A Supervised Artificial Intelligence Framework for Schizophrenia Detection Using Twitter Data

Résumé fourni par la source

Detecting schizophrenia in individuals who do not actively seek clinical care remains a significant psychiatric challenge. While traditional diagnoses rely heavily on clinical interviews, the widespread use of social media offers a novel, data-rich avenue for observing behavioral and linguistic markers. However, extracting diagnostic insights for complex mental health conditions from unstructured textual data presents unique analytical difficulties. This study proposes a supervised machine learning approach to diagnose schizophrenia using natural language processing (NLP) on Twitter data. We constructed a dataset of 320,856 tweets from 328 users, consisting of 164 individuals with self-disclosed schizophrenia and 164 matched control subjects. By extracting textual and user-centric features via specialized psychological dictionaries, we evaluated the efficacy of logistic regression and random forest classifiers. The results demonstrated that the random forest model achieved a highly promising classification accuracy of 90.12%. Furthermore, profound linguistic differences were identified: users with schizophrenia frequently utilized first-person pronouns and vocabulary associated with health, religion, death, anxiety, auditory experiences, and biological processes. Conversely, the control group predominantly favored terms related to friendships, leisure, success, and time. Psychological evaluations further indicated that individuals with schizophrenia exhibited higher arousal levels but lower scores in self-control and values compared to the control group. These findings underscore the potential of NLP and machine learning techniques as supportive, non-invasive tools for early detection and monitoring of schizophrenia through digital footprints.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.