American developers have learned to predict the development of psychosis in a patient. To do this, they used an automatic speech analysis algorithm, which included two metrics: an assessment of the semantic fullness of speech and the use of words related to sounds. It turned out that those patients who had a psychosis soon after the interview were characterized by poor vocabulary and frequent use of words like “sound” and “whisper”. The accuracy of such a diagnosis, as reported in an article published in Nature, was 93 percent.
Many mental disorders (including, for example, schizophrenia and bipolar disorder) are accompanied by psychosis – an altered state of mind, in which communication with the objective external world is lost, and a person begins to see, hear and feel what is not happening. As such, there is no treatment for psychosis: depending on the accompanying diagnosis, antipsychotics or therapy can help, but they rarely get rid of the condition forever.
At the same time, psychosis does not necessarily arise as a result of the development of a disease and is included in its pathogenesis, or it may even be the primary syndrome. Psychosis in a mild form can be controlled, due to which it is possible to get rid of the subsequent deterioration of the mental state. To facilitate the diagnosis of psychosis, automatic methods are being actively developed: many of them are aimed at determining the likelihood of developing psychosis from a person’s speech, which changes due to a change in mental state. In January last year, for example, developers managed using automatic speech analysis to identify a number of factors that signal the likely development of psychosis (violation of semantic connectivity, reducing the length of sentences, the absence of indicative pronouns) and achieve a prediction accuracy of 83 percent.
A new method was proposed by scientists under the leadership of Phillip Wolff of Emory University (Atlanta, USA). Unlike their predecessors, they focused on semantics, or rather, on two aspects of the speech of people who are responsible for the meaning of what was said and can signal the risk of developing psychosis.
The first aspect is the semantic poverty of speech: the presence of obsessive thoughts, lack of connection with reality can lead to the fact that a person’s speech is impoverished, the dictionary used is greatly reduced, certain repetitive patterns appear. The second aspect is directly related to one of the most prominent symptoms of psychosis – auditory hallucinations: their appearance, of course, does not go unnoticed, because of which a person experiencing them may more often talk about what he hears, using words associated with sounds and by voices.
In their study, the researchers used clippings from interviews of 40 people who had been seen by a university psychiatrist for two years: 12 of them had a psychosis during observation (and after the interview). The researchers used the word2vec algorithm, with which each word is represented as a vector in multidimensional space; in it, the proximity of two vectors to each other indicates the semantic proximity of two words.
In order to determine a certain field of the semantic norm of ordinary conversations and, accordingly, to preliminarily train the model, the scientists compiled a vector representation of the words from the posts on Reddit of 30 thousand users. After the vector representation of each word, the algorithm makes a vector representation of each sentence, summing up the values of the vectors denoting the words in it, on the basis of which the main idea-sentence is distinguished, which is also represented as a vector corresponding to a certain word in the sentence. For example, the unifying and basic meaning of the sentences “I want an apple”, “he wants to live” and “they want freedom” is the meaning of need for something, which is expressed in the verb “want”. The semantic fullness of the sentences used is calculated by calculating the proximity of the main word in it to everyone else.
On the basis of both models, a classifier was then trained, which, based on the semantic fullness index and the presence of words associated with sounds, determines whether a person falls into a group of people who have developed psychosis. It turned out that he and another indicator correlate with whether a person is diagnosed with psychosis. A joint analysis of patient interviews using two metrics made it possible to correctly predict the appearance of psychosis with an accuracy of 93 percent.
Despite the fact that the algorithm used works on the basis of speech analysis – just as the diagnosis works by a psychiatrist – the new method can be much more efficient: an objective statistical analysis of semantic parameters of speech is based on working with more material. It also means that two separate possible markers for the development of psychosis — the use of words related to sounds and the poverty of the dictionary — cannot be used in diagnosis for psychotherapeutic practice.
Psychosis is not the only thing that people can now diagnose using automatic speech analysis. In the fall, American developers, for example, presented an algorithm that, by analyzing speech, determines depression with an accuracy of 77 percent.