Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Node-weighted Graph Convolutional Network for Depression Detection in Transcribed Clinical Interviews

Sergio Burdisso, Esaú Villatoro-Tello, Srikanth Madikeri, Petr Motlicek

TL;DR

The paper addresses depression detection from transcribed clinical interviews by introducing an omega-GCN, a node-weighted Graph Convolutional Network that reweights self-connections to capture non-local semantics with low computational cost. It uses a two-layer inductive GCN on a word-document graph, incorporating PMI for word-word edges and PageRank for self-loops, with initial features comprising one-hot word vectors and TF-IDF document representations. Empirical results on DAIC-WOZ and E-DAIC show omega-GCN outperforms vanilla GCN and several baselines, achieving Macro-F1 scores up to 0.84 while using far fewer parameters than transformer models; qualitative analyses demonstrate interpretability via word clusters and LIWC-aligned psychological dimensions. The approach offers a practical, interpretable, and data-efficient option for AI-assisted mental health diagnostics, with future work exploring additional node types (e.g., acoustic) to further enhance performance and interpretability.

Abstract

We propose a simple approach for weighting self-connecting edges in a Graph Convolutional Network (GCN) and show its impact on depression detection from transcribed clinical interviews. To this end, we use a GCN for modeling non-consecutive and long-distance semantics to classify the transcriptions into depressed or control subjects. The proposed method aims to mitigate the limiting assumptions of locality and the equal importance of self-connections vs. edges to neighboring nodes in GCNs, while preserving attractive features such as low computational cost, data agnostic, and interpretability capabilities. We perform an exhaustive evaluation in two benchmark datasets. Results show that our approach consistently outperforms the vanilla GCN model as well as previously reported results, achieving an F1=0.84 on both datasets. Finally, a qualitative analysis illustrates the interpretability capabilities of the proposed approach and its alignment with previous findings in psychology.

Node-weighted Graph Convolutional Network for Depression Detection in Transcribed Clinical Interviews

TL;DR

The paper addresses depression detection from transcribed clinical interviews by introducing an omega-GCN, a node-weighted Graph Convolutional Network that reweights self-connections to capture non-local semantics with low computational cost. It uses a two-layer inductive GCN on a word-document graph, incorporating PMI for word-word edges and PageRank for self-loops, with initial features comprising one-hot word vectors and TF-IDF document representations. Empirical results on DAIC-WOZ and E-DAIC show omega-GCN outperforms vanilla GCN and several baselines, achieving Macro-F1 scores up to 0.84 while using far fewer parameters than transformer models; qualitative analyses demonstrate interpretability via word clusters and LIWC-aligned psychological dimensions. The approach offers a practical, interpretable, and data-efficient option for AI-assisted mental health diagnostics, with future work exploring additional node types (e.g., acoustic) to further enhance performance and interpretability.

Abstract

We propose a simple approach for weighting self-connecting edges in a Graph Convolutional Network (GCN) and show its impact on depression detection from transcribed clinical interviews. To this end, we use a GCN for modeling non-consecutive and long-distance semantics to classify the transcriptions into depressed or control subjects. The proposed method aims to mitigate the limiting assumptions of locality and the equal importance of self-connections vs. edges to neighboring nodes in GCNs, while preserving attractive features such as low computational cost, data agnostic, and interpretability capabilities. We perform an exhaustive evaluation in two benchmark datasets. Results show that our approach consistently outperforms the vanilla GCN model as well as previously reported results, achieving an F1=0.84 on both datasets. Finally, a qualitative analysis illustrates the interpretability capabilities of the proposed approach and its alignment with previous findings in psychology.
Paper Structure (8 sections, 4 equations, 3 figures, 2 tables)

This paper contains 8 sections, 4 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Graph neural network architecture
  3. Experimental setup
  4. Datasets
  5. Implementation details
  6. Results
  7. Exploring the model's interpretability
  8. Conclusions

Figures (3)

  • Figure 1: A two-layer GCN with nodes represented at three levels: initial (e.g. one-hot), $H^{(0)}$, intermediate/hidden, $H^{(1)}$, and output, $Z$, representations with the probability of each output label. Circles $\rightarrow$ document nodes & triangles $\rightarrow$ word nodes.
  • Figure 2: Node embbedings learned for DAIC-WOZ. As in Figure \ref{['fig:GCN_architecture']}, circles denote documents, triangles words, and colors denote class ([D] - depression, [C] - control). The gray rectangle in (a) indicates the zoomed region (b). Graph edges are also included.
  • Figure 3: Psychological dimensions present in the model.