Orygen researchers are investigating whether machine learning and computer analysis of how a young person talks can be used to accurately predict if they will develop psychosis.
Currently clinicians manually rate certain aspects of speech to determine psychosis risk, but Orygen researchers are working to automate the process with the help of machine learning.
Orygen’s head of ultra-high risk for psychosis research, Professor Barnaby Nelson, is leading the so-called SPEAK study and said early research findings were encouraging.
“Machine learning analysis of speech has already been shown to be more accurate than clinical ratings at predicting who goes on to develop a psychotic episode over time,” Professor Nelson said.
“But those studies have been mainly in small US samples of about 30 to 40 cases.
“What the SPEAK study is doing is recruiting a much larger sample to see whether we can replicate that finding.”
How a person speaks is important because it reflects thinking processes, Professor Nelson said.
“Thought disorder occurs frequently in psychotic disorders. Thought disorder can consist of things like having difficulty putting sentences together, going off topic or having difficulty even producing speech in the first place,” he said.
Clinicians commonly use two methods to manually analyse speech and determine psychosis risk: latent semantic analysis, which involves examining the flow of meaning in a narrative; and part-of-speech tagging, which looks at grammar, syntax and sentence complexity.
Generally, the more complicated and less coherent a person’s speech, the more likely it is they are experiencing disordered thought and may be at risk of developing psychosis.
The SPEAK study will involve 450 participants, including 150 young people (aged 12–25) at ultra-high risk of psychotic disorder, 150 young people experiencing first episode psychosis and 150 healthy young people. Participants are being recruited from Orygen’s clinical services and also through the Centre for Addiction and Mental Health, Toronto.
“Study participants will take part in an open-ended conversation with a research assistant for about 20–30 minutes,” Professor Nelson said. “The conversation will be recorded, uploaded to a secure platform, and transcribed.”
Transcripts will be analysed using an automated machine-learning approach developed in partnership with global technology company IBM and Dr Cheryl Corcoran from the Icahn School of Medicine at Mount Sinai.
“The machine analysis classifies which participants may be at high risk for a psychotic episode and those who are at risk of relapsing after a first psychotic episode, based on their speech patterns,” Professor Nelson said.
Even if the study does not find that machine learning is better at predicting psychosis onset – the technology may still be useful, Professor Nelson said.
“It could be developed into something like an app-based solution where a young person speaks into the app at regular intervals and the app does an automatic analysis of their speech to see how they’re going from a symptom and mental state point of view,” Professor Nelson said.
“That could be really useful in situations where a young person doesn't have access to a clinician – they might be discharged from clinical services, or they might not be turning up to sessions for whatever reason.”
The SPEAK study is funded by the National Institute of Mental Health (R01 MH115332-01).