Infant Cry Signal Diagnostic System Using Deep Learning and Fused Features.

Area of Science:

• Medical Diagnostics
• Bioacoustics
• Machine Learning

Background:

• Infants cannot verbalize symptoms, making early diagnosis difficult.
• Infant crying is a primary communication method for distress and needs.
• Accurate diagnosis of conditions like neonatal respiratory distress syndrome (RDS) and sepsis is vital due to high mortality rates.

Purpose of the Study:

• To develop and evaluate a medical diagnostic system for interpreting infant cry audio signals (CAS).
• To improve the accuracy and classification rate of diagnosing infant pathologies using cry analysis.
• To leverage deep learning algorithms and fused audio features for enhanced diagnostic capabilities.

Main Methods:

• Utilized a dataset of labeled infant cry audio signals, including those with RDS, sepsis, and healthy cries.
• Extracted audio features: harmonic ratio (HR), Gammatone frequency cepstral coefficients (GFCCs), and spectrograms via a pre-trained CNN.
• Fused these features and applied machine learning models (RF, SVM, DNN), with a focus on deep learning for feature extraction and fusion.

Main Results:

• The system achieved a highest accuracy of 97.50% using fused spectrogram, HR, and GFCC features processed through a deep learning model.
• Feature fusion through the learning process, particularly with spectrograms, significantly improved classification compared to simple concatenation.
• The deep learning approach effectively extracted sparsely represented features, enhancing the separation between different infant pathologies.

Conclusions:

• Fusing diverse audio features, especially spectrograms, within a deep learning framework is crucial for accurate infant cry analysis.
• The proposed system demonstrates a promising, non-invasive method for the early diagnosis of critical infant medical conditions.
• This approach outperforms previous benchmarks by enabling multi-classification of pathologies using advanced feature engineering and deep learning.