Your smartphone may know more about your mental health than you

“How have you been feeling since we last met?”

As a psychiatrist, this is how I often open a session with a patient. But as a Silicon Valley-based neuroscientist, I am struck by how wholly inadequate the question is for the task.

Not only is mood a complex and hard-to-define entity, but the capacity of a person to self-report on even simple facets of their experience is surprisingly limited. For example, we have found that the way depressed patients rate their own attention and concentration has little to do with objective laboratory tests of those capacities, and instead reflects their overall psychiatric distress.

While there is no doubt that mental illness involves individual subjective experience, the reality has been that asking about these experiences has simply not worked as a method for determining the cause of the illness or guiding its treatment. Given that mental illnesses reflect alterations in various brain functions, psychiatry has long wanted good measures of those alterations.

A dashboard for diagnosis

Imagine the clinician and patient having a dashboard at their fingertips that showed daily fluctuations in objective metrics of the patient’s behaviour, physiology and brain function, with machine learning algorithms flagging concerning trends or early patterns indicative of response to a treatment.

Arguably only through better objective measurement can we hope to improve our ability to manage mental illness. This sea change in the way we characterize individuals for the purpose of better diagnosis and treatment – often termed ‘digital phenotyping’ – is an exciting and rapidly emerging area of both academic and industry-based research efforts. This approach may also help us reach the substantial proportion of people across the globe with psychiatric conditions who never seek treatment due to barriers to access, stigma or other reasons.Progress in digital phenotyping has come through two sources. One is technological. The ubiquity of smartphones has created an ability to capture passively a torrent of data about the daily behaviour, non-content aspect of speech, geolocation and even word use of people across the globe. Given how much of our lives are lived through smartphones, our daily ‘digital exhaust’ is a rich source of information about us. Even how we interact with the smartphone itself, including factors such as typing speed, errors and response times, can index basic cognitive capacities in continuous and passive fashion.

There has likewise been a proliferation of consumer devices, such as smart watches and activity monitors, that can provide complementary data about physiology. Wearable technologies are also rapidly evolving, including for measurement of brain electrical activity in relatively unobtrusive ways. ‘Smart pillows’ are being developed that can measure the quality of our sleep, an activity that represents the single biggest component of our day.

Researchers are even unobtrusively extracting meaningful behavioural and physiological indices during daily driving commutes, as cars can make well-controlled measurement environments. Using some combination of these approaches could reveal ongoing fluctuations in mood, cognitive functioning, likelihood of a drug relapse, emergence of a psychotic episode and countless other critical facets of mental health.

The second source of progress, in light of the vastness of our digital footprints throughout daily life, is the rapid advances in machine learning, which have already revolutionized many facets of the technology industry. Every year has broken new ground with respect to the power and flexibility of these algorithms. This is anticipated to continue for some time. Critical advances in machine learning, such as for visual object recognition in pictures, has come through the availability of large amounts of well-curated data. That obstacle is being rapidly addressed by digital phenotyping efforts.

The bigger challenge will now be how to extract the most meaningful signals for each person. Unlike in other industries – advertising, for example – in which machine learning algorithms can improve outcomes in the aggregate, in medicine we must be able to make firm conclusions about individual patients. Doing so requires far greater confidence in the output of our algorithms.

This will also mean that some digital phenotyping applications, such as those that ‘rule in’ a person as a good candidate for a particular intervention, will naturally evolve before those that ‘rule out’ a person for an intervention. In most cases, greater certainty is required for denying than giving a particular intervention.

Another challenge is the fact that most medical tests typically do not change over time, whereas digital phenotyping-based measurements are likely to continue evolving as additional data is gathered and new technologies are leveraged. Regulatory agencies have only just begun contemplating how to hand rapidly evolving ‘software as a medical device’ applications, with only general guidance available now.

Despite many reasons to be excited, digital phenotyping has so far remained at the proof-of-concept level. We are still waiting for the first successful large-scale implementation with tangible health benefits. The ultimate success of digital phenotyping, especially its acceptance by patients, will also require addressing core challenges around privacy and data security, since some of this data may reflect the digital footprints of highly personal and private individual experience.

Nonetheless, the arc of digital phenotyping clearly points toward its importance for the future of psychiatry, as well as for behavioural health more generally. This is in no small measure due to the fact that little else has changed in psychiatry for the past few decades, and technological platforms for phenotyping are now rapidly emerging.

Ironically, while the introduction of digital phenotypes as a complement to traditional subjective descriptions of an individual will involve substantial disruption of traditional clinical decision-making and workflows, it is the very ubiquity of the mobile technologies that these efforts draw on that makes this disruption seem inevitable.