This article is part of Future Tense, a collaboration among Arizona State University, New America, and Slate. On Thursday, Sept. 28, at 9 a.m., Future Tense will hold an event in Washington on mental health and technology. For more information and to RSVP, visit the New America website.
Somewhere at MIT, a computer screen flickered, and Eliza was online. A Rogerian psychotherapist of sorts, she was always ready to listen and quick with subsequent questions. Never revealing much about herself, she instead encouraged her clients to proverbially “tell her more.” Yet Eliza never really knew more, never knew what was happening in the lives of those she conversed with, and never was able to contextualize their issues outside of scheduled sessions.
It was not her fault. It was 1964, and she was a computer program confined to clunky machines of the day. Now, half a century later, the concept of Eliza has transformed into a menagerie of apps that can carefully monitor your smartphone to track and even predict your mental health based on things like word choice in text messages, the speed of your speech, and usage patterns. Several startups and researchers today report they can use smartphone phone voice data to diagnose post-traumatic stress disorder, identify patients who might develop psychosis, and spot early signs of manic episodes in patients with bipolar disorder. There is even an effort to deliver therapy via a chatbot on Facebook. The unifying theme is that smartphone sensing will replace the need to “tell her more,” because Eliza-like apps and artificial intelligence programs might even know your mood state before you are even aware of it.
Smartphone sensing for mental health goes by lots of names—including digital sensing, passive data monitoring, digital phenotyping, and more—but the basic concept is simple. In using your smartphone throughout your daily life, you automatically generate a series of digital signatures regarding your activities and behaviors, which can be recorded. If you walk down the street to the local store, GPS sensors in your phone likely know how far you went, as well as how often you have left the house this week, the number of places you visited, and even whether you are following your routine. When you send a text message or answer an incoming call, the phone’s data logs know if you are more socially engaged today or perhaps calling one particular person much more in the past week. What about text messages? Do you make more spelling mistakes or type messages slowly when feeling tired or sluggish?
These kinds of smartphone data relating to mobility, sociability, and even cognitive state are readily accessible and offer the potential to contextualize daily life through a digital breadcrumb trail of clues about where we go, who we interact with, and how we are thinking. But does following this trail lead toward better mental health?
If you ask an entrepreneur, the answer is patently yes. Mental health disorders like depression are defined in terms of symptomatic irregularities such as alterations in mood, sleep, activity, social engagement, and thoughts—changes that can now be detected in real time by our smartphones. But this aspirational vision neglects several inconvenient truths and obfuscates the road ahead. Digital sensing is not yet a crystal ball unto human behavior, offering clairvoyant insights into our mental state. Rather, it is a new approach that offers potential benefits as well as risks, opportunities as well as challenges.
Thanks to the rapid proliferation of smartphones, wearable devices, and other sensor-laden technologies, amassing large quantities of digital data isn’t difficult. But large quantities of data do not necessarily translate to high therapeutic value, especially for mental health. This problem isn’t unique to apps for mental illness diagnosis or monitoring. For example, consider electroencephalogram, a technology to directly measure real-time electrical activity in the brain. Despite literally being able to record brain waves of those with mental illnesses, the clinical utility of this data in routine care remains limited. Even after decades of research on EEG, it is unlikely that a psychiatrist today will order one to help make a diagnosis, pick the right medication, or monitor recovery. That’s because it is tremendously difficult to turn raw data into useful information.
One oft-cited solution to the big data challenge of digital mental health data is to use artificial intelligence approaches like deep learning to help make sense of the raw data. Deep learning is the art and science of building enormous computer models—neural networks—that can be used to predict, classify, edit, describe, and create videos, images, and text. For example, apps can organize your photos by topic, even if you never labeled them yourself. But analyzing and helping human beings is much more difficult than helping a machine recognize the difference between a selfie and a landscape photo. Artificial intelligence programs still struggle with cancer diagnoses, even when complete medical records are available and even with medical knowledge of that cancer well characterized at the genetic level. Creating meaningful categories of mental illnesses is complex, making it difficult to create or train diagnostic algorithms. Simply put, psychiatric diagnosis is complex even when done by trained professionals. What appears as a new mental illness may actually be masking a medical disorder, such as a brain tumor causing changes in personality and behavior. If someone is suddenly making more errors made while typing into a smartphone, does it mean he is experiencing the early onset of schizophrenia? Or is he instead busy or distracted?
Furthermore, much of the smartphone and artificial intelligence research done today involves learning from people who are professional online survey-takers, often using Amazon Mechanical Turk, asking them to report their mental health symptoms without ever verifying their reported symptoms. Verifying symptoms is critical for ensuring that we are studying the right condition and training algorithms against the actual illness. For example, a recent study of smoking cessation found that more than 40 percent of people self-reporting one week of abstinence failed biochemical verification of their smoking status. Machine learning is only as good as its training data.
While digital mental health and its new data will eventually bring about a paradigm shift in how we diagnose, monitor, and treat mental illness, today we are only at the early stages of this transformation. Ultimately, digital technology will only advance our understanding and management of mental health if the right tools and analytical methods are developed. This calls for applying what is already known about both psychiatry and digital technology to build tools of measurement that are cheap, ubiquitous, reliable, and measure the right thing. Doing all of this is not rocket science, but that doesn’t mean it’s easy. It needs to be done carefully and in a transparent manner so that results can be verified, expanded upon by others, and shared to benefit the most people. Big data and deep learning will play critical roles, but they are not shortcuts and will require we do the same careful research and experimentation that one would expect for any new medical innovation. Until then, Eliza will still need to ask “tell me more.”