Skip to main content

IBM Research Is Working to Improve Mental Health Diagnoses

Imagine being able to predict the onset of a mental health condition before it hits. Or receive a diagnosis at home instead of a psychiatrist’s office. Or being able to adjust your medications based on patterns in your speech when you casually talk with family and friends.

Within the next five years, you may not have to imagine that at all. According to Guillermo Cecchi, principal research staff member, manager of Computational Psychiatry and Neuroimaging at IBM Research, often-debilitating mental health issues may be much easier to spot and manage, just based on speech and writing alone. According to Cecchi, an approximate minute of speech—thanks to advanced analytics, machine learning, natural-language processing and computational biology—will offer a quantifiable and accurate peek into a patient’s mental health, allowing doctors to more quickly address mental health concerns.

IBM Systems Magazine (ISM): What are the current problems with diagnosing neurological and mental conditions?

Guillermo Cecchi (GC): Well, there are two main problems. One is the consistency of the diagnosis. If I diagnose you with a certain condition, let’s say depression, a colleague from another institution may have a different diagnosis. Guidelines exist, but very few issues can be quantified in terms that are independent of the background of the patient and the context of the evaluator. The second problem is that even if the evaluation was consistent, it cannot be conducted except in the context of a clinical setting, whether that’s someone coming to a clinic or sending a social worker or psychiatrist to the patient. Bringing the patient to the clinic can cause real problems for some critical conditions, where, for example, they might be scared of being outside or being in the clinic. Sending someone to see the patient can also present a difficult proposition, especially in terms of accessibility. The patient might be in a remote location and the evaluator in the city. It may take several hours to get to the patient. Also, these professionals are chronically overworked, even in places with a relatively high density of treatment providers.

ISM: It’s pretty standard psychiatrists or psychologists sit down with patients and chat with them to help determine a diagnosis. How are you benchmarking against that, if that’s the right word for it?

GC: We take patients we know have a certain condition as part of the same group of psychiatrists in the same institution and use that as the gold standard to say, ‘Well, whatever we find in the speech patterns that predict the diagnosis is what we’ll consider a marker for the condition.’ In some cases, this is a bit less problematic. We did a study on the predictives of psychotic onset, and because the event itself is so dramatic, it’s hard to mistake it for anything else. No one is saying this is a potential diagnosis. It’s the actual prognosis. We’re still using the gold standard of the psychiatrists, but before the events take place, we actually beat them to the prognosis.

ISM: How do you quantify mental conditions?

GC: We’re using a combination of machine learning—or what we would call a data-driven or blackbox approach—to try to extract a large number of features to see which have predictive power, and at the same time, we’re trying to formalize ideas that have been around in psychiatry, neurology and psychology for a long time and try to quantify that in speech.

One example is the idea that it’s very important to define psychosis as the “flight of ideas” or incoherence of speech. So the idea that someone who is psychotic may be talking about something and all of a sudden—poof!—goes in a complete different direction, switching topics in an unrelated and inconsistent manner. So that’s something we were able to quantify in certain ways. Of course, this is not an exact science, but we found ways of formalizing that and measuring it in speech that had predictive power. We validated though independent means.

Another one is the idea of poverty of speech, which is important in psychosis and autism spectrum disorders. We formalized that by looking at the syntactic structure that’s produced by the patient and measuring how complex or how simple that structure is. In fact, we found that this is a very predictive feature in some people. We’re taking psychiatrists’ collectively acquired knowledge and formalizing it into something that’s measurable.

ISM: How long would it take to reach a conclusion regarding someone’s mental state?

GC: In the case of prediction of psychotic outbreak, we worked off interviews that were around 30 to 45 minutes. In the case of fully developed psychosis, one to two minutes are enough. If you’re already psychotic, schizophrenic or manic, it really takes very little time to identify your condition, but this study is more time-intensive because we’re making a prediction about the future. This is more complicated, but there seems to be enough information available when you analyze a half an hour of speech.

ISM: Does this work with writing as well?

GC: Yes. When you’re writing, you have more time to edit, but changes to speech and written expression are essentially the same. If you’re speaking and you become psychotic, you become more incoherent—and that also applies to your writing.

ISM: How will this work in a clinical setting?

GC: I think there are two ways of thinking about this. One is in terms of inpatient/outpatient access. Let’s look at a concrete example: The department of internal medicine at Mount Sinai School of Medicine in New York. Why the department of internal medicine? Because a strong connection exists between depression and the immune system. Our goal is to provide physicians with a very quick assessment on which they can act accordingly.

We are also working with Pfizer to investigate if speech—in conjunction with other wearable and environmental sensors—can be used to passively and unobtrusively monitor subjects during the course of the day as drug levels vary. The goal of such continuous monitoring is to improve clinical trials, and ultimately such technologies could enable personalized and closed-loop therapies.

ISM: Does this involve the use of a mobile device of some sort?

GC: Yes. It would have the speech analysis application on it—we’re working with Watson Health* on this.

ISM: Would patients have these devices, or would they be confined to a clinical environment?

GC: The patients would have the mobile device or even a smart device in their home on top of a kitchen table. It could be entirely dedicated to this or to pick up speech patterns or changes in your voice to accrue information regarding your state or provide you feedback for adjusting your medication. So this would be in the context of your daily life as opposed to the context of the clinic with the doctor in a white coat. I cannot emphasize enough being able to do this while you’re at home where you are more naturalistic and there’s less external context that may be interfering with your speech.