Using Machine Learning to Predict Psychosis

Self driving cars, human-like robots and robust medical technologies aren’t the only types of artificial intelligence that are making advances in people’s lives. In the mental healthcare industry, clinicians and patients alike are benefitting from advances of machine learning that are making strides in diagnosing mental illness.

Current psychosis prediction models used by human clinicians are riddled with errors and gaps in knowledge. At present, only 22 percent of clinical high-risk patients are accurately predicted to develop psychosis. Because of this diagnostic flaw, more attention has been turned towards machine learning in hopes algorithms can bring more accuracy to the process. A total of 52 researchers and clinicians at a consortium of universities across Germany, Switzerland, United Kingdom, Finland and Italy were able to use machine learning to accurately detect who would develop early stage psychosis based off of psychosis risk factors. Genetic information, previous psychiatric history and early stage psychotic signs were the predictive criteria for the algorithm’s magic. The data were collected through a European group that included participants at risk for the early stage of psychosis, which is called the prodromal period, for its status as a significant time when symptoms first start to appear in a given individual.

When considering a psychotic population, clinicians must be aware of the tendency for patients who start hearing or seeing things other people aren’t, signs of grandiosity, paranoid thoughts and other signs classically seen in people with psychosis. To test the effectiveness of machine learning against human clinician ratings, the at-risk patient group was followed up every three months by clinicians for a total of 18 months, then nine months until the 36 month mark. The researchers used questionnaires to gauge how active psychosis was in the participants during these follow-ups. Assessments performed by the clinicians were then compared with machine learning models to determine just exactly who was better at predicting the onset of psychosis.

Machine learning technology then analyzed risk factors such as childhood adversity, behavioral functioning, early stage psychosis symptoms, genetic information and brain scans using functional magnetic resonance imaging (fMRI), a tool used to model a three dimensional picture of the brain. By virtue of the AI, the algorithm was able to accurately predict 85.9 percent of the time who would develop psychosis from the data—which is significant. The findings from the study revealed just how powerful our total knowledge of psychosis, when put into our computer counterparts, can positively impact diagnostic processes both in speed and accuracy. These results could provide better turn-around times for people waiting to receive a psychotic diagnosis, or otherwise serve as a preventative strategy for warning families with a genetic history of schizophrenia.

This study, and others like it, suggest incredible shifts in the ways we could potentially diagnose and treat mental illness. People who live with schizophrenia often don’t know they have the disorder until it appears in its stereotypical form: a person who is considered “crazy” and “lunatic” from onlookers who poorly understand the disorder. But with these new computational techniques, psychosis may be able to be detected earlier and in younger patients who are at risk for developing psychotic disorders.

The implications of this research are beneficial to the field of mental illness at large. Having a machine accurately predict mental illness gives researchers insight to how mental illness actually functions. Clinicians may learn a thing or two about what makes a mental illness an illness from their soon-to-be machine learning peers.