just PPGs, and research on a wider patient population.

The datasets presented in this article are not readily available because of local ethical restrictions. Requests to access the datasets should be directed to shaun@kangwon.ac.kr.

The studies involving humans were approved by Institutional Review Board of KAIST (KH2020-027). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

YJ: Writing – original draft, Validation, Formal analysis, Conceptualization. SL: Writing – review & editing, Methodology, Investigation, Funding acquisition, Conceptualization. SP: Writing – review & editing, Software, Methodology, Investigation, Data curation, Conceptualization. JL: Writing – review & editing, Methodology, Funding acquisition, Formal analysis, Conceptualization.