Validação da medida de variabilidade da frequência cardíaca por smartphone frente ao eletrocardiograma em pacientes com depressão

Abstract

Heart rate variability (HRV) describes time fluctuations between consecutive heart beats, providing insight into the sympathetic and parasympathetic branches of the autonomic nervous system. In depressed patients, HRV has shown reduction due to autonomic dysregulation. Traditionally, its measurement is conducted using electrocardiography (ECG). A novel approach is measurement through a conventional smartphone via photoplethysmography (SPPG), which has not yet been explored in depressed patients. Thus, we developed PULSAR, an SPPG application which measures the user’s pulse wave using the smartphone camera and flash, without the need for external devices, to calculate HRV. To validate PULSAR, simultaneous 5-minute resting-state ECG and SPPG measurements were conducted in 15 healthy individuals and 15 depressed patients. In terms of analytical validation, the SPPG-derived HRV parameters demonstrated high correlations to their ECG- derived counterparts while systematically overestimating most variables. This misestimation is attributed to physiological pulse wave characteristics, motion artifacts and the low sampling rate in SPPG measurements. Regarding clinical validation, HRV parameters in the depressed population were significantly reduced compared to healthy controls in both ECG and SPPG recordings. The frequency-domain parameter LF power, which primarily reflects sympathetic nervous system activity, emerged as the most robust variable for both clinical and analytical validation. Its SPPG-derived value showed a strong correlation with its ECG counterpart (r = 0.84) and a significant reduction in depressed patients (p = 0.021). Future iterations of the PULSAR application should include covariate adjustment, focus on addressing its overestimation of HRV parameters, implementing longitudinal measurements in depressed patients, and potentially incorporating biofeedback techniques to introduce a therapeutic dimension.

Keywords: HRV, SPPG, Smartphone, Depression, Biomarker, ECG

Introduction

Heart rate variability and depression

Depression is a common illnesses, with an estimated 5.7% of adults affected globally [¹]. The average lifetime prevalence of major depressive disorder is projected to be around 15% in high-income countries [²]. While traditionally being diagnosed clinically and by exclusion, the search for appropriate cost-effective, easy-access diagnostic and predictive biomarkers is an ongoing and challenging task [³].

Heart Rate Variability (HRV) describes the time fluctuations in between consecutive heart beats. The heartbeat is regulated by the sinoatrial node, which in turn is influenced by the sympathetic and the parasympathetic nervous system (SNS and PNS) [⁴]. Thus, HRV offers a non-invasive insight into the autonomic nervous system (ANS) whose main branches are the SNS and PNS. A decrease in HRV is associated with a predominant SNS activity (“fight or flight”), whereas an elevated HRV indicates increased PNS activity (“rest and digest”) [⁵]. Different approaches of quantifying HRV have been established, whereas the time-domain and the frequency-domain parameters constitute the most common groups in addition to non-linear methods [⁴]. Various time-domain indexes quantify the amount of variability in inter-beat intervals and can be used to interpret both the general ANS and the specific PNS activity. For calculating the frequency-domain, the fluctuations of the heartbeat signal are assigned to different frequency bands using spectral analysis. The high frequency (HF) band can be used to draw conclusions about PNS activity, while the low frequency (LF) band is an indicator of both SNS and PNS activity, with a predominance of SNS influence [⁶, ⁷].

HRV is an active subject of research for psychiatry in general [⁸–¹⁰] and depression in particular [⁶, ¹¹, ¹²]. It has been demonstrated that patients with major depression show multiple significantly reduced indexes of the time-domain as well as significantly reduced LF and HF parameters in the frequency-domain [⁶]. The effects of altered SNS and PNS activity in depressed patients can be explained through the neurovisceral integration model, according to which emotional dysregulation is defined by interactions between cerebral affective systems and the ANS [¹³]. Although antidepressant medication seems to play a role in influencing HRV [¹⁴], its decrease in depressed patients appears to be independent of pharmacological treatments [⁶, ¹⁵].

Smartphone measurement of heart rate variability

The standard measurement method of obtaining HRV is electrocardiography (ECG) [⁴], which through high sampling rate and distinctive inter-beat intervals between R peaks allows for ideal measurement accuracy. An alternative, simpler way of detecting heart beats is photoplethysmography (PPG), which, in contrast, detects a pulse wave instead of an electrical signal [¹⁶]. The usage of pulse rate variability (PRV) has largely been applied as a validated surrogate of HRV but is not without challenges [¹⁷, ¹⁸]. In psychiatric research in particular, PPG derived HRV has shown high diagnostical potential [¹⁹]. The big advantage of PPG over ECG is its simplicity and widespread availability. Modern smartphones can turn into a PPG device by using the integrated flash as light source and the camera as photodetector [²⁰, ²¹]. The validity of approximating HRV by the use of smartphone PPG (SPPG) has been suggested in small-scale studies with healthy individuals [²²–²⁴].

As previously stated, the topic of HRV in psychiatry is intensely researched, and today’s universal use of smartphones potentially offers a simple and cost-effective way of estimating a depressed patient’s HRV. Although suggested for psychophysiological research [²⁵, ²⁶], to this date, the promising idea of approximating a depressed patient’s HRV via SPPG alone has not yet been implemented. Thus, we developed PULSAR, an SPPG-based HRV application, and validated its performance in depressed patients and healthy individuals through simultaneous ECG-based HRV assessment.

In this study, it was hypothesized that.

1. HRV parameters derived from SPPG correlate with those derived from ECG (analytical validation).

2. SPPG-derived HRV parameters differ significantly between depressed patients and healthy individuals (clinical validation).

Methods

Design and participants

We conducted a comparative, observational, cross-sectional validity study with 15 depressed patients and 15 healthy individuals. Depressed subjects (mean age 35.5 years, SD = 12.8 years, 9 males, 6 females) were recruited from the hospital’s specialized wards for depressive disorders. Inclusion criteria were age 18–65 years and a moderate to severe depressive episode following the criteria of ICD-10 as confirmed by a senior psychiatrist, with severity levels additionally assessed by the Montgomery–Åsberg Depression Rating Scale (MADRS) [²⁷]. MADRS is a ten-item questionnaire on depressive symptoms, assessed by a mental health professional. It yields a score from 0 to 60 and allows the classification of a depressive episode from mild to severe. Exclusion criteria were any additional currently untreated major psychiatric conditions, any current untreated/unstable non-psychiatric conditions, any forms of chronic cardiac arrhythmia or treatments interfering with ANS regulation of heartbeat. Healthy individuals (mean age = 35.7 years, SD = 13.6 years, 9 males, 6 females) were recruited from the private and professional network and matched precisely by age group and sex (See Table 1), being subjected to the same exclusion criteria.

This study was approved by the local ethics committee (Ethikkomission Zürich, BASEC ID 2023 − 01273). All participants provided written informed consent prior to participation, and their privacy rights were observed throughout the study.

Measurements

We conducted a short-term, 5-minute measurement of heart rate by ECG and smartphone simultaneously. Participants were resting in a 30° semi-recumbent position on an examination table and instructed to remain as immobile possible. A total of three consecutive measurements per participant was conducted to ascertain the measurement with the ideal conditions (absence of artifacts and noise in both measurement modalities).

ECG

For calculating ECG derived HRV, only the R peak intervals are relevant [²⁸]. Therefore, we used the bipolar limb recordings as sufficient measurement modalities [²⁹]. For ECG measurement we used Kalamed KEC-1000^®, which recorded the electrical signal with a sampling rate of 500 Hz (Hz). ECG curves were plotted with the corresponding Kalamed ^® software and exported in the CSV file format. We then chose the ECG channels with the most prominent R peak for HRV analysis visually.

SPPG

For recording the SPPG curve, we developed the Android ^® application PULSAR. Upon launching the application, it assigns the participant a distinctive identifier and shows user instructions. The user’s index finger is placed firmly on the smartphone camera in a way that the finger fully covers the camera, and the smartphone flash illuminates the finger (see Fig. 1). During measurement, pulse peaks are detected in real-time and beats per minute are calculated and displayed. In addition, the user is given feedback on each detected pulse peak by slight vibration of the smartphone and a visual pulsating signal on the display. After the 5-minute measurement, HRV parameters are calculated (see HRV Analysis below) and displayed to the user and are ready for export for further analysis.

PULSAR works by recording a video of the finger placed on the camera with a sampling rate of circa 30 Hz and adding up the amount of the red color in each video pixel (sum of red). Whenever a pulse wave passes the finger, the amount of red decreases due to the increased blood volume in some pixels, which results in a decrease of the sum of red in each video frame. Meanwhile, the smartphone flash illuminates the user’s finger and thereby accentuates the changes in red color. The SPPG pulse wave is therefore recorded by calculating the alterations of the color red over time (See Fig. 2) in the user’s index finger. Under ideal conditions, both the systolic pulse peaks and diastolic peaks can be distinctly recorded. The delay between the ECG’s R-peak (ventricular depolarization) and the pulse peak arises from the time required for the pulse wave to propagate from the heart to the fingertip. For pulse wave detection, we used pre-existing open-source code from Erdosi Balint’s HeartBeat project [³⁰], which we adapted and expanded for the PULSAR application.

We used the same Android ^® phone (Realme C11 ^®) for all measurements to ensure the same testing conditions for all participants. Recordings were performed using default camera exposure and illumination settings without manual parameter adjustments.

HRV analysis

Analysis of HRV parameters (see Table 2) was conducted using HeartPy (Version 1.2) [³¹], an open-source Python library, previously validated for ECG and PPG assessment of HRV [³²–³⁴]. Initially, ECG curves were cropped to match the exact measurement timespan of the 5-minute SPPG measurement. ECG and SPPG curves were analyzed using the process function of HeartPy, using Welch’s method [³⁵] to extract the frequency spectrum for the frequency-domain. For PPG signals, pulse peaks were used as fiducial points for pulse-to-pulse interval estimation. For ECG, R-peaks were detected and NN intervals derived. HeartPy’s standard bandpass filtering and artifact rejection algorithms were applied. In some cases, SPPG curves showed some baseline wandering [³⁶], which posed a challenge to the calculation of peak-peak intervals. We applied HeartPy’s filtering function remove_baseline_wander to correct the affected curves. In rarer instances, SPPG pulse peak amplitudes were not distinctive enough to obtain an accurate peak-peak interval. To solve this, the enhance_peaks function of HeartPy was applied. No interpolation, upsampling, or exclusion of recording segments was performed.

Methods (cont.)

Statistical analysis

For comparison between ECG measurements and Smartphone-based PULSAR analysis: Descriptive statistics were calculated to summarize the distributional aspects of both ECG and smartphone data. For each assessed HRV variable, the mean, median, standard deviation, and interquartile range were computed. A paired t-test was conducted for each variable to identify statistically significant differences between ECG and smartphone measurements. P-values were reported to assess the significance of these differences. Pearson correlation coefficients were used to quantify the linear association between the two methods and results were visualized in a heatmap to provide an overview of potential discrepancies. To visually inspect the relationship between measurements from both methods, scatterplots were created for each variable.

To evaluate systematic deviations and biases between the two measurement methods, metric-specific Bland-Altman analysis was performed for all assessed HRV parameters (time-domain, frequency-domain, and non-linear measures). This method plotted the mean of the two measurements against their difference for each variable, providing visual assessment of parameter-specific agreement, systematic deviations, outliers and potential proportional bias or heteroscedasticity beyond correlation-based assessments. Additionally, a statistical bias analysis was conducted to compare mean biases between healthy and depressed groups using t-tests to determine whether observed biases differed significantly.

For method comparison analyses, paired-sample effect sizes (Cohen’s dz) were calculated to quantify the magnitude of systematic differences between ECG and smartphone measurements.

Comparison between patients with depression and healthy controls: To assess the validity of the smartphone method in distinguishing between healthy and depressed individuals and its agreement with ECG data, a comprehensive analysis was conducted. Descriptive statistics and mean comparisons were performed for each variable between healthy and depressed individuals, separately for each method (ECG and smartphone) to determine whether differences between the groups were statistically significant. To account for multiple comparisons, the Benjamin-Hochberg correction (FDR-BH) was applied, balancing Type-I and Type-II errors while maintaining statistical power. Given the high correlations between HRV parameters, a stricter Bonferroni correction would have been too conservative, potentially masking true group differences. FDH-BH retained the highest number of significant results, reinforcing the robustness of observed differences without inflating false positives. Z-scores were calculated for ECG and smartphone variables for standardized comparisons, and the results were illustrated using box-and-whisker plots for healthy and depressed individuals, providing a visual representation of group differences.

Finally, to evaluate the predictive accuracy of both methods in distinguishing between depressed and healthy individuals, receiver operating characteristic (ROC) analyses were performed. AUC values, sensitivity, and specificity were calculated for each HRV variable.

In our statistical analysis we omitted the VLF parameter since its measurement is discouraged in short-term (e.g. ≤ 5 min) analysis of HRV. The HF_NU parameter was excluded as it is mathematically complementary to the LF_NU parameter [⁴], resulting in identical correlations and providing no additional analytical value. RMSSD and SD1 likewise quantify redundant information [³⁷]. Both parameters were nevertheless retained in the analysis, as SD1 is required for calculation of the SD1/SD2 ratio, and reporting both metrics facilitates comparability with existing HRV literature.

Results

Comparison of smartphone and ECG results

Smartphone derived HRV values yielded higher mean values for most HRV parameters, exhibiting significant paired differences (see descriptive results in Table 3). Effect sizes (Cohen’s dz) indicated large systematic differences for most time-domain parameters, whereas LF showed only negligible effects. However, moderate to high correlations were found for most variables (e.g. BPM, P_TOTAL, LF), indicating a strong concordance between the two measurement methods (see heatmap in Fig. 3 for further details). In contrast, variables with lower associations, such as SD1/SD2 or LF_PERC, suggest greater variability between methods.

Accordingly, scatterplots in Fig. 4 demonstrate a clear alignment along the trendlines, indicating a strong overall agreement between the two methods. For variables such as BPM, LF, or P_TOTAL, data points closely follow the line of best fit, suggesting consistent measurement performance. Systematic deviations are observed in variables such as SDNN, SDSD, PNN20, or SD1, where smartphone measurements tend to be slightly higher at lower ECG values but converge and align at higher levels. In contrast, variables such as SD1/SD2 and LF_PERC exhibit a greater degree of scatter, indicating potential measurement variability while still suggesting an underlying agreement trend.

Similarly, the Bland-Altman analyses in Fig. 5 demonstrate an overall strong agreement between the two methods for most variables. Variables such as BPM, SDNN, and SD2 show differences that are consistently centered close to the zero line, with only a few outliers outside the 95% limits of agreement.

In contrast, systematic deviations are observed for variables such as PNN20, RMSSD, and SDSD, where smartphone values tend to be higher at lower mean levels but align more closely with ECG measurements as the means increase. These patterns suggest a systematic (and thus predictable) shift, which further supports the trends identified in the scatterplots.

The bias analysis confirms a consistent level of agreement between the ECG and smartphone measurements across healthy and depressed groups for most variables (see Table 4). A clear exception is SD1/SD2, where a significant group-specific difference in bias suggests a systematic deviation. Additionally, SDSD and PNN20 variables are close to significance, suggesting trends toward group-specific bias that warrant further investigation.

Results (cont.)

For most variables, including BPM, SDNN, LF, HF, and P_TOTAL, the mean biases remain comparable across groups, underscoring the robustness of both measurement methods. These findings complement the Bland-Altman results, supporting the reliability of the smartphone as a viable alternative to ECG for most HRV parameters.

Comparison of healthy and depressed population

The results of the group comparisons indicate that ECG-based HRV variables effectively differentiate between depressed and healthy individuals, with smartphone measurements largely replicating these findings (See Table 5; Fig. 6a and b). To control for multiple comparisons, p-values were adjusted using the FDR-BH correction, ensuring a balanced approach between Type-I and Type-II errors. BPM and LF show strong agreement between methods, with similar means and consistent group differences, indicating their reliability in detecting depressive states. Systematic shifts were observed for SDNN, SD2, and P_TOTAL, where smartphone measurements tended to be higher than ECG values, yet group differences remained comparable between the two assessment methods. These differences remained statistically significant after correction, supporting the robustness of these HRV markers. In contrast, RMSSD, HF, and PNN20 did not remain significant after correction, indicating limited discriminatory power.

Results

Building on these findings, we assessed the predictive accuracy of both methods by analyzing AUC values, sensitivity, and specificity for each variable. (See Table 6). Overall, AUC analysis shows high AUC values and moderate- to high sensitivity and specificity across time- and frequency domain HRV variables in both ECG and smartphone measurements, suggesting a strong discriminatory power for distinguishing between our healthy and depressed population. However, some variables, such as LF_PERC, HF_PERC and LF_NU showed inconsistent results. While the AUC values for these metrics are moderate in some cases, their overall discriminative ability (and thus clinical utility) remains limited due to minimal and statistically non-significant group differences observed earlier.

Results (cont.)

Notably, SD1/SD2 emerges as a variable where smartphone measurements outperform ECG by means of group differentiation, underscoring the smartphone’s potential competitive performance for selected metrics.

For BPM, the AUC, sensitivity and specificity were identical across ECG and smartphone recordings. This reflects the almost perfect rank correlation between both measurements (Spearman ρ = 0.99, p < 0.001; not tabulated), resulting in virtually identical participant ordering and therefore identical ROC characteristics.

Discussion

This study presents a novel modality for approximating reduced HRV in depressed patients using only a standard smartphone without external devices, validating it against the state-of-the-art assessment of HRV via ECG. We confirmed Hypothesis 1 by demonstrating that SPPG-derived HRV parameters are highly correlated to their ECG based counterparts during simultaneous 5-minute resting measurements in patients with depression and healthy individuals. Additionally, our depressed population exhibited a significant reduction in HRV parameters across time- and frequency domain measures in both ECG and SPPG recordings compared to the healthy group, confirming Hypothesis 2. While previous studies have either extensively examined altered HRV in depression or the usage of (S)PPG as a surrogate of the gold standard ECG, to the author’s knowledge, this is the first application of a smartphone-only approach in clinically depressed patients.

When using PPG-derived pulse signals as a surrogate for HRV, it is important to consider the fundamental physiological differences between cardiac electrical activity and the peripheral pulse waveform [¹⁷, ¹⁸]. PPG yields a smooth signal, and its fiducial point detection proves an important challenge. Moreover, peripheral pulse is subjected to autonomic tone and peripheral perfusion. Thus, pulse rate variability (PRV) can be influenced by numerous physiological influences, such as pulse transit time variability, vascular characteristics (e.g. vascular tone or influence of medication), baroreflex influences, site of PPG measurement and ectopic or arrhythmic events. These physiological considerations underscore that PRV should be interpreted as a related but not identical construct to HRV.

Previous studies, however, have demonstrated strong correlations between ECG- and SPPG derived HRV parameters across time- and frequency domain measures [²², ³⁶, ³⁸]. Regarding analytical validation, our findings replicated these results for BPM and the time- and frequency domain parameters with one exception in each case. However, SPPG measurement tended to overestimate most HRV variables, particularly at lower values, leading to significant differences between both measurement methods. An exception to this trend was the LF power parameter, which exhibited both a high correlation and a non-significant difference between SPPG and ECG. The robustness of this parameter was confirmed by additional analysis (Bland-Altman and bias analysis). We partly attribute this systematic overestimation to motion artifacts during our SPPG measurements. Involuntary finger movement during (S)PPG measurement can introduce gaps and falsely detected pulse peaks, significantly altering HRV results [³⁸].

PULSAR has a relatively low sampling rate of 30 Hz, while established HRV literature recommends substantially higher sampling rates (≥ 250 Hz) for accurate interbeat discrimination [⁴], especially for frequency-sensitive parameters (e.g. RMSSD or HF). Previous studies have shown a systematic overestimation of PPG-derived HRV parameters at lower sampling rates, with the magnitude of overestimation increasing as the sampling rate decreases [³⁹, ⁴⁰]. This effect is attributed to a decrease in distinction of the pulse peak in lower sampling rates, leading to reduced precision in quantifying the interbeat intervals [³⁹]. Modern smartphones video cameras can reach a sampling rate of up to 240 Hz, which can provide higher temporal precision for pulse interval extraction [⁴¹]. Moreover, upscaling is possible through interpolation techniques (e.g. parabola interpolation, cubic-spline, and linear regression methods) and can be recommended for PPG sampling below 32 Hz [⁴²]. However, those modalities do not increase the intrinsic temporal resolution of the recorded signal. To accurately quantify HRV parameters using PULSAR, interpolation and calibration algorithms to correct for overestimation need to be implemented in the application. Nevertheless, current results remain sufficiently reliable for intra- and interindividual comparison of SPPG derived HRV parameters.

Recent systematic reviews and meta-analyses have consistently shown significant increases in BPM and decreases in HRV parameters across time- and frequency domain measures in depressed patients [⁶, ¹¹, ³⁹]. In terms of clinical validation, our study fully corroborates these findings through ECG-derived parameters and validates them for SPPG variables except for SDSD. This not only confirms previous research highlighting reduced HRV in depression but more importantly underscores the potential of SPPG in measuring said reduced HRV. Notably, the LF power parameter stood out again due to its strong performance: It exhibited a significant reduction in descriptive group comparisons (p = 0.021), a narrow interquartile range with minimal overlap in the box-and-whisker plots, and robust diagnostic accuracy in the ROC analysis for detecting depression (AUC = 0.81, sensitivity = 0.87, specificity = 0.80). Similarly, BPM demonstrated identical ROC characteristics across ECG and smartphone recordings. This reflects the very high rank concordance between both measurement modalities, indicating that discriminative performance is preserved despite minor systematic differences in absolute values. In contrast, non-linear parameters (SD1, SD2) as well as LF_PERC, HF_PERC and LF_NU showed limited discriminative relevance in our sample, which is consistent with the comparatively minor emphasis placed on these parameters in the existing literature.

These findings prompted a closer examination of the frequency domain of HRV. ANS influences lead to a fluctuating heart rate signal over time. Through spectral analysis, these oscillatory components can be broken down into different frequency bands [⁴³], with the LF band being set at 0.03–0.15 Hz and the HF band at 0.15–0.4 Hz [⁴⁴]. The LF band is proposed to be influenced by both PNS and SNS activity, whereas SNS activation appears to play a dominant role in its increase. The HF band is considered to be solely influenced by PNS activity [⁷]. In our depressed population, both LF and HF power parameters showed significant decreases, which suggests reduced activity of SNS as well as PNS branches of the ANS in depressed patients. Given LF power’s robustness as the most reliable variable measured by the PULSAR application, SPPG-derived measurements of this parameter can be proposed as a potential biomarker for gaining meaningful insights into the PNS and SNS activity of depressed individuals.

HRV has been suggested as a viable diagnostic biomarker for mental disorders in general [⁴⁵–⁴⁷] and depression in particular [¹¹, ⁴⁸, ⁴⁹]. Patients with recent suicide attempts have exhibited significant reduction in HRV parameters across time- and frequency domain measures [⁵⁰]. Beyond its diagnostic utility, heart rate analysis has demonstrated predictive potential for antidepressant medication with serotonin-norepinephrine reuptake inhibitors (SNRIs) [⁵¹]. Furthermore, a recent study utilized HRV parameters to discriminate between depressive responders and non-responders to ketamine therapy, underscoring its potential as a predictive biomarker for treatment outcomes [⁵²]. Given that SPPG measurement of HRV offers a non-invasive, easy-access and cost-effective modality [²¹], its potential in detecting, stratifying and analysis of the predictiveness of HRV-derived biomarkers merits significant consideration.

Strengths and limitations

We investigated a total of 14 common HRV parameters in addition to beats per minute (BPM) and conducted rigorous descriptive and visual analyses. This allowed us to perform a broad analytical and clinical validation process and ascertain the specific SPPG derived HRV parameters with the highest (e.g. LF power) and the least (e.g. SDSD) potential.

For this study, we used a relatively small dataset of n = 30 participants. While this sample size was deemed appropriate for an initial feasibility study in this research area, future studies should include a larger cohort to strengthen our findings.

Several important covariates known to influence BPM and HRV, including antidepressant medication, age, sex, physical fitness, smoking and time of the day [⁵³] were not systematically accounted for in this study. Owing to the feasibility-oriented design and limited sample size, covariate adjustment or subgroup analyses were not deemed statistically justified. Future studies with larger cohorts should address these covariates in line with current HRV reporting guidelines.

Skin pigmentation was not systematically assessed in this study and represents an important limitation. Increased melanin concentration alters light absorption and reflectance properties of the skin, which can reduce PPG signal amplitude and signal-to-noise ratio [⁵⁴]. This may introduce systematic measurement differences or increased data loss in individuals with darker skin tones, thereby limiting the generalizability of the present findings. Given the feasibility-driven design and limited sample size, no participant skin tone information (e.g. Fitzpatrick scale or objective melanin index) was collected. Future studies should explicitly assess skin pigmentation and evaluate its influence on SPPG-derived HRV metrics and should consider mitigation strategies such as wavelength selection or algorithmic compensation to improve equity and external validity of measurements.

Identical measurement conditions (same smartphone, reclining position, finger placement, room lighting) were maintained for all participants. In practical applications, however, PULSAR would be used in a variety of settings where external factors, particularly lighting conditions, could affect measurement precision.

The study did not include complementary agreement statistics (e.g. ICC or error-in-variables regression models), which may further refine reliability estimates in future larger cohorts.

Physiological contributors to PRV–HRV divergence were not explicitly modeled and represent an important limitation of this feasibility study. Additionally, PULSAR’s SPPG measurements seem to be susceptible to movement artifacts, which can distort the recorded pulse wave and the derived HRV parameters. To mitigate this, we selected the result least affected by artifacts from three measurements and applied HeartPy’s filtering functions. For future iterations, we aim to implement detection and correction algorithms into PULSAR to automatically reduce movement artifacts. Moreover, PULSAR’s relatively low sampling rate of 30 Hz presents a limitation to measurement accuracy. Future iterations should aim for higher sampling rate for acquisition of SPPG signal and additionally implement interpolation algorithms for signal upscaling.

The study did not include additional comparison with a clinical-grade PPG sensor, which limits the ability to disentangle smartphone-specific measurement effects from fundamental physiological differences between PPG and ECG derived HRV. Moreover, in our current study, no supplementary SSPG signal and waveform characteristics (such as pulse amplitude, pulse arrival time, pulse shape indices, perfusion-related measures and quality metrics) were systematically analyzed or reported. Future multimodal research incorporating ECG, SPPG, and dedicated PPG sensors as well as incorporation of PPG signal waveform characteristics and quality indicators may enable a more precise characterization of sensor- versus physiology-driven discrepancies.

Future research should focus on longitudinally assessing HRV parameters in depressed patients using PULSAR, with particular emphasis on the LF power parameter. Such studies would not only reinforce our findings but also provide patients with meaningful insights into their ANS by offering a user-friendly, HRV-derived score. Numerous studies have demonstrated the efficacy of HRV biofeedback training [⁵⁵, ⁵⁶]. Future iterations of our smartphone application could integrate biofeedback techniques, such as breathing exercises [⁵⁷], adding a therapeutic dimension to PULSAR.

Conclusion

In terms of analytical validation, our study demonstrated strong to very strong correlations between most SPPG-derived HRV parameters and their ECG counterparts. However, SPPG measurements exhibited systematic overestimations, particularly at lower HRV values. We attribute this to physiological PRV-HRV divergence, motion artifacts and reduced sampling rate. Regarding clinical validation, our depressed population revealed a significant reduction in HRV measured by SPPG- and confirmed by ECG-based measurements. Among the assessed variables, the frequency-domain LF power parameter (r = 0.84, p = 0.021, AUC = 0.81, sensitivity = 0.87, specificity = 0.80), which is chiefly dependent on sympathetic nervous system activity, emerged as the most reliable metric, demonstrating strong potential for detecting HRV reductions in depressed patients. These findings support the theory of autonomic dysregulation with altered sympathetic and reduced parasympathetic nervous system activity in depression.

In conclusion, our study demonstrated the feasibility of approximating HRV reduction in depression using SPPG alone, without the need for external devices. This further reinforces the role of HRV as a biomarker in depression and highlights smartphone-based measurement as an accessible and cost-effective modality for its acquisition. Future studies should include covariate adjustment, honor the impact of skin pigmentation, aim to address the systematic overestimations observed in SPPG measurements (e.g. signal upscaling, additional PPG sensors), conduct longitudinal assessment of HRV in larger groups of depressed patients and incorporate biofeedback techniques into our smartphone application.

Author contributions

Lion D. Comfort : Writing—original draft, investigation, software, conceptualization, data curation. Mario Müller : Writing—original draft, formal analysis, visualization, validation, data curation. Erich Seifritz : Writing—review and editing, supervision. Sebastian Olbrich: Writing—review and editing, project administration, supervision, conceptualization, validation.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Open access funding for this publication was provided by the University of Zurich. Open access funding provided by University of Zurich.

Data availability

Data will be made available upon reasonable request to the corresponding author.

Declarations

Conflict of interest

The authors declare no competing interests.

Ethical approval

This study was approved by the local ethics committee (Ethikkomission Zürich, BASEC ID 2023 − 01273) on November 23, 2023.

Informed consent

All participants provided written informed consent prior to participation, and their privacy rights were observed throughout the study.

References

Associated Data

Data Availability Statement

Data will be made available upon reasonable request to the corresponding author.