Applying NLP to Mental Health Diagnosis

In a relatively recent field of study, natural language processing (NLP) is being used to help diagnose mental health conditions, such as schizophrenia, by analyzing patients’ speech. In a presentation at the Future of Text Analytics summit, Kfir Bar, Ph.D., Chief Scientist at Babel Street, highlighted the findings of researchers working with Hebrew-speaking adults with mental illnesses.

People with mental illnesses often display disordered speech patterns with distinct characteristics. By applying computational tools that process human language to detect these patterns, clinicians may be aided in their diagnoses and subsequent treatment plans.

With major thought disorders, like depression and schizophrenia, certain categories of words are prevalent. The Linguistic Inquiry and Word Count (LIWC) contains around 2,300 words in categories such as body-related, affect-related, and with positive and negative connotations. Analyzing the frequency of LIWC words in spoken text can be an indicator of someone’s mental state. A usage report shows the breakdown of words into categories. Diagnosticians can use this information to see correlations between specific signatures of word usage and some mental health disorders.

Another type of analysis focuses on the complexity of speech, since some mental illnesses present with poverty of speech. Machine learning tools are able to tag the parts of speech, and are also able to perform semantic analysis for grammar, syntax and word usage.

In his group’s research, Bar and his colleagues looked at some specific speech disturbances with a correlation to schizophrenia. These include:

• Derailment – when a speaker jumps between topics as they move away from the main topic of the conversation
• Single word generation problems – when a speaker changes a known word or makes one up, but still uses it in context so that it appears to make sense
• Incoherence, or “word salad” – when a speaker streams forth illogical or incomprehensible speech filled with semantic and grammatical errors

In conducting their study, the team collected speech samples from patients who were diagnosed with schizophrenia, and from a control group. The patients and control subjects were asked 18 questions, out of which 14 were thematic-apperception-test (TAT) pictures that participants were requested to describe, followed by four questions that required the participant to share some personal thoughts and emotions.

As part of the research, the team processed the speech samples with NLP tools in order to train a machine learning model to detect schizophrenia. A scoring system was employed to determine the degree of derailment or coherence in the speech. When focusing only on nouns, verbs, adjectives, and adverbs, the difference in scores between patient and control groups becomes statistically significant. In fact, the machine model was 81.5% accurate in detecting schizophrenia from the speech samples.

While the work is ongoing, these findings are encouraging for the use of NLP tools in mental health. A diagnostic tool that is tuned to detect subtleties in speech patterns could have more far-reaching uses than diagnosing schizophrenia, and could possibly aid in the detection of other mental and neurological illnesses.

The published work of Bar and colleagues was part of the proceedings of the Sixth Workshop on Computational Linguistics and Clinical Psychology.