Towards Unsupervised Subject-Independent Speech-Based Relapse Detection in Patients with Psychosis using Variational Autoencoders

Generative models, such as Variational Autoen-coders, are being increasingly utilized for various acoustic modeling tasks, such as anomaly detection from audio signals. Motivated by this, in this work we propose a Convolutional Variational Autoencoder (CVAE), in order to detect and predict the appea...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2022 30th European Signal Processing Conference (EUSIPCO) S. 175 - 179
Hauptverfasser: Garoufis, C., Zlatintsi, A., Filntisis, P. P., Efthymiou, N., Kalisperakis, E., Karantinos, T., Garyfalli, V., Lazaridi, M., Smyrnis, N., Maragos, P.
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: EUSIPCO 29.08.2022
Schlagworte:
Aggregates
Anomaly Detection
Convolution
Europe
Interviews
Mental disorders
Predictive models
Protocols
Psychotic Disorders
Relapse Prediction
Spontaneous Speech
Vari-ational Autoencoder
ISSN:2076-1465
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Generative models, such as Variational Autoen-coders, are being increasingly utilized for various acoustic modeling tasks, such as anomaly detection from audio signals. Motivated by this, in this work we propose a Convolutional Variational Autoencoder (CVAE), in order to detect and predict the appearance of relapses in patients with psychotic disorders, such as schizophrenia and bipolar disorder. The proposed system utilizes speech segments of patients, isolated from interviews conducted with their clinicians containing spontaneous speech, and represented as log-mel spectrograms. The results from the analysis of each segment are then aggregated in a per-interview basis. We explore the performance of our system in both a personalized and a universal (patient-independent) setup. Evaluation of our method in data from 13 patients and 375 interviews, with a total duration of 30509 sec of isolated speech, indicate that the CVAE achieves similar results to a Convolutional Autoencoder (CAE) baseline in a personalized setup. Further-more, the proposed model significantly outperforms the CAE baseline when considering a universal relapse detection setup.
ISSN:2076-1465
DOI:10.23919/EUSIPCO55093.2022.9909841