Digital phenotyping: a global tool for psychiatry

In 2050, when psychiatrists look back at the first two decades of the 21st century, what will they recognize as having the greatest impact? No doubt the revolution in genomics, which has given us new insights into the risk architecture of mental illness, and the revolution in neuroscience, which has given us a new view of mental illnesses as circuit disorders, will be considered important. But perhaps the revolution in technology and information science will prove more consequential for global mental health. If this sounds like hyperbole, consider two supportive data points. First, in the past decade smartphones have become nearly ubiquitous. There are over three billion smartphone Internet subscriptions, each device with the information processing capacity of the supercomputers of the 1990s. In many parts of the world that lack credit cards, phones have become the primary way to conduct commerce. Second, broadband access to social media and search platforms is becoming global. In 2016, 3.3 billion people had Internet access, one third of whom were in India and China. Even in areas without easy access to clean water, ownership of a smartphone and rapid access to information have become the symbols of modernity. The smartphone and the Internet can solve specific problems that we face in psychiatry, but their clinical use also raises new ethical challenges. What specific problems can be addressed by the smartphone? Our lack of objective measurement has handicapped both diagnosis and treatment in psychiatry. As just one example, our assessment of depression depends largely on selfreports of sleep, appetite and emotional state, although we recognize that people with depression are biased in their assessments. The smartphone offers us an objective and ecological source of measurement. This approach, now called digital phenotyping, is based on sensors (activity and location), voice and speech (sentiment and prosody), and, perhaps most important, human-computer interaction. Human-computer interaction measures not what you type but how you type. Subtle aspects of typing and scrolling, such as the latency between space and character or the interval between scroll and click, are surprisingly good surrogates for cognitive traits and affective states. If this seems improbable, remember that many of our neuropsychological tests, such as the Trails A and B tests or the Digit Symbol Substitution, are not substantially different from the psychomotor requirements of operating a smartphone. In a sense, those gold standard tests of cognitive control and information processing are attempting to assess how we function. In a world where we spend so much of our lives on our smartphones, could it become possible to assess how we function directly and continuously rather than using laboratory measures at a single point in time? The promise of digital phenotyping is that this objective measure happens in the context of a patient’s lived experience, reflecting how he/she functions in his/her world, not in our clinic. Signals from a new mother struggling with depression may look quite different during a 3 am feeding compared to what she reports to her clinician the next day. This kind of ecological and continuous measurement addresses some of the central issues that challenge our field. We know that most people with a mental illness do not seek help, and those who do seek help usually arrive after considerable delay. For populations at risk, such as post-partum women or victims of trauma, could digital phenotyping signal the transition from risk to the need for care? For people in care, too often we fail to preempt relapse. For patients in treatment, could digital phenotyping serve as a “smoke alarm” providing early signals of relapse or recovery? Digital phenotyping is still being developed as a clinical tool. It seems clear from the early results that, although activity and geolocation data are non-specific and noisy, for some people changes in activity can be an early sign ofmania or depression. Speech and voice may also yield clinically relevant signals. We have known for decades that when people are depressed their pronouns shift to first person singular. But again, the sensitivity and specificity of these findings still need to be defined. Putting sensor data, speech and voice data, and human-computer interaction together might provide a digital phenotype that could do for psychiatry what HgbA1c or serum cholesterol has done for other areas of medicine, giving precision to diagnosis and accuracy to outcomes. The opportunity of this new approach to measurement is matched by an ethical challenge.When doesmeasurement become surveillance? Is tracking geolocation or collecting speech too intrusive? How can patients trust that digital phenotyping data will be protected? Even if patients consent to have their smartphone monitored, is there full transparency and a deep understanding of what data will be collected and how these data will be used? Who owns the data? For psychiatry, one of the most informative phone signals might reside in the “digital exhaust”, such as search history or social media posts. Those signalsmight confess suicidal intent or early signs of psychosis. Does the value of this information outweigh the intrusion of privacy required to obtain it? All of these issues are part of an active debate, as merits any new promising technology. To be clear, digital phenotyping is still a research project conducted on small numbers of consented volunteers. While researchers hope this approach will solve global mental health problems, the scientific and ethical issues need to be resolved before digital phenotyping becomes a tool for population health. Some of the most vexing issues may have technical solutions. For instance, human-computer interaction is “contentfree”. This approach collects how you type, not what you type and, therefore, might be less intrusive than monitoring geolocation or search history. Tools that can analyze smartphone