Anxiety is one of the leading causes of mental health disability around the world [1]. It includes feelings of excessive worry and negative thoughts, accompanied by physical symptoms such as heart palpitations and increased blood pressure [2]. Anxiety is also associated with a high degree of functional impairment [3] leading to poor quality of life [4] and high health care utilization [5]. Despite being one of the leading causes of mental health disability (1 in 4 people according to the World Mental Health Survey [6]), the detection of generalized anxiety disorder (GAD) is very low in primary care settings [7-9]. These challenges stem from the problems regarding diagnostic processes and inaccuracies [8,10-16] as well as overlapping comorbidities [9,17,18] and physical symptomatology [5,19]. The diagnosis is also vulnerable to the observer’s state of mind [20] and biased self-perception [21] of symptoms. Whether it is the diagnosis of GAD as a singular condition or as a comorbidity, the validity of the diagnostic classifications and instruments in themselves has been rigorously debated. Newson et al [22] highlighted the heterogeneity in DSM-5 classification, where it failed to diagnose a specific disorder from random. Zimmerman et al [23] demonstrated how a physician can diagnose depression and its comorbidities in 227 different ways and Phillips [15] has highlighted the ambiguities in DSM-5 criteria for disorder classification. A recent analysis [10] of eHealth data, patient records, and physician reports in psychiatric cases has highlighted the presence of diagnostic errors in two-thirds of the sample.

With the advancement of technology, researchers have employed multisource data and advanced data analysis techniques to refine and improve mental health diagnosis. One such opportunity to use an upcoming method to improve screening of anxiety is to harness the power of smartphones using the principles of digital phenotyping [24]. Digital phenotyping is a novel computational approach that relies on real-time quantification of human behavior through continuous monitoring of digital biomarkers [25-27]. Mobile and wearable digital devices offer the opportunity to track a multitude of parameters such as mobility (through GPS and accelerometer) [28,29], societal interactions [30] (number of calls, voice tone detection, number of messages sent), digital interactions (access to certain apps), phone usage frequency (screen turned on/off) [27], and health monitoring parameters (heart rate, blood pressure, and oxygen saturation) [31]. However, most digital phenotyping approaches present limited applicability due to the lack of standardized data processing approach for big data exploitation and lack of a specific pattern of unique features for complex mental conditions such as anxiety disorder.

Smartphones hold huge potential in redefining the ability to understand mental health behavior. Sensors embedded in smartphones allow for both passive and continuous data collection, which enhances the possibility of understanding human behavior daily [32-34]. Longitudinal monitoring of passive sensors and phone usage has been linked to tracking mental health behavioral trends [24]. Digital phenotyping of mental health has proven successful in dealing with the challenges associated with a diagnosis such as biases in self-reporting and lack of time in primary care settings, thus paving the way for new and novel methods of screening and monitoring [35].

Most previous studies have focused on using digital phenotyping and passive sensor data to predict social anxiety rather than generalized anxiety [28,29,32,36]. In addition, the passive data used in previous research were intrusive of the users’ privacy and collected identifiable data points such as GPS, audio, message logs, and Bluetooth. Jacobson et al [29] demonstrated that sensor data such as accelerometer, call log, and text message data from smartphones could predict social anxiety symptom severity. Another study found that people with high social anxiety had much lower call and text message logs, and used more health and fitness apps and less camera apps as compared with the low social anxiety group [36]. A clinical review on digital phenotyping and the mental health of college students found that sensors such as accelerometer, Bluetooth, and social information can help in understanding clinical symptomatology [37]. By contrast, Meyerhoff et al [28] found that GPS-based sensor features can be useful in predicting depression severity, but it was not significant in predicting anxiety. Other studies that have researched generalized anxiety have been grouped along with other disorders such as depression and social anxiety. The sensors that have been utilized included location sampled every 5 minutes, call and message log data, duration, and length. Interestingly, these studies also found that there was no significant relationship between GAD and location sensors [28,38]. A more recent study investigated how features extracted from smartphones can be used to predict GAD, social anxiety disorder, and depression. The authors found that their machine learning models and features were able to predict social anxiety disorder and depression severity but not GAD [25]. Such findings have paved the way to explore more ways to map generalized anxiety using nonintrusive and nonidentifiable smartphone data.

In this study a novel mental behavioral profiling metric, derived from smartphone usage, is defined for the identification and tracking of GAD. The accuracy of this metric is evaluated in relation to the standardized anxiety assessment protocol using the 7-item Generalized Anxiety Disorder Scale (GAD-7) questionnaire scoring.

Participants were recruited via an advertisement through social media campaigns on Facebook and Google. Research has shown that this is an effective means of recruitment and provides more generalizability than a clinic-recruited study [39]. Interested participants responded to the advertisement by reading about the study and signing the informed consent form. They then downloaded the “Behavidence Research App” from the Google Play store and filled in a demographic questionnaire, followed by the GAD-7 scale. These data were collected at a single time point only during the onboarding process. The app continued to passively collect nonintrusive data from the smartphone such as screen time and app usage, with no engagement requirement from the user. There was absolutely no private information collected, making this solution completely nonintrusive and secure. Data were collected between October 2021 and January 2022. The participants were informed about the type of nonidentifiable passive data collected in the consent form.

A total of 238 globally distributed users responded to the online advertisement. The inclusion criteria were (1) participants should be over 18 years of age; (2) participants must be able read, speak, and write in English; and (3) participants must have an Android smartphone. Of the enrolled participants, 229 completed the entire on-boarding process. There were no restrictions on gender, ethnicity, or the participant’s location.

The advertisement, informed consent, and the study protocol were approved by the independent Western Institutional Board Copernicus Group (WIRB-CG) institutional review board (Approval Number 20216225).

Self-reported demographic data from the 229 participants (Table 1) show that 85 (37.1%) identified as females, 142 (62%) identified as males, and 2 (0.9%) identified as nonbinary or preferred not to disclose their gender. For the participants’ age distribution, 102 (44.5%) were aged between 18 and 25, 66 (28.8%) between 26 and 35, 56 (24.5%) between 36 and 55, and 5 (2.2%) between 56 and 64. A majority of the participants that completed the questionnaire were of Asian race (104/229, 45.4%), and had education levels between some college diploma and a bachelor’s degree (158/229, 69%). The participants in this study were from different locations around the globe. Most were in Asia (84/229, 36.7%) followed by Africa (76/229, 33.2%). The remaining participants were from America, Europe, and Australia.

Table 2 represents the distribution of the 229 recruits and their GAD-7 scoring. The GAD-7 was completed at the start of recruitment at a single time point during this study, which spanned from October 2021 to January 2022. The distribution of the GAD-7 scores was as follows: 23/229 (10%) were none with GAD-7 scoring less than 5, while 206/229 (89.9%) showed signs of anxiety by scoring between “mild” and “severe.” The mean GAD-7 score was 11.8 (SD 5.7).

As seen in Table 3 16% (14/88) of self-reported healthy “none” group participants scored “none” on the GAD-7, whereas the greatest percentage (29/88, 33%) of participants in this group scored “moderate” anxiety. Table 3 also shows that 52% (13/25) of participants with self-reported anxiety had severe anxiety on the GAD-7. Further, 61% (31/51) of participants with self-reported depression had “severe” anxiety and only 2% (1/51) had no signs of anxiety.

Each GAD-7 item was tested using nonparametric permutation testing with Pearson correlation against MHSS on the day that the GAD-7 was filled (Table 7). The highest correlated items belonged to Items 1, 3, and 7: (1) “Feeling nervous, anxious, or on edge” had a correlation of 0.54 (P<.001), (3) “Worrying too much about different things” had a correlation of 0.59 (P<.001), and (7) “Feeling afraid, as if something awful might happen” had a correlation of 0.55 (P<.001).

Smartphone technology has certainly become a primary platform not only for communication but also to receive, manage, and share multiple kinds of data. Recently, the application of smartphones and their sensing capabilities have demonstrated huge potential in health information acquisition and analysis [25-30,34-38]. Mining smartphone data to represent digital behavior can be used for delivering informed clinical decisions and early risk stratification of mental health disorders. Through this study, we demonstrate the application of digital phenotyping in the identification and remote monitoring of GAD.

A novel mental behavioral profiling metric called MHSS was derived by engineering 34 digital features to serve as a marker for GAD. This was accomplished using smartphone usage data mined in a passive manner without the use of any private information. The smartphone usage data comprised active app usage time and frequency collected through the Behavidence app for an average period of 14 days per user. A single observation that consists of 24 hours of smartphone usage data had a typical size of 30 KB. During the course of the study, the engagement with the Behavidence app (number of times the app was opened per day) had an average of 0.78%, highlighting the benefit of zero respondent burden. Answering the GAD-7 questionnaire was only for the purpose of training the models and testing its performance. Models created in the study explored the ability of the MHSS to predict the GAD-7 outcome at 3 levels of granularity. The regression model explored the conformance of MHSS to GAD-7 on an individual score level (0-21) and achieved an RMSE of 4.508. The multiclass classification model encoded 4 levels of anxiety severity with an overall accuracy of 50%, whereas the binary classification model distinguished individuals with severe anxiety from the ones without any anxiety with an overall accuracy of 76%.

Although there can be a substantial within-subject variability in scoring across time as mentioned by Meyerhoff et al [28], the reported SD for GAD-7 (3.50) is less than the RMSE achieved in this study. In a clinical use case, the GAD-7 score–based anxiety category is more relevant than the individual scores. Interrater reliability of anxiety disorder diagnosis is shown to have a κ value of 0.20 [44]. A key performance indicator for MHSS would be its ability to differentiate individuals across the anxiety categories with an accuracy over 70%. Each anxiety category (ie, none, mild, moderate, and severe) has a range of 4 points in the GAD-7 scale. As the RMSE in this regression model exceeds this range, this model would result in very low accuracy of anxiety category prediction.

The GAD-7 multiclass model achieved an overall accuracy of 50%, with a sensitivity of 63%, 37%, 41%, and 52% and specificity of 80%, 84%, 93%, and 74% for the none, mild, moderate, and severe classes, respectively. Prior studies performed in primary care clinics have noted that a cut-off score of 10 or higher on the GAD-7 scale has a sensitivity of 89% and specificity of 82% [45]. Although GAD-7 may be particularly useful in assessing symptom severity, a score of 10 or greater on the GAD-7 is most reliable for identifying cases of GAD. This supports the case for developing a binary classification model as an effective screening tool. With the available number of participants in the study, the statistical power for differentiating participants with severe anxiety from ones without anxiety using the digital phenotype as a marker was the strongest (76%). Based on testing various modeling algorithms including random forest, logistic regression, K-nearest neighbors, and RF, the GAD-7 binary XGBoost model achieved 76% accuracy with a sensitivity of 62% and specificity of 86%. These accuracy levels are higher than published results that use intrusive markers to predict generalized and social anxiety disorder [25], or that have used physiological markers to predict anxiety severity [46]. Along with the accuracy levels, sensitivity and specificity results for the GAD-7 binary model are also higher than studies done by Nemesure et al [47] and Fukazawa et al [48], which used binary classification for prediction of anxiety.

One of the key findings was the higher use of certain app categories such as “passive information consumption apps,” “games,” and “health and fitness” among participants with anxiety as compared with those without. Feature importance analysis has been performed by various previous studies, and they have demonstrated the usefulness of knowing these predictors [49]. Previous studies have stated various features such as daily screen time [25] as useful predictors. This study highlights certain app categories as important predicting features, allowing a deep dive into the digital usage patterns of people with and without anxiety. Whether the increased usage of such apps is a result or a cause of elevated anxiety is a topic for further exploration.

The correlation analysis performed between the items of the GAD-7 scale found that the highest correlated items were 1, 3, and 7. This has been a very interesting finding because the 2-factor structure of the GAD-7 scale has been suggested in previous studies such as Beard and Björgvinsson [50], where Items 1, 2, 3, and 7 belonged to the cognitive and emotional component of anxiety and 4, 5, and 6 to the somatic component. This points to the result that machine learning algorithms employed to generate MHSS are more sensitive in picking up the emotional/cognitive component of anxiety.

The MHSS for anxiety has the potential to serve as a complementary continuous metric to the GAD-7 questionnaire as well as clinical assessment of anxiety disorder. This metric has the advantage of being able to monitor daily anxiety levels with no respondent burden. This enables the use of smartphone-based sensing to overcome any “state-of-mind” biases. Given the metric’s sensitivity to the emotional/cognitive component of anxiety, it can help in overcoming those undiagnosed cases where somatic symptoms of anxiety result in a conflict in diagnosis. This is especially useful in cases where there is an overlap of physical symptoms (shortness of breath or palpitations) and cognitive symptoms (such as insomnia, restlessness) as well as an overlap with depression [9,19]. Another potential use for MHSS is outlining and differentiating the risk of comorbidities. Anxiety disorders are mostly comorbid with depression. A recent study using the same Behavidence research app was able to predict depression severity with the MHSS for depression. Choudhary et al [26] found that machine learning models that generated an MHSS for depression had high accuracy metrics (≥89%) and were able to distinguish between users with depression and those without. Coupled with the findings of this study, MHSS can distinguish between comorbid depression and anxiety, thereby improving clinical decision making.

One of the limitations of the study was that the GAD-7 questionnaire was collected at only 1 time point during the study. In this study the sample size was average, with unequal amounts of gender proportions and education background, which can affect the generalizability of the study, as GAD is a very commonly observed phenomenon. Although the study had almost equal proportions of mild, moderate, and severe groups of anxiety, this was an online recruited sample. With accurate model metrics, further studies should aim for having clinical samples and populations. Therefore, future models should focus on recruiting larger sample sizes and clinical populations to further test the applicability of such findings. Although the machine learning models indicate a higher accuracy of the GAD-7 binary model, the MHSS may have different thresholds for various levels of anxiety severity, which should be subjected to further research. Given the existence of comorbidities, particularly depression, a dedicated study to assess the correlation between MHSS for depression and MHSS for anxiety could generate valuable insights and shed light on how different interventions may be impactful.

The lack of access to mental health care can be addressed through the ubiquitously available smartphone and the development of passive and widely available screening technologies for detecting the most common mental health disorders. Objective smartphone-collected data contain enough information about an individual’s digital behavior to infer his/her mental states and screen for anxiety, and is a technology that provides remote, longitudinal, and continuous monitoring as an integrative and agile solution. Machine learning serves as an effective technique to mine such big data to derive accurate biomarkers for mental health conditions such as anxiety.