Towards Explainable Infant Cry Reasoning Model.

Area of Science:

• Machine Learning
• Infant Health
• Signal Processing

Background:

• Infant crying is vital for assessing health and well-being.
• Current infant cry analysis models struggle with generalization due to voice variations and lack interpretability.

Purpose of the Study:

• To develop an explainable infant cry reasoning model (EICR) addressing generalization and interpretability challenges.
• To enhance the understanding of infant cry reasons through a novel approach.

Main Methods:

• Proposed an explainable infant cry reasoning model (EICR) using a CNN with representation consistency.
• Integrated a frequency- and energy-based fuzzy decision tree for interpretability via knowledge distillation.
• Utilized energy features from different frequency bands as node attributes in the decision tree.

Main Results:

• EICR demonstrated improved accuracy by 2.39% and F1-score by 1.48% in cross-subject experiments.
• The fuzzy decision tree provided step-by-step decision rules for distinguishing cry reasons.
• Enhanced model generalization across different infant voiceprints.

Conclusions:

• The EICR model offers a more accurate and interpretable solution for infant cry analysis.
• The explainable nature of the decision tree provides new insights into infant cry interpretation.
• This approach supports improved infant caregiving and early-warning systems.