Schizophrenia affects 21 million people globally, with 80% experiencing relapses that threaten their health and safety. Detecting these relapses would require very frequent contact with clinicians, which is neither desirable nor feasible. An accurate online relapse predictor would alert clinicians in time for early interventions. Thus, we're developing an AI monitoring system that uses spoken language processing (SLP) and natural language processing (NLP) to analyze speech recorded at home and calculate relapse risk. This tool is being validated across six languages in a longitudinal cohort and tested in a multicenter randomized trial.

Building trust with clinicians and service users is crucial for the system's success. We're ensuring this by adhering to the EU's criteria for trustworthy AI, making our tool reliable and acceptable to users and carers. TRUSTING is an ambitious project with 12 partners from 10 different countries, with the aim of creating a foundation for innovative technology that promises equitable healthcare for those at risk of relapse.

In the frame of this grant, the GraC lab is in charge of Work Package 2, which tests for the generalizability and validity of the language-based model of psychosis across languages, tasks and sociodemographic groups, using a database of patient speech and symptom scores from 7+ languages. 

Project frameworks

This overview highlights the research lines that GraC is involved in. For more detailed information about the TRUSTING project and the research conducted by our partners, please visit the dedicated TRUSTING website.

Understanding the semantic space

We focus on semantic changes in psychosis. In this research line, we have explored semantic distances in speech according to different language models, and we have found that people with schizophrenia show a shrinking semantic space. We have also looked at the geometry of the semantic space, and shown altered semantic navigation patterns in individuals with schizophrenia and depression. Lastly, we have also created a single index of semantic behavior that is related to schizophrenic symptoms.

Automatic detection

Another framework consists of developing accurate models for the detection of schizophrenia symptoms, based on features extracted from spontaneous speech. In this line, we have combined different speech tasks to enhance the accuracy of schizophrenic spectrum disorder detection.

One project we wish to highlight is the PANSS prediction project, where we want to predict specific symptoms from acoustic-prosodic variables that can be extracted automatically from speech:

• In this study, we have preprocessed a cross-sectional dataset of speech in psychosis across 7 languages, and tried 8 different machine-learning regressors, as well as a deep learning model.
• We have found that we can predict specific symptomatology, such as mannerism and posturing, conceptual disorganization, and unusual thought content.

Psychometric properties

We aim to test for psychometric properties of NLP variables, across feature domains including acoustic-prosodic, semantic, syntactic, and information-theoretic. So far we have been testing for the test-retest validity of such metrics, in a crowdsourced setting, assessing reliability with intraclass correlation coefficients: ICCs (collaboration with Derya Cokal and Martín Villalba, University of Cologne). So far, we have seen that acoustic-prosodic metrics have high ICCs, while semantic ICCs are very low. Similarly, in a collaboration with Yan Cong, we investigate whether computational semantic measures are stable across languages.