Quantization of speech disorganization for PTSD and speech disorders detection
Résumé
Psycholinguistic literature shows that Post-Traumatic Stress Disorder (PTSD) and its specific criteria as stated in
the DSM-5 affect language. We aim to analyze these effects using automated methods while keeping
interpretability, to quantify them and their importance. However, automatic PTSD detection is an open problem,
with NLP-based approaches being even rarer. With psychiatrists often having interviews with the patient, the
problem itself of detecting PTSD from a single discourse without interaction is a complex one and gold standard
tests have only 80% accuracy in doing so. In this study, we propose an approach, focusing on speech
disorganization effects only and their association with PTSD. Our method, based on sentence similarity, used state-
of-the-art models to underline speech organization in transcribed discourse. Using such representation allowed us
to create features inferred from the way psychiatrists read texts. On the 13-November Cohorts, our method reached
accuracy comparable to human standards for our dataset, both in PTSD prediction and Criterion D (negative
alterations in cognitions and mood) detection. According to the literature, criterion D is the most related to speech
organization. We also added a second contribution to our work: a method extracting parts of texts that contain the
speech disorder phenomenon.
Origine | Fichiers éditeurs autorisés sur une archive ouverte |
---|