Detection of Behavior Change in People with Depression

Major Depressive Disorder (MDD) is the most common mental health disorder and remains a leading cause of disability and lost productivity with huge costs for society. MDD has high rates of relapse and recurrence, and it has strong correlations with feelings of low social support and disrupted sleep. However, MDD is also commonly misdiagnosed by primary care providers, which leads to delayed treatment and unnecessary suffering. Changes in technology now make it possible to cheaply and effectively monitor social and sleep behaviors, offering the potential of early detection of the onset of MDD. We report on the design of Big Black Dog, a smartphone-based system for gathering data about social and sleep behaviors. We also report on the results of a pilot evaluation to understand the feasibility of gathering and using data from smartphones for inferring the onset of depression. Introduction Each year 7% of Americans experience an episode of major depression [NIMH, 2011]. As a leading disability, depression has huge costs in terms of reduced productivity and absenteeism [RAND, 2008]. Most people seek help from their primary care provider (PCP); however, PCPs fail to recognize depression symptoms 65% of the time [Jencks, 1985; Coyne et al, 1995]. The delay in diagnosis and treatment increases the time people suffer from this condition. Research has shown that early detection of a first episode or of a recurrent episode can have a major positive impact [Halfin, 2007; Kupfer et al., 1989]. MDD has high rates of relapse and recurrence, and it has strong correlations with feelings of low social support and disrupted sleep. For example, a lack of social support has been found to predict depression as well as many other health related issues [Sias and Bartoo, 2007]. Past work has also found that sleep disorders are correlated with depression [Livingston et al., 1993], and disrupted sleep has been found to predict recurrences [Perlis et al., 1997]. The meteoric adoption of smartphones offers a tantalizing opportunity, in that many people now carry a networked, sensor-rich device almost everywhere they go. These changes make it possible to cheaply and effectively monitor people’s activities and behaviors, which could then be used to detect the early onset of MDD. Towards this end, we have developed a system called Big Black Dog (BBD) to detect the onset of major depression, allowing for earlier diagnosis and treatment of first episodes, relapses, and recurrences. Our angle of attack is to capture data about and model social behaviors and sleep behaviors. This paper reports on the design and on a pilot study to understand the feasibility of our approach. Related Work Researchers have investigated a number of behavioral signals to detect the mental state of people, using such approaches as brain signals (Stewart et al., 2010), heart rate (Vikram et al. 2002) blood pressure (Shinn et al., 2001), voice prosody (Cohn et al., 2009; France et al., 2000), and facial expression (Cohn et al., 2009) as proxies for psychophysiological information. EEGs, heart rate trackers, and skin conductors provide rich streams of data; however they are cumbersome to wear, often difficult to use, and typically limited to being used in clinics. Text mining has also been investigated as a method to detect depression. De Choudhury et al. used tweets to detect depression (De Choudhury et al., 2013). Important indicators included a decrease in social activity, raised negative affect, highly clustered ego networks, heightened relational and medicinal concerns, and greater expression of religious involvement. Smartphones offer rich a set of built-in sensors including accelerometers, location (GPS, WiFi ID, signal strength), light, and microphone. In past work, we used call logs, text logs, and contact list data to model social behaviors. In other past work, we used smartphone sensor data to model sleep quantity and sleep quality. BBD uses many of the same features from these two systems and applies them in a pilot study specifically for depression. The most relevant piece of past work is Mobilyze (Burns et al., 2011), a smartphone app that collects sensor data to detect the user’s cognitive state. Mobilyze uses machine learning models to predict mood, emotions, cognitive/motivational states, activities, environmental context, and social context. This system has been used for intervention and not for depression detection. Our research builds on this previous work, focusing on detecting the early onset of depression. Unlike Mobilyze that constantly prompts users for ground truth data, we only ask for a weekly survey. Our goal is to track behavior without any effort from the users. In the following sections, we report on design of the BBD mobile application followed by the pilot study and primary results. Overview of the Data Collection app We designed and implemented BBD as an Android app that uploads captured data every day to our server (see Fig. 1). BBD collects sensor data that might reveal behavioral and environmental factors, including noise amplitude (from microphone), location, WiFi SSIDs, light intensity (from ambient light sensor), and movement (from accelerometer). To minimize power consumption, each of these sensors captures data on a relatively light duty cycle (see Table 1). If the battery charge decreases below 30%, the phone samples the information less frequently. If the battery is very low (below 15%), we pause the logging. BBD also captures device states, such as screen on/off, apps currently running, and the battery-charging state. For example, “screen on” in the middle of night is a signal that a user is probably not asleep. Lastly, BBD collects daily call logs and text messages from the phone. BBD stores captured data in a database in the protected storage area of the phone. It creates a new database each day and uploads the previous database to the server. This strategy reduces the risk of data loss and complications that can come when attempting to upload large files. Figure 1. System overview of BBD. Raw data Frequency Low Battery Very low Battery Sound (1hz) Every 2 minute (1min break) Stop Apps When the screen is turned on, every 10