Behavioral Indicators on a Mobile Sensing Platform Predict Clinically Validated Psychiatric Symptoms of Mood and Anxiety Disorders

Background There is a critical need for real-time tracking of behavioral indicators of mental disorders. Mobile sensing platforms that objectively and noninvasively collect, store, and analyze behavioral indicators have not yet been clinically validated or scalable. Objective The aim of our study was to report on models of clinical symptoms for post-traumatic stress disorder (PTSD) and depression derived from a scalable mobile sensing platform. Methods A total of 73 participants (67% [49/73] male, 48% [35/73] non-Hispanic white, 33% [24/73] veteran status) who reported at least one symptom of PTSD or depression completed a 12-week field trial. Behavioral indicators were collected through the noninvasive mobile sensing platform on participants’ mobile phones. Clinical symptoms were measured through validated clinical interviews with a licensed clinical social worker. A combination hypothesis and data-driven approach was used to derive key features for modeling symptoms, including the sum of outgoing calls, count of unique numbers texted, absolute distance traveled, dynamic variation of the voice, speaking rate, and voice quality. Participants also reported ease of use and data sharing concerns. Results Behavioral indicators predicted clinically assessed symptoms of depression and PTSD (cross-validated area under the curve [AUC] for depressed mood=.74, fatigue=.56, interest in activities=.75, and social connectedness=.83). Participants reported comfort sharing individual data with physicians (Mean 3.08, SD 1.22), mental health providers (Mean 3.25, SD 1.39), and medical researchers (Mean 3.03, SD 1.36). Conclusions Behavioral indicators passively collected through a mobile sensing platform predicted symptoms of depression and PTSD. The use of mobile sensing platforms can provide clinically validated behavioral indicators in real time; however, further validation of these models and this platform in large clinical samples is needed.


eMethods -Participants
During an additional in-person screening before consent, participants filled out the Beck Scale for Suicide Ideation (BSS;Beck et al., 1993). If participants had a positive response on any item, they met with a clinical social worker who administered the Mini-International Neuropsychiatric Interview (M.I.N.I.) suicide module and excluded participants at immediate risk of self-harm or selfreported suicidality. No participants were excluded for this reason.

eMethods -Mobile sensing platform architecture
Cogito Corporation developed a mobile sensing platform for use in data gathering, storage, and analysis. The software allows for data collection, secure transfer, ecological momentary assessment, and audio recording. Digital trace data was collected from predefined, configured probes built into the phone's operating system. Data from probes were gathered on intermittent fixed schedules. Database files were encrypted on the mobile device before being securely transferred to centralized servers. On the server, data was decrypted and then validated to assure data structural integrity. This step was critical as the underlying mobile operating system can update at any time and change the structure of the data files. Validated data were then stored in databases and encrypted at rest. These data are later transformed into feature and modeling results by the computational engine embedded within the platform software. The platform also includes an ecological momentary assessment system with the ability to present, collect, and store self-reported survey results from individuals. An audio recording functionality allows users to record audio check-ins to gather voice data.

eMethods -Data security & privacy Protecting Identity
All data gathered and stored on the system were marked by a unique identification number (UID). The identities of the individuals with whom participants communicated were also protected, using a salted hash approach where a random string represented the same phone number per participant's social contact throughout the study. While a single hash was generated per social contact on each device, a different hash was created for the same contact on another individual's device. Thus, the salted hash method protected the identities of those in participants' social network, and, by extension, the identities of subjects themselves.

Data Storage on the Device
Once data were gathered on the participant's device, they were immediately encrypted, according to military grade AES, in a publicprivate key pairing. As the data were always stored in an encrypted format on the subject's device, the research application and any other app, as well as the user him or herself, could not access the data. Given the encrypted nature of the stored data, if a participant lost his or her mobile phone, there would have been an extremely minimal risk of privacy breach.

Data Storage on the Platform
Data were transferred from the participant's phone to Cogito servers through a Secure Sockets Layer (SSL), an encrypted connection. Once data reached Cogito servers, the information was decrypted for analysis, and encrypted at rest. All data were stored locally at Cogito, and identified by unique identification number only. Data access was limited to named study staff.

eMethods -Digital trace data
The mobile sensing platform collected six main categories of digital trace data over the 12-week study period: Activity, Social, Location, Device Interaction, Device Information, and Vocal Cues.
Activity data included Accelerometer and Gyroscope probes, both of which were gathered every 30 minutes. The Accelerometer variables were represented by the rate of rotation of the phone around the X, Y, and Z axes. The average amount of Accelerometer data gathered per week for each participant was 54,889 readings, and the total amount of Accelerometer data collected over the 12-week study period across all participants was 47,006,966 readings.
Social data was captured through Contacts, Call Log, and SMS Log probes. Data from the Contacts probe were gathered after every phone re-start while Call Log and SMS Log data were collected once each day. The average amount of Call and SMS Log data for each participant per week was 107 calls and 190 SMS messages. The total amount of gathered Call and SMS Log data was 90,976 calls and 166,948 SMS messages.
Location data was measured via a Location probe, from which data were gathered every 30 minutes. The average amount of Location data gathered per week for each participant was 250 GPS readings, and the total amount of data collected was 218,822 GPS readings.
Device Interaction was assessed through a Screen probe, which collected data every time the the phone screen was turned on/activated and turned off. The average amount of data gathered per week for each participant was 1,179 screen readings, and the total amount of data collected was 1,033,121 screen readings.
Device Information data were measured through the probes Battery, Android Info, and Hardware. Hardware and Android Info data were gathered once a day, while Battery data were collected every hour.
Vocal cues were sampled through participants' audio diaries. The average amount of data collected per week for each subject was approximately 1 audio diary, and the total amount of data gathered was 847 audio diaries.

eMethods -Survey instrument
Participants were asked if they felt the application violated personal privacy, if they consciously changed cell phone use because of study participation, if they would be interested in using a similar application in the future, and if they would like to receive personalized feedback. They were also asked if they would be willing to share personalized health information collected through the application with friends, family, healthcare providers, mental health providers, similar patients, researchers, insurance providers, or government health organizations. These were averaged to create a comfort with personal sharing score. Participants were also asked if they would be willing to share anonymized health-related data with researchers, government organizations, other patients, support groups, and insurance providers. These were averaged to create a comfort with anonymize sharing score. All responses were on a 5-pt scale from not at all to extremely.
In the survey below, we refer to the application as Cogito VetGuard.