Optimizing depression detection in clinical doctor-patient interviews using a multi-instance learning framework

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: Feb. 24, 2025

Abstract In recent years, the number of people suffering from depression has gradually increased, and early detection is great significance for well-being public. However, current methods detecting are relatively limited, typically relying on self-rating scale (SDS) interviews. These influenced by subjective or environmental factors. To improve objectivity efficiency diagnosis, deep learning techniques have been applied to field automatic (ADD), providing a more accurate objective approach. During interviews, transcribed interview data one most commonly used modalities in ADD. previous studies only utilized response texts selected question–answer pairs, resulting information redundancy loss. This paper first apply multiple instance (MIL) framework textual data, aiming overcome issues inadequate text representation ineffective extraction long texts. MIL framework, each undergoes an independent feature process, ensuring that local features fully captured. not enhances overall capability but also alleviates issue sample imbalance dataset. Additionally, this improves upon aggregation strategies introducing two hyper-parameters accommodate uncertainties sentiment. An ensemble model MT5 RoBERTa (referred as multi-MTRB) was constructed extract output confidence scores indicating presence depressive instances. Due unique design proposed method highly interpretable able identify specific sentences depressed patients, while LIME provide in-depth interpretation negative sentences. provides promising approach context patterns. We evaluated DAIC-WOZ E-DAIC datasets with excellent results. The F1 score 0.88 dataset 0.86

Language: Английский

Multi-view graph-based interview representation to improve depression level estimation

Brain Informatics, Journal Year: 2024, Volume and Issue: 11(1)

Published: June 4, 2024

Abstract Depression is a serious mental illness that affects millions worldwide and consequently has attracted considerable research interest in recent years. Within the field of automated depression estimation, most researchers focus on neural network architectures while ignoring other directions. this paper, we explore an alternate approach study impact input representations learning ability models. In particular, work with graph-based to highlight different aspects transcripts, both at interview corpus levels. We use sentence similarity graphs keyword correlation exemplify advantages graphical over sequential models for binary classification problems within estimation. Additionally, design multi-view split transcripts into question answer views order take account dialogue structure. Our experiments show benefits based encodings provide new state-of-the-art results gold standard DAIC-WOZ dataset. Further analysis establishes our method as means generating meaningful insights visual summaries can be used by medical professionals.

Language: Английский