Mithilfe von Computer-Vision-Techniken an Tonbildern kurzer Sprachfragmente („Spektrogramme“) trainierten Forscher ein neuronales Netzwerk auf Sprachaufzeichnungen von Menschen mit und ohne Schizophrenie. Die Ergebnisse legen nahe, dass die Analyse alltäglicher Sprachmuster bei der Diagnose und Verfolgung von Schizophrenie hilfreich sein könnte.
https://www.nature.com/articles/s44277-025-00040-1
3 Kommentare
That’s cool, but given that one of the primary symptoms of schizophrenia its effects on speech, that feels like it’s doing something a human can already do.
These kinds of studies are interesting, but could
never be used clinically in our current environment.
Schizophrenia, and other psychiatric disorders, is diagnosed based on criteria according to the DSM. You can’t make these diagnoses based on correlation with speech patterns.
Other disease, like cancer, is diagnosed when a biomarker or symptom prompts a more precise test. The hope with AI is that it would let us infer more from the readily available biomarkers, or help us choose the correct follow-up test.
In psychiatry, at least for schizophrenia, the “follow-up test” would be mainly the psychiatric evaluation by a trained professional who considers the entirety of the person. The result of the evaluation is a classification, rather than a validation like with cancer.
So the usefulness of “tools” like this is pretty limited. The bottleneck is still a human examination, and the examination is going to be the only thing that determines the diagnosis of schizophrenia.
I’m currently writing a paper on the history of speech markers in a variety of mental illnesses. They have literally been suggested for a century now, in hundrets of publications. I’ve yet to hear of a psychiatric hospital where any of this stuff is actually used.
Why? Many reasons, but as can be read in the paper I’ll link below clinical models always need to address a clear clinical decision point and have to do it *better* than other diagnostic markers.
I’ve seen a paper which literally classified patient speech for whether the person is acutely suicidal, from what kind of speech sample? Recorded suicide notes.
The situation in this case is similar: what’s more available as a datapoint: the content of what a patient says and their behaviour otherwise, or a spectrogram and a Neural Net which clinicians would have to run?
So I’m super sceptical when I come across the 1000th paper suggesting some speech based biomarker for mental illness.
Interesting and very acclaimed paper making similar arguments for oncological prediction models:
https://www.nature.com/articles/s41698-024-00553-6