This study recruited a total of 127 participants, comprising university students and staff members, from the last author’s affiliated institution in Shenzhen, China (72/127, 56.69% male; mean age 22.78, SD 1.97 years). Participants were recruited through on-campus posters and electronic posters distributed via student social media platforms. Eligibility was restricted to university students and staff, ensuring that all participants were at least 18 years of age and had access to university facilities. Participants were required to have no clinically diagnosed mental disorders or cardiovascular conditions. No additional exclusion criteria were applied.

During data collection, participants were instructed to rest their arm on a table, hold a Mi 8 SE smartphone (Xiaomi) in their left hand, and place their index finger over the rear camera. Using the Heartily Happy app, developed by our team and publicly available on Google Play at the time of the study, participants recorded 7 fingertip videos, each lasting 4 minutes, at a resolution of 120×160 pixels. To further evaluate the accuracy and practical applicability of the proposed approach, participants were additionally instructed to place the index finger of their right hand on a US Food and Drug Administration–cleared fingertip oximeter (FDA 510(k): K082641; CMS50D, Contec Medical Systems Co Ltd) to simultaneously record pulse waveforms. This study is part of a broader research program examining the relationship between pulse waveforms and mental health. The same dataset has also been used in studies investigating blood pressure [13] and HRV [37].

The Heartily Happy app used the smartphone’s built-in flashlight to illuminate the fingertip and captured video at approximately 30 frames per second (Figure 2) [13,39]. Raw video frames, initially recorded in YUV format, were converted to red green blue color model for subsequent analysis. Waveform signals were extracted by computing a SD–weighted average across the 3 color channels.

Following peak detection, both photoplethysmography and oximetry signals were segmented into beat-to-beat intervals (BBIs). Waveform features were calculated for each BBI, and the median value of each feature across all BBIs was used to represent the sample. Detailed signal-processing procedures are described in [13,39].

Because pulse waveform length is inversely related to HR, following a conventional, although controversial, practice in aortic blood pressure research, waveform features were normalized to a reference HR of 75 beats per minute to control for HR-related effects [40-42]. In addition, since the timing of reflected waves depends on vascular path length, which is associated with body height, waveform features are often further normalized by stature; for example, the stiffness index is defined as body height divided by peak-to-peak time (PPT) [43]. Therefore, in this study, waveform features were normalized to a reference height of 170 cm to reduce stature-related interindividual variability. Sex was not included as a separate normalization factor, as its effects on waveform morphology are largely mediated by differences in body height.

This study considers 3 categories of waveform features commonly cited in pulse waveform analysis but underexplored in psychological research (Table 1 [13]). Time-domain indicators encompass measures of intervals between characteristic waveform points, amplitude differences among these points, and the total area under the waveform curve. Curvature-domain features are derived from the second derivative of the photoplethysmography waveform, often termed acceleration photoplethysmography, and quantify waveform contour and curvature [44]. Frequency-domain features are obtained via Fourier transform decomposition, with power spectral density functions characterizing the strength of each frequency component (Figure 3) [13].

To assess well-being, 3 self-report measures were administered: the Satisfaction With Life Scale (SWLS) [53], the Subjective Vitality Scale (SVS) [54], and the positive and negative affect schedule (PANAS) [55]. The SWLS comprises 5 items rated on a 7-point Likert scale ranging from “strongly disagree” to “strongly agree.” The SVS comprises 7 statements scored on a 7-point scale from “not at all true” to “very true.” The PANAS consists of 20 adjectives assessing positive and negative affect over the past week, with responses provided on a 5-point scale (1=“very slightly or not at all,” 5=“extremely”). Separate scores for positive affect (PANAS-P) and negative affect (PANAS-N) were computed to enhance interpretability.

Common mental health measures included the Patient Health Questionnaire-9 (PHQ-9) [56] and the Generalized Anxiety Disorder-7 (GAD-7) scale [57]. The PHQ-9 includes 9 items rated from 0 (“not at all”) to 3 (“nearly every day”), with total scores categorizing depressive symptom severity from minimal to severe. The GAD-7 comprises 7 items using the same response scale and classifies anxiety levels as normal, mild, moderate, or severe based on the summed score.

Momentary emotional states were captured using the Self-Assessment Manikin (SAM) [58], a nonverbal pictorial instrument representing affective valence and arousal. Participants selected the figure that best reflected their current level of pleasure (valence) and activation (arousal). The dominance dimension was omitted, as it was not directly relevant to the study’s objectives. All questionnaires were completed before signal collection.

Associations between individual pulse-waveform features and psychological measures were first examined using simple linear regression, which provides a transparent and interpretable baseline for evaluating feature–outcome relationships and facilitates comparison with prior psychophysiological studies relying on univariate analyses.

The data exhibited a nested structure: psychological scales were administered once per participant, whereas pulse-waveform features consisted of multiple repeated measurements per participant. Treating all observations as independent under standard ordinary least squares assumptions would therefore be inappropriate and could result in underestimated standard errors and biased statistical inference. To explicitly address this nonindependence, 2 complementary analytic strategies were implemented.

The first strategy involved participant-level aggregation, whereby waveform features were summarized for each individual using median values and were entered into ordinary least squares models. This approach mitigates within-subject dependence by aligning the unit of analysis for physiological features with that of the psychological measures. The second strategy applied cluster-robust standard errors at the original observation level, allowing statistical inference to account for repeated measurements within participants while preserving within-person variability.

From a theoretical perspective, participant-level aggregation may be more appropriate in this context, as cluster-robust standard errors can be less suitable when the dependent variable is constant across observations within the same cluster [59]. However, this aggregation approach necessarily entails a loss of within-participant information. Taken together, these 2 analytical strategies offer complementary protections against bias arising from nonindependent observations.

In addition, because multiple waveform features were tested following feature selection, statistical significance was evaluated using a Bonferroni-corrected threshold [60,61]. Although conservative, this correction was adopted to control the error rate, with the acknowledgment that it may exclude features of potential substantive relevance [62].

To evaluate predictive utility beyond univariate linear models, machine learning methods were further used to model psychological outcomes from multivariate pulse-waveform features. These methods are well-suited to this context because they can capture nonlinear relationships, complex feature interactions, and high-dimensional structures that may not be adequately modeled using traditional regression. Model performance was assessed using 5-fold cross-validation to obtain stable estimates of generalization performance while maintaining sufficient training data in each fold. Among the evaluated algorithms, random forest (RF) models were selected due to their robustness to multicollinearity, resistance to overfitting in moderate sample sizes, and ability to model nonlinear associations without strong parametric assumptions.

In each fold, the RF model was trained on data from a subset of participants and evaluated on held-out participants. Performance was quantified using out-of-sample metrics, including mean absolute error and the correlation between predicted and observed values, averaged across folds. The RF models were implemented using the default hyperparameters provided by scikit-learn. No manual hyperparameter tuning was performed. Specifically, the number of trees was set to 100, node splitting used the mean squared error criterion, trees were grown to full depth (no maximum depth specified), and bootstrap sampling was enabled. Other parameters remained at their library default values.

To prevent data leakage and preserve independence between training and evaluation, all observations from a given participant were assigned exclusively to either the training or test set. For all data analyses, an IQR multiplier of 1.5 was applied for outlier removal.

Agreement between predicted and observed values was further assessed using Bland–Altman analysis. Unlike correlation-based metrics, this approach evaluates absolute agreement and identifies potential systematic bias between predictions and ground-truth values, providing a complementary assessment of model validity that is particularly relevant for evaluating practical predictive performance.

For oximeter-photoplethysmography agreement analysis, 3 time-domain pulse waveform features commonly reported in prior studies, RI, FN, and diastolic time, were selected as representative indicators. Pearson correlation analyses were conducted to assess the linear association between the 2 measurement systems for each feature. Because correlation does not imply measurement agreement or interchangeability, Bland–Altman analysis was additionally performed to evaluate method agreement by plotting the differences between measurements against their means, thereby visualizing systematic bias and the 95% limits of agreement. The mean difference indicates directional bias, whereas the limits of agreement define the expected range of measurement differences.

Agreement was considered acceptable when most observations fell within the limits of agreement without systematic trends across the measurement range. Together, correlation and Bland–Altman analyses enabled a comprehensive evaluation of both association strength and measurement agreement between the smartphone-based photoplethysmography system and the clinical-grade fingertip oximeter.

The study protocol was approved by the Ethics Board of the Department of Psychology at Tsinghua University (number 201910). All study procedures were conducted in accordance with institutional and national ethical guidelines for human subjects research. Written informed consent was obtained from all participants prior to participation.Each participant received 50 renminbi (approximately US $7) as compensation for participation. Participants were informed about the study procedures and their right to withdraw at any time without penalty. All data were deidentified prior to analysis to protect participant privacy. Access to the original identifiable data were restricted to the first author, and no personally identifiable information was included in the analyses or reported results.

Of the 127 participants recruited, 121 completed all questionnaires (71/121, 58.68% male; mean age 22.79, SD 1.87 years), and 113 of these participants completed the pulse-waveform recording (67/113, 59.29% male; mean age 22.74, SD 1.83 years). The 2 participants who exhibited severe depressive symptoms, as indicated by questionnaire assessments, were excluded at this stage.

From the remaining 113 participants, a total of 766 waveform samples were collected (mean 6.78 samples per participant), comprising 229,165 BBIs (BBIs; mean 299 BBIs per sample). Of these, 100 samples were excluded due to partial feature loss (666 remaining; 86.9%), primarily resulting from poor signal quality that prevented reliable peak and valley detection (eg, HR could not be computed), although all recordings were successfully completed. Additionally, 19 participants were excluded because of problematic questionnaire responses identified through manipulation checks or severe depressive symptoms, and 5 samples with a signal quality index below 0.5, as proposed by Liu et al [39], were removed. After all exclusions, 496 valid samples from 91 participants were retained for analysis (Figure 4).

Features exhibited strong within-domain correlations across the time, curvature, and frequency domains, whereas cross-domain correlations were comparatively weak (Figure 5). To address high collinearity in the data, a systematic feature-selection procedure was applied prior to statistical modeling to reduce multicollinearity and improve model stability, parameter estimation, and inferential validity. Within each domain, feature pairs with absolute correlations exceeding a prespecified threshold of 0.5 were considered highly collinear and addressed through sequential elimination, after which the top 2 variables were retained for further analysis. Following this procedure, the final feature set used in subsequent analyses comprised estimated reflection index, crest time (CT), the third curvature minimum (F/A), the fourth curvature minimum (H/A), the first power spectrum density component (rPSD1), the baseline of Fourier decomposition (V0), and systolic blood pressure.

In univariate regression and correlation analyses, we examined associations among 8 psychological measures, 9 pulse waveform features, and blood pressure. The results indicated that negative psychological states were primarily associated with time- and curvature-domain waveform features (Figure 6). Depressive symptoms (PHQ-9) were significantly associated with F/A and V0, with the association between PHQ-9 and estimated reflection index remaining significant after Bonferroni correction. Generalized anxiety (GAD-7) also showed a Bonferroni-corrected significant association with F/A. Negative affect (PANAS-N) was associated with both CT and F/A.

In contrast, positive affect exhibited fewer significant associations with waveform features: SWLS was correlated with F/A, and PANAS-P was correlated with CT. Regarding momentary emotional states, valence was associated with F/A and H/A, whereas arousal was associated with CT and H/A.

In this study, 9 features selected through feature engineering were used as inputs to predict mental health outcomes. The machine learning results (Figure 7) indicated that waveform-related features were effective in predicting mental disorders (P<.001 for PHQ-9, GAD-7, and PANAS-N), whereas predictive performance for positive emotional states was substantially weaker. With respect to short-term emotional fluctuations, heartbeat waveform features showed no clear association with valence but were evidently influenced by arousal. Bland–Altman analysis further demonstrated that, for outcomes with significant predictive correlations, the machine learning predictions exhibited slight but not substantial systematic bias. The 3 most important features identified by the RF models across all dependent variables were CT, F/A, and systolic blood pressure.

The results (Figure 8) indicate reasonable agreement between waveform features obtained from the oximeter and those derived from the smartphone-based photoplethysmography method. For all 3 features analyzed, correlation coefficients were statistically significant (P<.05), and Bland–Altman analyses revealed minimal systematic bias. However, consistent with the optical boundary effects inherent to smartphone photoplethysmography acquisition discussed in the waveform metrics section, the reflection index exhibited some degree of distortion. In contrast, time-related waveform features were largely unaffected by these optical limitations and showed more consistent agreement across the 2 measurement modalities.

Overall, the findings support the study’s hypotheses and objectives by demonstrating that fingertip pulse-waveform signals recorded using a standard smartphone can be used to predict individual psychological health status. This work makes 3 principal contributions.

First, we introduce and validate a cost-effective, noninvasive approach that relies solely on a smartphone camera to capture subtle skin color variations, reconstruct digital arterial pulse waveforms, and infer mental health status. Across both univariate linear regression and multivariate machine learning analyses, waveform features showed consistent and statistically significant associations with psychological indicators. Beyond extending smartphone-based waveform analysis to psychological research, these findings provide additional empirical support for pulse-waveform–based psychological assessment more broadly [34,35], as well as for the use of smartphones in related physiological applications such as HRV and blood pressure analysis [63,64].

Second, our results reveal associations between pulse-waveform features and psychological variables across multiple feature domains. Consistent with findings reported by Kontaxis et al [34] and Celka et al [65], negative psychological conditions were strongly linked to waveform characteristics, suggesting that long-term psychological states may induce structural or functional changes in the arterial system. However, contrary to our initial expectations, these associations were primarily observed in curvature- and time-domain features rather than in the frequency domain. As one of the first studies to incorporate frequency-domain pulse-waveform features into mental health research, we suggest that such features may encode relevant information, although this was not supported by the present dataset.

The relatively weaker associations observed between positive affect and pulse-waveform features were also unexpected but appear reasonable in retrospect. Given the well-documented negative correlation between positive and negative affect and evidence that positive affect plays a protective role against adverse psychological states [66], we anticipated that positive affect would exhibit associations with waveform features comparable to those observed for negative affect. Although the observed relationships were directionally consistent with this hypothesis, they were generally weaker and less statistically robust. This pattern remains theoretically consistent with prior psychological research, which emphasizes that positive and negative affect are distinct, partially independent constructs that can co-occur within individuals despite their negative correlation [67]. Together, these findings suggest that both affective dimensions relate to cardiovascular functioning, but that negative affect exerts a more pronounced influence.

Short-term emotional fluctuations appear to exert limited influence on vascular wall properties. However, because arousal is closely linked to HRV [68], which directly affects the timing of cardiac contractions, it is reasonable that arousal was associated primarily with time-domain waveform features. This interpretation aligns with earlier work in affective computing by Healey and Picard [69], who demonstrated that physiological signals reliably capture arousal but are less effective at distinguishing affective valence. Although the present findings are supported by emerging evidence, further studies using higher-quality data and more controlled experimental designs are needed to confirm and extend these results.

Despite the theoretical and practical innovations of this study, 3 limitations warrant cautious interpretation of the results.

First, although the machine learning models achieved significant predictive performance, the level of agreement between the proposed method and the target psychological measures remained limited. Similarly, comparisons between waveform features derived from the smartphone-based photoplethysmography method and those obtained from a reference oximeter demonstrated statistically significant associations, but the magnitude of agreement was insufficient for precise interchangeability. Taken together, these findings indicate that, while pulse waveform features are systematically influenced by mental health status, the current models primarily capture correlational relationships rather than yielding predictions with practical accuracy.

Second, of the 766 waveform samples collected, 100 (13.1%) were classified as corrupted and excluded, a substantial loss rate that may undermine the study’s validity. In uncontrolled, real-world settings, slight finger movement or fluctuations in ambient illumination can distort the photoplethysmography signals [70]. Future research can therefore use enhanced image capture protocols [71], integrate advanced signal processing and noise reduction algorithms [72], and implement quality filters to discard low-fidelity recordings [73]. Moreover, the limited and variable frame rates of smartphone cameras (typically 10-60 Hz) constrain temporal resolution compared with medical-grade devices sampling up to 1000 Hz [74]. Because smartphone camera frame rates are not fixed, this instability further exacerbates data-collection challenges [75]. The effects of reduced and unstable sampling rates on photoplethysmography signal integrity remain poorly understood and warrant further investigation [76].

Third, the sample consisted primarily of young adults with similar educational backgrounds, which may limit the generalizability of the findings. Because vascular changes associated with psychological factors often develop over time, associations between mental health and pulse-waveform morphology may be more pronounced in older adults or individuals with chronic conditions. Future studies should therefore include participants across broader age ranges and with more diverse health profiles. Additionally, the slight gender imbalance in the sample may have introduced bias in sex-specific waveform effects. Other potential confounders, such as sleep quality, smoking, diet, and medication use, were not controlled for and should be addressed in future research to reduce unexplained variance.

Pulse waveform analysis has traditionally been applied to the assessment of physical health, whereas its extension to mental health, via links between psychological states and vascular stiffness, remains relatively novel. Conventional waveform acquisition methods are often costly and inaccessible, which has limited their adoption in psychological research. This study uses smartphone-based photoplethysmography to analyze digital arterial waveforms, thereby providing a noninvasive and accessible approach for psychological assessment. The findings demonstrate significant associations between psychological variables and pulse-waveform features, particularly within the time- and curvature-domains, and thus contribute to the advancement of psychophysiological research grounded in hemodynamic principles. Nevertheless, technical limitations (eg, signal variability and low, unstable sampling rates) and sample-related constraints (eg, demographic homogeneity and uncontrolled confounders) should be addressed in future studies.