An unusual cooperation brings together medicine and the humanities in the search for a new way to diagnose psychosis. The approach makes use of artificial intelligence to gain data on what is actually ‘normal’.

Eva Schissler

It begins gradually: A face on a billboard appears to send a secret message, but its content is unclear. Weeks later, the impression may arise that one’s thoughts are being controlled. Added to that, strange voices in one’s head, the feeling of something crawling on one’s skin or seemingly distorted faces on the street.

We humans rely on our sensory impressions and our ability to process and interpret everyday perceptions correctly. If someone doubts this ability, we might think we are being ‘gaslighted’: The person wants to convince us of a false reality, even though we actually see, hear and feel the opposite.

In psychosis, it is the other way around. The brain really does receive false sensory information; it can no longer distinguish between actual and imagined perceptions and draw logical conclusions. It is a frightening situation in which family and friends expressing doubts no longer get through to the person. The construct with which psychotic patients try to put together a meaningful overall picture seems all too logical to them.

There are no exact figures, but it is estimated that around three percent of all people develop psychosis at least once during their lives. The reasons for this are manifold: In addition to physical causes such as brain injury, hormone imbalance or the consumption of psychoactive substances, psychosis can also occur without external influences. Although there is a genetic predisposition, the exact trigger in these cases is unclear. Some people suffer only a single psychotic episode, others several, and yet others become chronically ill. In most cases, medication and psychotherapy help them regain their mental health.

More than communication

In order to diagnose psychosis, psychiatrists depend on patients self-reporting. Although the diagnosis is relatively reliable, it remains a subjective medical assessment even in the best of cases. With their research project ‘LAMBDA: Language Markers and Brain Dysfunction in Early Psychosis’, the linguists Professor Dr Klaus von Heusinger and Dr Derya Çokal from the Institute for German Language and Literature I as well as the psychiatrist Professor Dr Joseph Kambeitz want to develop better diagnostic tools. Joseph Kambeitz, head of the Early Recognition and Intervention Centre for Mental Crises (FETZ) at University Hospital Cologne, said: “Since many things in psychiatry are associated with language, it makes sense to cooperate with linguists.”

The team works on the assumption that, beyond the person’s self-assessment, language can provide important indications of psychosis. Language is not only a means of communication, it also expresses whether the person speaking is able to think in a structured way. In addition, one of the typical signs of psychosis is that people hear voices or that they do not indicate a change of topic in a conversation, but rather jump back and forth between different topics. Thus, the connection with language is already evident in the disorder itself.

The researchers are searching for patterns that express specific errors in the processing of language. “We want to identify a linguistic biomarker that is typical of brain dysfunction leading to psychosis,” said Derya Çokal, a postdoctoral researcher at Collaborative Research Centre 1252 ‘Prominence in Language,’ in which Klaus von Heusinger heads two projects.

Ideally, genes, proteins or hormone changes can serve as biomarkers of diseases. In mental disorders, they are often not so specific. However, it has already been shown for other diseases that certain speech errors are typical of mental disorders or malfunctions. “Since the nineteenth century, there have been studies on aphasia, especially in patients who had suffered brain injury on the battlefield,” said Klaus von Heusinger. Sufferers from this disease can either no longer speak at all, or what they say makes no sense. This is also typical after a stroke. “The doctors concluded that certain cognitive abilities are controlled by certain areas of the brain and their interaction,” the linguist added. In recent times, the methods and research questions have become more differentiated, but the basic model still applies to many mental disorders. Studies on Alzheimer’s disease, for example, show typical errors in the use of pronouns that indicate the onset of the neurodegenerative disease.

Artificial intelligence and high-tech devices

The University of Cologne group is not the only one to look at psychosis from a linguistic perspective, but a special combination of methods makes this comparatively small project unique: The team collects language samples from patients in the early phase of a first-onset psychosis as well as from a control group. This focus on the early stage is important to capture the pure symptoms before a diagnosis, possible hospitalization or the further course of the disease affect the result.

These samples are then compared with so-called large language models. These models are based on a large data set of everyday language collected by artificial intelligence systems such as ChatGPT. They predict statistically probable repetitions of syntax and speech flow, and can thus identify typical characteristics of normal language. Together with the international project partner Professor Dr Wolfram Hinzen in Barcelona, the team also evaluated language samples from English, Spanish and Turkish – in accordance with the criteria of the DISCOURSE Consortium (Consortium for Research in Thought, Language and Communication in Psychosis), so that the results could be generalized. “Studies on schizophrenia have already shown that language disorders are always the same – across languages with completely different structures and syntax,” said Çokal.

In addition, the brain activities of the study participants are examined by magnetic resonance imaging (MRI): In the scanner, they hear a story whose narrative structure is distorted. This incoherence would normally cause stress for speech processing. The research team expect that the brains of healthy and psychotic participants will react differently to this dissonance, as the networks responsible for coherence are affected during the onset of psychosis.

The focus is on the so-called Default Mode Network. This network is typically idle and is suppressed as soon as we focus on specific tasks. In psychosis, changes are suspected especially here. In addition, participants are presented with different language sequences in the MRI scanner, allowing the researchers to see the brain’s response to changes in sentence structure. Outside the scanner, participants also solve tasks and answer questions. This enables the researchers to assess changes in free or controlled speech production in conditions of psychosis. “Here, too, we and our project partners set people in different countries the same tasks in their respective languages,” said Derya Çokal.

Better predictions

Among other things, the research project is an example of how artificial intelligence – here in the form of large language models – can be used in medicine. For Joseph Kambeitz and his colleagues, this is significant not only for diagnosis, but also for predicting the further course of the condition. “Prognoses are much more difficult,” said the psychiatrist. “The textbooks say that one third of patients suffer only one psychotic episode, one third have recurring episodes, and one third become chronically ill.” However, it is impossible to predict into which group an individual patient will fall – a state of limbo that can be devastating for them. He hopes that at some point in the future, AI will also make more precise predictions possible.

For now the team is working to develop a better method for early diagnosis, allowing for medical intervention before a full-blown psychotic episode. This would minimize further damage to patients. After all, the disease often goes hand in hand with alienation, job loss and social isolation. According to Derya Çokal, a diagnostic procedure in psychiatry based on linguistic findings would be “a significant advance in public mental health”.