Sch-net

Sch-net detects schizophrenia from speech by applying a deep convolutional neural network to identify impaired speech patterns associated with negative symptoms.

Key Features:

• End-to-End Learning: Automates feature extraction from raw speech signals without manual feature engineering.
• Convolutional Neural Network (CNN) Architecture: Implements a CNN-based framework as the core classification model.
• Skip Connections: Integrates low- and high-level features to enrich representations for classification.
• Convolutional Block Attention Module (CBAM): Applies learnable attention weights to emphasize features relevant for detection.
• Performance on Schizophrenia Dataset: Achieved 97.68% accuracy on a dataset comprising 28 patients with schizophrenia and 28 healthy controls.
• Validation on LANNA Database for SLI: Attained 99.52% accuracy on the LANNA children's speech database for specific language impairment (SLI) detection.
• Generalization Capability: Demonstrated applicability across different datasets and conditions, including SLI detection.
• Experimental Validation: Evaluated through ablation experiments and comparative analyses against traditional feature-engineered models.

Scientific Applications:

• Schizophrenia diagnosis (negative symptoms): Assists objective assessment of speech impairments to support diagnosis of schizophrenia with predominant negative symptoms.
• Specific Language Impairment (SLI) detection: Applied to children's speech in the LANNA database for SLI classification.
• Automated speech-based diagnostic assessment: Provides automated extraction and selection of speech features for clinical research and diagnostic support.

Methodology:

Trains a deep convolutional neural network enhanced with skip connections and a Convolutional Block Attention Module (CBAM) using end-to-end learning for automatic feature extraction, with validation via ablation experiments and comparative analyses.