Table 1.

Summary of passive sensing behavioral data collected throughout the study.

Behavior	Description	Derived hourly features
Acceleration	3-axis acceleration data were collected from a smartphone, sampled from 50-100 Hz. Previous CrossCheck studies utilized the Android activity recognition APIa, which classifies activity data as follows: on bicycle, still, in vehicle, tilting, or unknown. In this study, we chose to use raw acceleration features to make our anomaly detection system independent of a specific activity recognition API platform	Mean acceleration over the hour
App use	CrossCheck recorded apps running on a user’s smartphone every 15 min	Number of unique apps opened within an hour
Call	Phone calls can indicate social interaction and communication. We tracked when incoming, outgoing, missed, rejected, and blocked calls occurred	Number and duration of incoming, outgoing, missed, rejected, and blocked calls
Conversation	Previous studies have investigated the link between conversations, human voice, and mental health [12,33,34]. We detected human voices and conversational episodes using algorithms from our previous work [35]	Number and duration of conversations
Location	Previous research has shown that location can be associated with mental health [12,13,36]. We tracked location information from users through their smartphones	Time in primary, secondary, and all other locations as well as total distance travelled in the hour
Screen activity	The amount of time users spend on their phones can be tracked to learn normal daily behaviors. The time users’ screens were on versus off was recorded	Number of times the phone was used as well as the duration of use
Sleep	On each day, the sleep duration, onset, and wake time were detected. These calculations occurred using a combination of information based upon users' screen time, physical activity, ambient sound, and light [12,37]	Sleep duration, onset. and wake time. As we estimated only the longest sleep episode per day, this is technically a daily feature. We replicated these features across all hours within a single day
Text	Text messages are another indicator of social interaction. We tracked when texts were received, sent, drafted, left in a user's outbox, failed to send, and were queued for sending	Number of received, sent, drafted, outbox, failed to send, and queued messages

^aAPI: application programming interface.