Depression and anxiety are common mental health issues in children and adolescents. The prevalence of depression and anxiety among children has consistently risen, with 5.6 million children diagnosed with anxiety and 2.4 million with depression in the United States in 2020. These figures represent increases of 27% and 29%, respectively, compared to 2016 (1). The situation in South Korea mirrors this trend. According to the National Health Insurance Service (NHIS), the number of children and adolescents aged 18 and under receiving treatment for depression or anxiety disorders rose from 54,000 in 2019 to 63,000 in 2021, reaching approximately a total of 210,000 from 2019 to 2022 (2). Depression and anxiety often coexist in children (3). Some researchers argue that anxiety and depression in preadolescents are part of a larger structure of negative emotions, rather than being distinct emotional states (4). In adults, the cumulative effect of daily stress plays a significant role in predicting the development of depressive and anxiety symptoms (5, 6). This phenomenon is also observed in children (7).

Beck’s Cognitive Theory of Depression suggests that some individuals possess negative cognitive schemas, and when activated by factors such as stress, these schemas contribute to the emergence of negative perceptions and interpretations, potentially leading to depressive symptoms (8). Research on depression in children and adolescents also supports the Cognitive Vulnerability Theory of Depression (9). Stressors related to school or home environments are associated with higher levels of depressive symptoms and aggression in children (10). Children experiencing depression encounter a higher number of stressful life events, such as an increase in arguments with parents, compared to their non-depressed counterparts (11).

Furthermore, stress is associated with anxiety. Studies examining the relationship between non-traumatic life stress events (e.g., parental divorce, a grandparent’s death, or relocation) and anxiety disorders or symptoms have shown that children diagnosed with anxiety disorders have experienced stressors more frequently than non-anxious children (12). Similarly, families experiencing an increased number of stressors report higher levels of anxiety symptoms in both parents and children (13).

The early detection and evaluation of mental health issues in children are crucial for implementing prevention and intervention strategies (14). Mental health concerns in children under the age of eight years are at risk of being overlooked or neglected, as these young individuals may struggle to recognize their negative emotions due to limited emotional awareness (15). This is particularly true for vulnerable children, where issues related to self-reporting can impact the clinical assessment (16). Existing screening tools often consist of extensive parent-reported surveys, which inherently result in the under-reporting of children’s symptoms. This occurs because discerning a child’s thoughts and emotions can be challenging, even for adults who are intimately familiar with the child’s behavior (17, 18). These challenges underscore the need for an objective, accurate, and easily applicable mental health screening tool to detect anxiety and depression in young children (19).

Reliable measurements of mental health are crucial for patient care and clinical research. As such, the use of digital technologies for assessing mental health is increasing, aiming to overcome the limitations of conventional evaluations by objectively measuring mental health biomarkers (20). Digital biomarkers refer to objective, measurable physiological and behavioral indicators collected through portable, wearable, implantable, or ingestible digital devices (21). Across various medical apparatus, hardware, or software strata, digital biomarkers can capture physiological and behavioral cues from individuals (22). The use of these digital biomarkers enables the continuous remote monitoring and evaluation of reliable clinical data, thereby enhancing the accuracy of diagnoses and treatments (23, 24). Moreover, these devices are less susceptible to human bias and can reduce patient stress (25). In fact, a growing body of research is exploring tools that use digital biomarkers for assessing various aspects of depression, anxiety, and stress (26, 27).

In clinical or laboratory settings, a wide array of digital biomarkers is available for conducting mental health screenings. However, our study focuses on developing a user-friendly, non-clinical screening tool designed for home use, especially for children. To achieve this goal, we are exploring digital biomarkers that can be developed through software and implemented on universally accessible devices.

Our primary objective was to confirm the feasibility of using various digital biomarkers in children, specifically for monitoring childhood depression, anxiety, and stress. Based on previous research, the following hypotheses were formulated:

H1. HRV, eye-tracking, and voice features will account for the child’s depression score.

H2. HRV, eye-tracking, and voice features will account for the child’s anxiety score.

H3. HRV, eye-tracking, and voice features will account for the child’s stress score.

We initially examined previous studies that assessed depression, anxiety, and stress using HRV, eye-tracking, and voice data. Based on these references, we selected variables with a significance level of < 0.005 to avoid multicollinearity. The variables commonly used, as cited in previous academic papers, are summarized in Table 1 .

The application we developed performs a comprehensive evaluation by simultaneously incorporating three standard tests: HRV, eye-tracking, and voice ( Figure 1 ).

We conducted the study procedures in the clinical laboratory of Chung-Ang University Medical School. Participants were recruited voluntarily after they and their parents expressed interest through online postings and promotional materials, leading to the enrollment of 54 children. Given the focus on elementary school students, the age range of participants was seven to 11 years. There were no dropouts, and written consent was obtained from all participants. Upon their arrival at the laboratory, parents and/or guardians and children were guided through the consent procedures by researchers from Chung-Ang University. After giving consent, participants filled out demographic questionnaires and trait inventories, underwent a dominant hand test, and then used the digital assessment tool. Upon completion, each participant was informed that the test had concluded and they were reunited with their parents. We adhered to a heuristic of 50 people per condition, based on sample sizes from previous literature (59). Ethical approval for the study was obtained from the Institutional Review Board at Chung-Ang University.

The biomarker collection module, developed as a digital application compatible with the Android operating system, was designed for future real-world application. HRV and eye-tracking data collection was facilitated by Software Development Kits (SDKs), while voice data were captured using the tablet’s built-in recording function. All collected data were initially stored locally on the devices and later transmitted to our server upon completion of the process. For biomarker extraction, we employed Samsung tablet PCs, which facilitated efficient and accurate data collection in a controlled environment.

Additionally, participants were given a 2-minute break before each task in our assessment tool to ensure a stable state.

Data underwent partial correlation analyses, accounting for age and gender influences, followed by hierarchical regression analysis to determine the impact of digital biomarker features on mental health symptoms. In these analyses, age and gender were accounted for as covariates in the first block. Digital biomarker features were then introduced in a stepwise multiple regression in the second block. This covariate analysis was conducted to address the broad age range and potential gender influence on digital biomarker data. Accordingly, partial correlation analyses were carried out, considering the effects of age and gender on the observed associations. This analytical approach was used in a study investigating the relationship between multimodal digital biomarker features and major depressive disorder scores (65).

Furthermore, the sample data were categorized by the children’s age into two groups: elementary lower grades (1st, 2nd, 3rd grades) and elementary upper grades (4th, 5th, 6th grades). This categorization enabled the exploration of potential age-related influences on digital biomarker data. Given the hypothesis that the broad age range might impact digital biomarker outcomes, participants were divided into two groups for subsequent correlation analysis. The statistical significance level was established at α = 0.05.

The final sample consisted of 43 participants. Six participants were excluded due to missing HRV data. This exclusion was necessary to ensure a complete dataset for analyzing the potential of using digital biomarkers in a multimodal context. Outliers were identified and removed when five or more data points deviated more than three standard deviations (SDs) from the mean of each digital biomarker variable. Ultimately, the sample included 25 males and 18 females, with an average age of 9.05 years (SD = 1.308) ( Table 2 ).

To explore the relationships among children’s depressive symptoms, anxiety symptoms, and stress with digital biomarker variables, we conducted partial correlation analyses. This method enabled the identification of pure correlations by controlling for age and gender effects. No significant correlations were found between HRV variables and age or gender. However, significant associations were observed between eye-tracking variables and gender, with females spending more time viewing contemptuous and sad expressions. Additionally, we have detailed the interrelations among the digital biomarkers in Table 3 .

The K-CDI-2 and K-RCMAS-2 scores showed correlations with HRV variables. The K-CDI-2 was positively correlated with LF (r = 0.409, p = 0.008) ( Table 4 ) and negatively associated with eye-tracking variables (r = -0.374, p = 0.016) ( Table 4 ). Conversely, the K-RCMAS-2 displayed a negative correlation with HR (r = -0.310, p = 0.048), HF (r = -0.340, p = 0.030), and a positive relationship with LF/HF (r = 0.409, p = 0.008) ( Table 5 ). For the stress scale, pNN50 showed a negative correlation (r = -0.382, p = 0.014). Acoustic variables (voice) were not correlated with the clinical scales ( Table 6 ).

Further analyses were conducted to examine the impact of participants’ varied ages. Anticipating that the broad age range could influence digital biomarker outcomes, participants were stratified into two groups for subsequent correlation analysis: lower grades (grades 1 to 3) and upper grades (grades 4 to 6) of elementary school, with gender controlled in our analyses ( Tables 7 , 8 ).

The HF value, not significantly associated with depression scores, showed a significant positive correlation in the lower-grade group (r = 0.503, p = 0.028). Similarly, the LF value, not significantly correlated with anxiety scores, demonstrated a significant positive relationship in the upper-grade group (r = 0.489, p = 0.016). Conversely, jitter was positively associated (r = 0.464, p = 0.046) with stress in the low-grade group. These results reveal relationships for HRV and voice features that were previously not correlated.

We assessed the predictive capabilities of digital biomarker variables for depressive symptoms, controlling for age and gender as covariates. Initially, age and gender were included as covariates, and subsequently, HF, LF, Sad.TF:time(refers to the total duration of gazing at sad facial expressions), pitch, jitter, and shimmer were introduced stepwise. This approach was adopted because these variables have been identified as predictors of depression in prior research. LF and Sad.TF:time emerged as significant predictors for the total depressive symptom score, yielding the following results: F(4, 38) = 4.103, p = 0.007, R² = 0.302, and ΔR² = 0.228 ( Table 9 ).

We performed the same analysis for anxiety symptoms, following a similar procedure. This time, variables associated with anxiety symptoms in the literature, such as HF, LF, LF/HF, HR, pNN50, Contempt.TF:time(refers to the total duration of gazing at contemptuous facial expressions), pitch, jitter, and shimmer, were included in the second block. LF/HF, HR, and pNN50 remained significant predictors for the total anxiety symptom score, with results as follows: F(5, 37) = 5.565, p < 0.001, R² = 0.429, and ΔR² = 0.352 ( Table 10 ).

For stress symptoms, we analyzed variables linked to stress in previous studies (HF, pNN50, Contempt.TF:time, pitch, jitter, and shimmer) in the second block. Here, pNN50 stood out as a significant predictor for the total stress symptom score: F(3, 39) = 2.545, p = 0.03, R² = 0.164, and ΔR² = 0.099 ( Table 11 ).

Our research suggested the potential clinical utility of multimodal digital biomarkers in tracking childhood depression. While the effect size was modest, the total duration spent viewing sad faces was identified as an explanatory variable for child depression, particularly in relation to LF. These outcomes support our first hypothesis and underscore the potential of multimodal biomarkers in monitoring depression among children, marking a novel application of this method in pediatric populations. This discovery paves the way for further exploration of childhood depression using a multimodal approach in future studies.

Moreover, anxiety symptoms were most accurately explained by HRV features. Notably, HRV was one of the variables that explained both depression and stress. However, the variables explaining anxiety solely consisted of HRV features, which exhibited the greatest explanatory power in our findings. We propose that the effectiveness of HRV stems from its ability to capture variations in autonomic nervous system (ANS) activity. Prior studies have consistently demonstrated a significant relationship between ANS activity and depression, anxiety, and stress (66), lending support to our hypothesis.

Finally, we conducted additional analyses based on participants’ ages, dividing them into lower grades (grades 1 to 3) and upper grades (grades 4 to 6) of elementary school. Our findings align with previous research that investigated age-related variations in digital biomarker patterns (67, 68). Importantly, our research identified distinct digital biomarker patterns in children across different age groups, including both before and during adolescence. This finding highlights the presence of age-specific disparities in digital biomarker patterns.

Our study, while producing potentially beneficial results, was subject to several limitations.

First, our small sample was characterized by a broad age range. The diverse developmental stages of the children, corresponding to their ages, may have influenced our results. A sample size calculation based on a large effect size of 0.4 indicated the need for 49 participants, highlighting the insufficiency of our initial sample. However, the effect size derived from our findings was 0.6, suggesting the sample size was adequate. However, to enhance the credibility of our findings, recruiting children within a narrower age range and increasing the sample size are essential for more definitive conclusions. This limitation reflects the exploratory nature of our study and indicates the need for further research.

Second, our study did not include children clinically diagnosed with depression or anxiety, only healthy controls. While many previous studies have focused on diagnosed subjects, our research identified trends using clinical measurements in healthy children and relied on self-reported data rather than evaluations by caregivers or clinicians. Future studies will need to validate our assessment tool with clinically confirmed patient groups.

Third, we could not replicate prior research (68) indicating a significant decrease in LF in depressed children. However, our findings are consistent with those of Wang et al. (69), and the study by Jangpangi et al. (70), where LF was increased in the depressed group, supports our research. Further investigation is required to clarify these varied results.

As a result, our study suggests the feasibility of using multimodal biomarkers for monitoring children’s mental health in non-clinical settings but does not assert that our tool and results are suitable for clinical diagnosis.