Using CCA-Fused Cepstral Features in a Deep Learning-Based Cry Diagnostic System for Detecting an Ensemble of...

Area of Science:

• Medical Informatics
• Signal Processing
• Pediatrics

Background:

• Newborn cries are vital communication signals indicating health and emotional states.
• Distinguishing between healthy and pathologic infant cries is crucial for early medical intervention.

Purpose of the Study:

• To develop an automated, non-invasive Newborn Cry Diagnostic System (NCDS).
• To differentiate between healthy and pathologic newborns using cry signal analysis.
• To explore novel feature fusion techniques for enhanced diagnostic accuracy.

Main Methods:

• Extraction of Mel-frequency Cepstral Coefficients (MFCC) and Gammatone Frequency Cepstral Coefficients (GFCC) from newborn cries.
• Fusion of feature sets using Canonical Correlation Analysis (CCA).
• Classification using Support Vector Machine (SVM) and Long Short-term Memory (LSTM) models, with Bayesian and grid search for hyperparameter optimization.

Main Results:

• The CCA-fused feature set with LSTM achieved a 99.86% F-score for inspiratory cries.
• The GFCC feature set with LSTM achieved a 99.44% F-score for expiratory cries.
• Both feature sets and classifiers demonstrated high efficacy in identifying infant pathologies.

Conclusions:

• Newborn cry analysis is a valuable tool for detecting infant pathologies.
• The proposed NCDS framework demonstrates significant potential as an early diagnostic aid in clinical settings.
• Automated analysis of cry signals can improve the identification of at-risk newborns.