Venue: Amphithéâtre du laboratoire IMS à Talence,
et sur Zoom à l’adresse :
Defense in french
New speech biomarkers for automatic sleepiness detection
Voice is one of the most promising tools in digital medicine. In association with virtual medical companions, the estimation of symptoms based on voice features will allow both home monitoring of patients suffering from chronic neuropsychiatric diseases and access to personalized lifestyle advice for the general population. Sleepiness, occuring in many pathologies and being very prevalent both in patients suffering from chronic diseases and in the general population, is a key symptom for this approach.
The objective of the work presented in this manuscript is thus to complete the information collected by virtual assistants during the interaction of the subjects with them, by using vocal markers validated as being reliable markers of sleepiness. Our approach is the following.
First, we introduce the mechanisms of voice production and the different pathologies that can interfere with the involved muscular and neuro-muscular functions, with a focus on the methodologies used for the recording and annotation of the corpora.
Then, we attempt to establish a consensual definition of sleepiness using three reference dictionaries of the French language; two text mining approaches; and finally through an umbrella review of tools designed to measure it.
Subsequently, we present our own corpus of patients with hypersomnia, recorded at the sleep medicine center of the Bordeaux University Hospital on a reading aloud task, annotated with both subjective (questionnaires) and objective (sleep latency to the Multiple Sleep Latency Test) measures of sleepiness validated by the physicians of the University Hospital.
This corpus is then compared with other state-of-the-art corpora on voice sleepiness detection, from which we propose recommendations on the development of such corpora. Then, using a perceptual study, we validate the use of the MSLT database for the detection of sleepiness in speech.
Based on this corpus, we develop four categories of speech features, measuring two dimensions of the impact of sleepiness on speech. On the one hand, we study markers of acoustic voice quality; on the other hand, we design markers of reading quality, divided into three subcategories: reading errors made by patients, their automation through errors made by automatic speech recognition systems, and finally the durations and locations of reading pauses.
These features are validated on different forms of sleepiness (objective and subjective).
Finally, we present a methodology to train a classifier for the clinical use of these speech features for the detection of three symptoms related to sleepiness. We carry out a detailed analysis of the obtained results and of the descriptors used by the classifier. To go further, we then propose to bring the classification problem closer to the reality of clinical reasoning by classifying two syndromes derived from the previous symptoms.
Finally, in this same direction, we consider research perspectives around symptom networks, in the framework of digital medicine research on sleepiness and, in a more general way, on digital psychiatry.
– Pr. Corinne Fredouille, Univ. d’Avignon, Rapporteuse
– Pr. Isabel Trancoso, Univ. de Lisbonne, Rapporteuse
– Dr. Pierre-Alexis Geoffroy, Univ. de Paris, Rapporteur
– Dr. Véronique Delvaux, FNRS – Univ. de Mons, Examinatrice
– Dr. Guy Fagherazzi, Luxembourg Institute of Health, Examinateur
– Dr. Jean-Arthur Micoulaud-Franchi, Univ. de Bordeaux, Invité
– Dr. Jean-Luc Rouas, CNRS – LaBRI, Directeur
– Pr. Pierre Philip, Univ. de Bordeaux, Co-directeur
PhD Student – Sleepiness detection through voice – LaBRI / SANPSY
ENSEA Engineer – Multimedia signal processing