Detecting discomfort in infants through facial expressions

Area of Science:

• Medical technology
• Artificial intelligence
• Infant health

Background:

• Early detection of infant discomfort is crucial to prevent adverse outcomes like impaired brain development.
• Untreated discomfort can lead to long-term issues in neuroendocrine and immune system responses.
• Current methods for assessing infant distress may be limited.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (CNN) system for detecting infant discomfort.
• To analyze infant facial expressions for objective discomfort assessment.
• To improve the accuracy and efficiency of infant pain monitoring.

Main Methods:

• Utilized a dataset of 55 infant facial expression videos.
• Employed a pre-trained CNN model, fine-tuned with a public facial expression dataset and infant-specific data.
• Implemented a two-fold cross-validation strategy for performance evaluation.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.96 with the pre-trained and fine-tuned CNN model.
• Demonstrated superior performance compared to models without pre-training (AUC = 0.77) and methods using handcrafted features.
• Further improved AUC to 0.98 by fusing individual frame results.

Conclusions:

• The proposed CNN-based system shows high accuracy in detecting infant discomfort from facial expressions.
• This technology holds significant potential for continuous and objective discomfort and pain monitoring in clinical environments.
• Facial expression analysis via deep learning offers a promising avenue for improving infant care and outcomes.