The American Psychiatric Association defines depression as a “common and serious medical illness that negatively affects how you feel, the way you think, and how you act” [1]. It comprises symptoms such as low mood, guilt, suicidal ideation, and cognitive decline [1,2]. According to The Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2019, depression is one of the most disabling mental health disorders [3], and it poses a significant economic and medical burden. A study by Greenberg et al [4] calculated an increase in economic cost related to depression of 37.9% from US $236.6 billion to US $326.2 billion in 2020. These costs comprised direct, suicide-related, and workplace costs [4]. There also has been a global increase in the prevalence of depression. The percentage of adults in the United States with major depressive disorder increased by 12.9%, from 15.5 to 17.5 million, between 2010 and 2018 [4]. To further add to these increasing numbers worldwide, the COVID-19 pandemic has led to a substantial increase in mental health conditions, including depression [5], which has been aggravated by the uncertainty associated with the disease, isolation, and overall decreased social interaction [6,7]. Given this rise in depression rates and the immense costs associated with it, adequate diagnosis and timely intervention have become a pressing and urgent need [8].

Depression, as most other mental illnesses, is diagnosed via the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) [9], or the International Classification of Diseases, 11th Revision [10]. However, there is growing skepticism regarding their validity [11,12]. In a groundbreaking research study by Newson et al [11] in 2021, they were able to quantify the degree of heterogeneity within and across the DSM-5 symptom profile in that the DSM-5 criteria “fails to diagnose patients by symptom profile any better than random assignment.” This strongly supports Zimmerman et al [13], who found that there are 227 different ways to diagnose depression. The problem is further exacerbated by heterogeneity among scales used for depression screening and diagnosis [14,15], illustrated by a cross-sectional study that found that, in a small sample of 309 patients, there was a misdiagnosis in 55% of these cases [16]. In addition, there are no approved biomarkers as part of the diagnostic criteria for depression [17]. Compounding factors that contribute to the hurdles associated with adequate screening and monitoring of depression are lack of primary care physicians, low recognition of depression in primary care [18], delayed response to treatment [19], 12-week waiting period in the absence of a response, other comorbidities, and patient fear of stigma attached to depression [20]. There are several instruments for detecting depression in primary care [21], one of which is the Patient Health Questionnaire-9 (PHQ-9), which has been adopted as the gold standard for detecting depression and grading its severity [22].

To overcome these many challenges associated with traditional methods for the detection, management, and monitoring of depression, smartphone-based interventions have advanced as an available and alternate option. Middleweerd et al [23] found that the use of digital tools for physical health monitoring, such as fitness-based smartphone apps, was becoming increasingly popular. The use of digital tools for the management of mental health conditions became a key resource as the demand for mental health support exceeded the supply when the COVID-19 pandemic led to a rise in depressive disorders worldwide [6]. In addition to telehealth and remote therapy, a solution that emerged was digital health assessment using smartphones and their sensors [24], also known as digital phenotyping. Torous et al [25] define the term as “moment-by-moment quantification of the individual-level human phenotype in-situ using data from smartphones and other personal digital devices.” The use of passive sensors in the mental health industry has the potential to detect real-time changes in psychological factors, and this can be used to increase access to care [26], reduce stigma [27] improve diagnosis [28], and enable remote monitoring [29] as has been established by previous and ongoing research.

There is a growing body of research on passive data sensing and its use in modeling human behavior [30]. Previous work has shown that monitoring these digital biomarkers using machine learning models to assess passive smartphone data can aid in the screening, treatment, and remote monitoring of mental health disorders. In a study by Wang et al [31], the app Student Life was used to show the correlation between depression and accelerometer- and screen use–based biomarkers. In another study, Saeb et al [32] found significant correlations between depression and passive data such as phone use and GPS in a sample of 40 participants. Asare et al [33] found that age group and gender as predictors led to improved machine learning performance. Their study concluded that behavioral markers indicative of depression can be unobtrusively identified using smartphone sensor data [33]. Taking a machine learning approach, a study found that the predictive power of mobile device use patterns was significant to continuously screen for depressive symptoms or monitor ongoing treatments [34]. In line with this study, another study used the Remote Monitoring Application in Psychiatry to explore the validity of smartphone-based assessments for self-reporting mood symptoms and found high compatibility with nonsmartphone-based assessments [35], thus proving such tools to be helpful for clinicians and research.

This study focused on South Korea as it has consistently reported a low number of depression cases despite high suicide rates [36] and the dramatic worldwide increase in depression. In 2005, a study reported that the annual prevalence rate of depression in South Korea was 1.7%, whereas rates of depression were reported to be much higher in that 25.3% scored positive for depression in a nationwide sample study [37]. Researchers have debated various reasons for the low prevalence rates, including the difference in cutoff scores in South Korea versus other countries and the associated stigma [37] attached to mental health conditions, but have not yet reached a conclusion as to the cause. Digital phenotyping and passive monitoring can provide a timely opportunity to target issues such as low access to mental health diagnoses, stigma, and associated health consequences in South Korea and similar countries.

Smartphone sensors and passive data, when coupled with relevant statistical and machine learning models, provide an avenue to capture behavioral changes associated with mental health disorders in naturalistic settings [30,38]. Much of the previous work in this field has used sensors that are invasive and privacy-related such as GPS, call logs, SMS text message logs, and keyboards. This study demonstrates a novel mental behavioral profiling metric termed Mental Health Similarity Score (MHSS), derived from analyzing passively monitored nonintrusive and nonidentifiable smartphone use data, to identify and track depressive behavior.

We collected active and passive data in a longitudinal observational study using the Behavidence (Behavidence, Inc) mobile app, derived from a cohort of anonymous participants in South Korea. Participants were invited to take part in this study through social media advertisements and campaigns, which is an effective tool for recruitment in research studies [39]. The advertisement used a research code that the interested individual could use to enter the study by downloading the app (Figure 1). The data set was collected from 558 participants between November 2021 and December 2021. All the participants were Android-based smartphone users.

The inclusion criteria were as follows: (1) they must be aged >18 years and (2) they must have an Android device.

Depression severity was assessed using a patient-reported outcome measurement questionnaire. When the participants registered with a research code and enrolled in the study, they answered the PHQ-9 [40] in the app. The PHQ-9 scale is the gold standard for detecting and measuring the severity of depression worldwide [41]. It has been validated for use in community-based and general population settings and has sound psychometric properties [42]. Although the depression assessments were self-reported, the PHQ-9 has been clinically validated for the assessment of depression severity owing to its high internal reliability (Cronbach α=.89) and has been used in multiple studies as a self-reported questionnaire [22]. The PHQ-9 measures depression severity over the preceding 2 weeks. Each item of the PHQ-9 is scored on a scale of 0 (not at all) to 3 (nearly every day). The total PHQ-9 score ranges from 0 to 27, with a score of ≥10 indicating a major depressive disorder [22]. A score of <5 indicates no depression, 5 to 9 indicates mild depression, 10 to 14 indicates moderate depression, 15 to 19 indicates moderately severe depression, and 20 to 27 indicates severe depression [40]. The PHQ-9 has also been established to have good psychometric properties in the South Korean population, which is the focus of this study [43].

Behavidence [44] is a mental health screening app that passively collects personal device use data with zero respondent burden and no use of identifiable information. The app works as an always-on solution and can be downloaded from the Google Play store. Individuals can register or log in to the app with no supervision, and any required onboarding information was easily made available to the study participants remotely. In addition, demographic user profile information questions (gender, age bracket, and existing mental health indication) were answered within the app.

The Behavidence Research App was developed for smartphones running Android version 5 or higher and requires connectivity to send data to the back end for analysis and receive data analysis outcomes. It does not require connectivity to collect the data. For an app to run as a background process, it must obtain the Battery Optimization and Usage Data Access permissions from the user. These permissions are obtained during the onboarding process.

The app uses the principle of digital phenotyping to track user behavior. It displays an MHSS developed from phone use metrics such as time spent on various apps on a daily and weekly basis. The MHSS displays how similar the user’s digital behavior is to the digital behavior of someone who has been diagnosed with depression. This similarity score is a range from 0% to 100% (Figure 1). The MHSS is generated every 24 hours. The app also shows the user their weekly history of similarity scores. In this study, gyroscope readings from each participant’s smartphone device were collected in addition as sensor features measuring the direction and speed at which the phone was spinning around its axis.

The app and back end use strict data privacy and security protocols. The solution is compliant with the Health Insurance Portability and Accountability Act and the General Data Protection Regulation.

The data set used for model training only included participants aged ≥18 years with at least 24 hours of complete passive, nonsensor, personal device use data. The final data set contained 399 participants with an average of 10 (SD 25.21) days of mobile data. The addition of the use of gyroscope readings, measuring angular velocity as phone sensor features, was tested to improve the accuracy of the model. The purpose of this is to test whether sensor features provide additional insights than using only nonsensor features. The data set used specifically for the gyroscope model training had a reduced individual participant number of 193, in which at least 24 hours of raw device use with the additional sensor readings was available. This reduced number of participants was due to the inability to collect gyroscope readings from specific Android phones.

The raw data set collected contains daily behavioral patterns that were data-cleaned and transformed to reach independent features such as opening and closing apps with the start and end times in Coordinated Universal Time in milliseconds. These data were preprocessed by converting the time stamps to local dates and times according to the user’s time zone. Digital biomarkers, used as machine learning features in this study, were calculated per user, taking daily behavioral patterns on a 24-hour basis starting from midnight every day. The 3 main types of nonsensor features were average time on the phone per day, frequency of events per day, and app category use per day. The mapping of various apps into specific categories can be found in Multimedia Appendix 1. In addition, gyroscope data were collected and processed daily to generate sensor features such as mean activity, average gap activity, and total activity. In the end, a total of 37 features were extracted and merged on a per-user, per-day basis. Explanations of each feature can be found in Multimedia Appendix 2.

The participants downloaded the Behavidence app and answered a simple demographic questionnaire along with the informed consent form. They then completed the PHQ-9. The questionnaire was answered only as a 1-time data point and, thereafter, they were free to use the app on their own. The app generated a daily MHSS. The app added no further respondent burden and, therefore, the participants were able to check the score whenever they felt the need to or not at all.

Consent was voluntarily given on the participants’ smartphones once they were informed of the purpose of the study. The data set does not contain personally identifiable or any personal health information. The advertisement, informed consent, and study protocol were approved by the independent Western Institutional Review Board-Copernicus Group, Institutional Review Board (approval number: 20216225).

Self-reported demographic data from the 558 participants (Table 1) show that, of these, 286 (51.3%) identified as women, 254 (45.5%) identified as men, and 18 (3.2%) identified as nonbinary or preferred not to disclose their gender. Regarding the participants’ age distribution, of the 558 participants, 474 (84.9%) were aged between 18 and 25 years, 29 (5.2%) were aged between 26 and 35 years, 42 (7.5%) were aged between 36 and 55 years, 10 (1.8%) were aged between 56 and 64 years, and 3 (0.5%) were aged ≥65 years. The PHQ-9 questionnaire was administered to users in both English and Korean, with most of the participants belonging to the Korean-speaking population (487/558, 87.3%).

Table 2 presents the distribution of the PHQ-9 scores of the 558 participants. The PHQ-9 was collected at the start of recruitment at a single time point during this study. The distribution of the PHQ-9 scores was as follows: 11.3% (63/558) were in the none category (ie, they were not depressed) with PHQ-9 scores <5, whereas 88.7% (495/558) showed signs of depression by scoring between mild and severe. The mean PHQ-9 score was 12.5 (SD 6.29).

There is an imbalance in the gender distribution when looking into each severity group of depression. For the none and mild cohorts, men represented the majority, whereas, in the moderate, moderately severe, and severe cohorts, there was a female majority, as shown in Table 3.

Out of the 558 participants, 499 (89%) answered “no previous diagnosis” in the demographic questions collected at onboarding. Moreover, 65% (323/499) of the participants who reported that they had no previous diagnosis of any kind obtained a PHQ-9 score of at least moderate (≥10) to severe depression, as shown in Table 4.

A total of 37 features were computed from the raw passive smartphone data. Of the 37 features, 29 (78%) showed statistical significance in the 1-tailed t test results between the none and severe cohorts. Overall, 8 of the significant nonsensor and gyroscope (sensor) features are displayed in Table 5. The remaining list can be found in Multimedia Appendix 3. Effect size analysis showed that the most important features had moderate to high effect sizes when comparing the none and severe category populations [45].

Among the classification algorithms, random forest proved to have the highest predictive accuracy (87%). Extreme gradient boosting followed with an accuracy of 86%, whereas the support vector machine classifier had the lowest accuracy (44%), as shown in Table 6.

The top-performing algorithm, the random forest classifier trained on the PHQ-9 binary nonsensor model (none vs severe on the depression rating scale) on 34 of the nonsensor features mentioned in Table 5, achieved a precision of 85% to 89%, recall of 85% to 89%, F₁ of 87%, and overall accuracy of 87%, as shown in Table 7.

The feature importance plot based on Gini Impurity measurement analysis indicates the top passive digital features indicative of differentiating between none and severe cohort participants (Figure 2). The top 5 features are mean session time on the phone within a 24-hour period, average session time in the social interaction apps (app category 1), average session time within a 24-hour period in the miscellaneous and additional passive recreational apps (app category 11), number of times active messaging and communication apps (app category 3) were opened within a 24-hour period, and average time spent on unofficial or unregulated apps (app category 0). The top 5 features had Gini Impurity values ranging from 0.6 to 0.1.

The top 5 features from the feature importance list achieved statistical significance with P<.001, as shown in Table 8. Average overall session time and average time spent on social interaction apps, miscellaneous and additional passive recreational apps, and unregulated apps had greater mean values for the participants who scored as severe compared with the participants who scored as none on the PHQ-9. None participants opened active messaging and communication apps 110 times on average (SD 7.02), whereas severe participants opened this app category 74 times on average (71.05).

In addition, the model was tested to see whether it could accurately predict participants who had reported a previous diagnosis of depression. The model achieved an accuracy of 80% in detecting depression but only 26% in detecting the none group.

Additional validation of whether the PHQ-9 binary nonsensor-based model (none vs severe) could accurately predict the progression of depression was performed by calculating the percentage of participants in each group who had a majority of days with high MHSSs (>50%), with the MHSS as the model’s prediction of class probabilities. As shown in Table 9, the majority increases as severity increases, indicating that participants with severe depression had a majority of days with high MHSSs, supporting the model’s prediction ability.

When the gyroscope features were added as additional markers, the overall accuracy dropped to 76% with a precision of 74% to 78%, recall of 67% to 83%, and F₁ of 72% to 78%. This model was also tested on the self-proclaimed diagnosis cohort and achieved 27% accuracy in detecting depression and 0% accuracy in detecting the none group. When age and gender were added to see if demographics played a role in classifying none (not depressed) versus severe (depressed) cohort participants, the overall accuracy decreased slightly from 87% to 84%, precision increased from 82% to 87%, recall increased from 83% to 86%, and F₁ increased from 84% to 85%.

An additional random forest classifier trained on the PHQ-9 3-class model—none (PHQ-9 <5), moderate (10≤PHQ-9<15), and severe (PHQ-9≥20) depression on the 34 nonsensor features—achieved a precision of 74% to 86%, recall of 76% to 83%, F₁ of 75% to 84%, and overall accuracy of 78%, as shown in Table 7.

A significant positive Pearson correlation was found among PHQ-9 questions 2, 6, and 9 within the severe category users and the mental health behavioral profiling metric (r=0.73), as shown in Table 10. When a gyroscope sensor was added as a feature, the Pearson correlation among questions 2, 6, and 9 dropped from 0.73 to 0.46.

A binary model trained on questions 2, 6, and 9 was constructed to complement the PHQ-9 binary nonsensor model. The participants who scored 0 on all 3 questions were considered as the none class, whereas the participants who scored 3 on every question were considered as the depression symptoms class. The PHQ-9 questions model achieved an overall accuracy of 78% with a precision of 76% to 80%, recall of 75% to 81%, and F₁ score of 78% to 79%, as shown in Table 7. Figure 3 shows the feature importance plots for this prediction model. Top features include (1) number of times active messaging and communication apps (app category 3) were opened within the 24-hour period, (2) number of times active messaging and communication apps were opened or longer than 1 SD from the mean session time within the 24-hour period (app 3 upper), (3) number of passive information consumption apps (app category 2) opened within the 24-hour period, (4) average time spent on general utilities apps (app category 6), and (5) average time spent on passive information consumption apps (app category 2).

Table 11 displays the mean values of the top 5 features of the random forest model for PHQ-9 questions 2, 6, and 9. The number of times the participants opened active messaging and communication apps that had greater session lengths than the average was calculated for both the none and severe participants and proved to be both statistically significant (P<.001) and a top feature in the questions model. The none participants opened this app category 6.47 times on average compared with the severe participants, who opened it 3.25 times. In addition, none participants opened passive information consumption apps 2.13 times on average compared with severe participants, who opened them 0.46 times on average. Finally, severe participants had general utilities apps opened for longer (0.57 minutes) on average than the none participants (0.40 minutes), but none participants had passive information and consumption apps opened for longer (0.29 minutes) on average than severe participants (0.18 minutes).

The study objective was to demonstrate a novel machine learning mental behavioral profiling metric termed MHSS, derived from analyzing passively monitored and nonintrusive smartphone use data, to identify and track depressive behavior. This objective was met as the MHSS models reached an overall accuracy of 87%. In this study, an average of 10 days of smartphone data were used in addition to PHQ-9 results from 399 participants to demonstrate the ability to detect digital behavioral markers quantified from the participants’ smartphones to detect depression severity. We further focused on using these digital behavioral markers to develop predictive models to classify none (not depressed) and severe (depressed) symptom severity scores.

A mental behavioral profiling metric termed MHSS, developed from digital markers extracted from the participants’ smartphone data, was able to predict the participants’ depression state (ie, none or severe) with high predictive performance using machine learning models.

Demographic analysis found a higher number of women in the severe group compared with the none group, which is in line with previous literature on the prevalence of depression in South Korea [46]. Studies conducted in South Korea have shown that, in a sample, women were more depressed than men across all age groups [47]. Both regression (random forest regression and multivariate adaptive regression splines) and classification (random forest classification, extreme gradient boosting, and support vector machines) machine learning models were tested to evaluate the highest predictive accuracy between none and severe depression. The PHQ-9 binary nonsensor model (none vs severe) achieved the highest accuracy using a random forest classification algorithm with the following metrics: precision of 85% to 89%, recall of 85% to 89%, F₁ of 87%, and overall accuracy of 87%. The PHQ-9 3-class (none vs mild vs severe) model achieved the following metrics: precision of 74% to 86%, recall of 76% to 83%, F₁ of 75% to 84%, and overall accuracy of 78%. The effect size of the nonsensor features was moderate, and the effect size of the sensor features was low. The PHQ-9 gyroscope sensor model achieved the following metrics: precision of 74% to 78%, recall of 67% to 83%, F₁ of 72% to 78%, and overall accuracy of 76%. Although the results of this study have similar accuracies to previous studies [24,48], these models indicate that invasive features such as GPS tracking and audio information are not necessarily required to detect behaviors in individuals with depression.

The feature importance list was extracted based on the Gini Impurity measurement. In the PHQ-9 binary nonsensor model, results found that mean session time was the most important feature in predicting severe depression using nonintrusive passive sensors. Mean session time was higher in participants with severe depression compared with none. This result is in line with previous findings that people with higher phone use have a positive correlation with self-reported depression [49]. Another study found that high mobile phone use was associated with symptoms of depression in men and women at 1-year follow-up compared with people with low phone use [50]. Going down the feature importance plot, it was observed that app category 1 (social interaction apps) had higher use in participants with severe depression compared with participants with no depression. This aligns with studies that have shown that social media use is higher in people with depression [51,52] and that limited use can lead to a decline in self-reported feelings of depression [53]. In the PHQ-9 questions model, it was interesting to find that mean session time on active messaging and communication apps was lower in participants with severe depression (ie, their use of this app category was low compared with the none group). This finding is in line with previous findings that web-based communication is reported to be low in people with depression [54].

A per-item correlation was performed, and a significant positive Pearson correlation was found between PHQ-9 questions 2, 6, and 9 within the severe category users and the mental health behavioral profiling metric, that is, the MHSS (r=0.73). Users who had higher scores in the 3 questions also had higher MHSSs (>50%). Previous research has shown that items 2, 6, and 9 comprise the affective-cognitive component of the PHQ-9 scale [55]. The highest-correlated PHQ-9 items—2, 6, and 9—were the questions that indicated affective symptomatology; therefore, a separate PHQ-9 questions model was created, achieving a precision of 76% to 80%, recall of 75% to 81%, F₁ of 78% to 89%, and overall accuracy of 78%. This 2D PHQ-9 questionnaire (the other being the somatic component) approach has been used in previous studies and shown to have sound psychometric reliability and validity [55-57]. Thus, the results of this study add to the literature as, to the best of our knowledge, no previous studies have explored the 2-factorial approach of the questionnaire and used it to create digital behavioral markers.

We chose South Korea as the study site as there are low levels of reported depression despite the high number of cases of suicide. The study found that, of the 469 individuals who reported having no diagnosis as their current status in their demographics questionnaire, 307 (65.5%) scored as moderate to severe depression (PHQ-9 score ≥10). This result fits previous literature that states that the population in South Korea is often less likely to seek treatment and diagnosis for depression because of low awareness and stigma [58]. Our results also complement our intention to study a South Korean sample as a previous study on the prevalence rates of depression in South Korea found that, despite the high suicide rate in the country [59], the prevalence of depression has been reported to be much lower compared with other countries [60]. This can be mainly due to 2 factors: the access rate to services for depression has been reported to be low and the mental health treatment gap for major depression is 56.3% [61]. It is also interesting that studies have shown that the prevalence of depression rates is lower in Asian countries, such as South Korea, when compared with Western countries [62,63] owing to the stigma surrounding psychiatric illnesses [64]. This result demonstrates the feasibility of a daily mental health profiling metric using smartphone-based passive data to monitor symptoms, administer tests at home, and schedule interventions, which will help overcome the limitations hindering traditional methods of assessment such as stigma [64] and hesitation toward accessing mental health services because of low education levels [58].

This study also evaluated the use of another passive sensor (ie, gyroscope) to improve the accuracy of our models. This study found that, when a gyroscope sensor was added as a feature, the Pearson correlation among questions 2, 6, and 9 decreased from 0.73 to 0.46. Mean activity (P<.001) and average gap activity (P<.001) features from the gyroscope sensors showed statistically significant differences between none and severe individuals. Therefore, although gyroscope sensor data show some distinction between the 2 cohorts when including them as an additional feature, the gyroscope as a sensor alone does not add predictive power.

Previous researchers have established a relationship between depression and digital phenotyping using identifiable passive information such as GPS and HealthKit information [8,48] and have a high respondent burden, such as daily mood surveys and multiple assessments [24,48]. Our study adds an approach in which we show that high-accuracy models to detect depression can be achieved using nonintrusive data such as average time on the phone per day, frequency of events per day, and app category use per day, with only 1 baseline assessment and no further respondent burden. The behavioral profiling metric, called the MHSS, is easy to understand by the user and, therefore, is easily incorporated into various clinical and therapeutic scenarios. The findings about various app category uses provide a dive into behavior patterns of depressed and not depressed groups, which can be useful for risk profiling. This study also further complements the idea shared by Onnela [65] in his research that private passive data collected from smartphones present a big challenge and should be anonymized. The app used in this study collects only nonintrusive, passive data, and the data are encrypted from the time they are collected and then re-encrypted when they are stored in the servers, thereby guaranteeing an accurate and safe MHSS. To further address any concerns about security, the app provides the ability to obtain daily MHSSs as a completely anonymous user, ensuring zero traceability.

In the everyday clinical scenario, the MHSS can help with remote monitoring of symptoms as well as treatment or intervention efficacy. It is a simple, affordable, and accessible form of technology that is easily scalable. This proves especially useful in low- and middle-income countries, where there are multiple barriers to mental health care access. In our study, we found that the MHSS can detect individual patterns of behavior as well as population-based trends. However, further research is required to establish its use on an epidemiological level.

The strength of the study can be found in using nonintrusive, passive behavioral data to generate digital phenotypes for depression and, in the future, for more mental health disorders. In addition, web-based recruitment was used, which eased the onboarding process and allowed the users to participate in the study at their own comfort. This study design is easy to replicate for other digital phenotyping indications where it is possible to administer web-based self-report questionnaires and generate results.

The first limitation of the study was that the PHQ-9 was administered only once; thus, the depression symptom status was only collected at baseline. Future studies should aim to assess the symptoms at 2 time points and observe the changes in questionnaire scores alongside the changes in digital behavior. Another limitation was that our data were heavily inclined toward the 18 to 25-year age group, with 84.9% (474/558) of the participants belonging to it. Our study did not have clinical diagnosis of depression as an inclusion criterion, only a self-reported clinical diagnosis and the self-reported PHQ-9 scale. Using a more diverse age group in a more proportionate number could provide a better overview of how digital behavior symptom severity could change with age as a factor. Although the PHQ-9 as a patient-reported outcome measure is the gold standard method for diagnosing depression and is used worldwide to screen for depression, the results of this study will be further consolidated when tested in a clinically diagnosed population. This study was available only for Android users; therefore, further studies should look at incorporating the iOS operating system. Furthermore, future studies can include other locations and questionnaires.

Nonidentifiable passive smartphone data prove to be a suitable tool to assist with the remote screening and monitoring of depression. The strong privacy metrics and low respondent burden pave the way for further exploration in not only screening and even triaging patients but also measuring therapeutic outcomes through the MHSS as a metric. Finally, the aggregated measurement of a group as a health metric could further support larger epidemiological studies.