Classifying Depression Symptom Severity - Assessment of Speech Representations in Personalized and Generalized Machine Learning Models

Aug 3, 2026

Upload by Finn

Urgent demand exists for improved management and treatment of Major Depressive Disorder (MDD), and speech is a promising digital marker captured via machine learning. Many studies rely on cross-sectional data, leaving personalization underexplored for building more robust, reliable models. This work evaluates combinations of speech representations and machine learning models under both personalized and generalized settings for two-class depression-severity classification. Longitudinal results show personalization advantages, with the best model using self-supervised features plus a CNN-LSTM back-end.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages5

Reading Time13 min

AccessFree to download

Abstract

Introduction

Experimental Corpus

INTERSPEECH 2023 20-24 August 2023, Dublin, Ireland Classifying depression symptom severity: Assessment of speech representations in personalized and generalized machine learning models. Edward L. Campbell 1 ,2, Judith Dineley2, Pauline Conde2, Faith Matcham2 ,3, Katie M. White2, Carolin Oetzmann2, Sara Simblett2, Stuart Bruce4, Amos A. Folarin2 ,5 ,6, Til Wykes2 ,5, Srinivasan Vairavan7, Richard J.B. Dobson2 ,6, Laura Docı´o-Ferna´ndez 1, Carmen Garcı´a-Mateo 1, VaibhavA. Narayan8, Matthew Hotopf2 ,5 , Nicholas Cummins 2, The RADAR-CNS Consortium9 1 GTM research group, AtlanTTic Research Center, University of Vigo, Spain 2 Institute of Psychiatry, Psychology and Neuroscience, King’s College London, UK 3 School of Psychology, University of Sussex, Falmer, UK 4 RADAR-CNS Patient Advisory Board, King’s College London, UK 5 NIHR Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London, UK 6 Institute of Health Informatics, University College London, UK 7 Janssen Research and Development LLC, Titusville, NJ, United States 8 Davos Alzheimer’s Collaborative [9](9 www.radar-cns.org)[ www.radar-cns.org](9 www.radar-cns.org) [ecampbell@gts.uvigo.es](ecampbell@gts.uvigo.es) , [nick.cummins@kcl.ac.uk](nick.cummins@kcl.ac.uk) Abstract There is an urgent need for new methods that improve the management and treatment of Major Depressive Disorder (MDD) . Speech has long been regarded as a promising digital marker in this regard, with many works highlighting that speech changes associated with MDD can be captured through machine learning models. Typically, findings are based on cross-sectional data, with little work exploring the advantages of personalization in building more robust and reliable models. This work assesses the strengths of different combinations of speech representations and machine learning models, in personalized and generalized settings in a two-class depression severity classification paradigm. Key results on a longitudinal dataset highlight the benefits of personalization. Our strongest performing model set-up utilized self-supervised learning features and convolutional neural network (CNN) and long short-term memory (LSTM) back-end. Index Terms: Major depressive disorder, personalization, selfsupervised learning, remote monitoring technologies 1. Introduction Due to the prevalence and high socioeconomic costs associated with Major Depressive Disorder, several digital health initiatives have started to explore new ways to improve the management and treatment of MDD [1, 2, 3] . Speech is uniquely placed as a health signal in such projects due to its pyramidal structure of information [4, 5] . This structure runs from acoustic information at the lowest level, then onto prosodic, phonetic and finally conversational at the highest level [4, 5] . The acoustic, phonetic and prosodic levels have been of particular interest in speech-based depression detection, with a rich set of supporting literature strengthening the case for speech to be considered a valuable marker of depression [6, 7] . The majority of machine learning works in this field have focused on developing generalizable machine learning models to detect the presence or absence of depression in speech samples from cross-sectional datasets [8, 9] . However, the complexity of speech and the natural variety of human voices make robust extraction of speech patterns associated with depression a highly non-trivial task. Adding this is the ordinal nature of depression scores, meaning we cannot assume a continuous and well-behaved relationship between changes in speech features and assessment scores [10] . Given these difficulties, there is a strong case for exploring personalization to improve the performance of speech-based systems [11] . Herein, we compare the performance of different speech representations and machine learning models in generalized and personalized settings. The mai

FAQs

What problem does the study address in MDD speech-based modeling?

It targets the need for better methods to manage and treat Major Depressive Disorder and evaluates speech-based machine learning approaches beyond cross-sectional, generalized modeling.

How does the study compare personalized versus generalized machine learning settings?

It assesses multiple combinations of speech representations and machine learning models in both personalized and generalized settings using a two-class depression-severity classification paradigm.

Which model setup performed best on the longitudinal data?

The strongest performance used self-supervised learning features with a convolutional neural network (CNN) and a long short-term memory (LSTM) back-end.

Aug 3, 2026

Upload by Finn