Mental-health professionals have long known that disorganized thought patterns present themselves in spoken language. Disjointed speech, where one thought is not well-connected to the next, is common among people with schizophrenia.
Analyzing the speech of patients for clues that indicate psychopathology is not new. Back in 1979, Sherry Rochester's book Crazy Talk studied the topic in depth. The 1990s saw many guidelines developed to help doctors predict psychosis from listening to dialogue. They can do so with remarkable accuracy—nearly 80% of the time, they are right.
But this is the kind of problem that computers are very good at, and the machines can do better: A new study published in Nature Schizophrenia  showed not just that computers were good, but that they were perfect. Algorithms correctly predicted which at-risk youth would go on to develop psychosis over a 2.5-year period with 100% accuracy.
The algorithms did this by analyzing the subjects' spoken dialogues, measuring the coherent flow from one sentence to the next. The programs measured disruptions by analyzing the structure of sentences. If there was a single jarring disruption, it was a sign that psychosis might follow.
Guillermo Cecchi, one of the study authors, told The Atlantic:
“When people speak, they can speak in short, simple sentences. Or they can speak in longer, more complex sentences, that have clauses added that further elaborate and describe the main idea. The measures of complexity and coherence are separate and not correlated with one another. However, simple syntax and semantic incoherence do tend to aggregate together in schizophrenia.”
The algorithms have an advantage over humans in that they don't lose focus. A doctor listening to a patient speak may jot down a note or lose deep focus on what is being said and miss one of these subtle episodes. The computer doesn't face that risk.
This first study was small, with 34 subjects, so one would expect the algorithms won't maintain a perfect record as they are deployed on a wider scale. However, the results are promising in several ways: