Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust and Explainable Depression Identification from Speech Using Vowel-Based Ensemble Learning Approaches

Kexin Feng, Theodora Chaspari

TL;DR

Pre-trained vowel-based embeddings, that integrate semantically meaningful linguistic units, are used for identifying depression from speech, demonstrating robustness and reduced susceptibility to dataset mean/median values.

Abstract

This study investigates explainable machine learning algorithms for identifying depression from speech. Grounded in evidence from speech production that depression affects motor control and vowel generation, pre-trained vowel-based embeddings, that integrate semantically meaningful linguistic units, are used. Following that, an ensemble learning approach decomposes the problem into constituent parts characterized by specific depression symptoms and severity levels. Two methods are explored: a "bottom-up" approach with 8 models predicting individual Patient Health Questionnaire-8 (PHQ-8) item scores, and a "top-down" approach using a Mixture of Experts (MoE) with a router module for assessing depression severity. Both methods depict performance comparable to state-of-the-art baselines, demonstrating robustness and reduced susceptibility to dataset mean/median values. System explainability benefits are discussed highlighting their potential to assist clinicians in depression diagnosis and screening.

Robust and Explainable Depression Identification from Speech Using Vowel-Based Ensemble Learning Approaches

TL;DR

Pre-trained vowel-based embeddings, that integrate semantically meaningful linguistic units, are used for identifying depression from speech, demonstrating robustness and reduced susceptibility to dataset mean/median values.

Abstract

This study investigates explainable machine learning algorithms for identifying depression from speech. Grounded in evidence from speech production that depression affects motor control and vowel generation, pre-trained vowel-based embeddings, that integrate semantically meaningful linguistic units, are used. Following that, an ensemble learning approach decomposes the problem into constituent parts characterized by specific depression symptoms and severity levels. Two methods are explored: a "bottom-up" approach with 8 models predicting individual Patient Health Questionnaire-8 (PHQ-8) item scores, and a "top-down" approach using a Mixture of Experts (MoE) with a router module for assessing depression severity. Both methods depict performance comparable to state-of-the-art baselines, demonstrating robustness and reduced susceptibility to dataset mean/median values. System explainability benefits are discussed highlighting their potential to assist clinicians in depression diagnosis and screening.

Paper Structure

This paper contains 20 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background & Related Work
  3. The Patient Health Questionnaire-8 (PHQ-8)
  4. Ensemble Learning for Depression Identification
  5. Depression Identification from Speech
  6. Limitations of Prior Work and Study Contributions
  7. Data Description
  8. Proposed Method
  9. Vowel-Based Speech Embeddings
  10. Data Augmentation
  11. Bottom-up Ensemble System Design
  12. Top-down Ensemble System Design
  13. Results
  14. Depression Identification Performance
  15. Comparison between Actual and Estimated PHQ-8
  16. ...and 5 more sections

Figures (3)

  • Figure 1: A schematic illustration of the bottom-up system. Separate PHQ-8 items are estimated based on individual models and are further aggregated into the final PHQ-8 score.
  • Figure 2: A schematic illustration of the top-down system. Each expert is trained on a depression severity class separately. The router selects an expert (e.g., severe depression in the example) that predicts a score within the corresponding range (e.g., 20 to 24 for severe depression).
  • Figure 3: 2D scatter plot between actual and predicted PHQ-8 score from different systems. The x-axis depicts the participants of the test sample ordered in ascending order of the actual PHQ-8 score.