Improved Digital Therapy for Developmental Pediatrics Using Domain-Specific Artificial Intelligence: Machine Learning...

Area of Science:

• Computer Vision
• Machine Learning
• Developmental Psychology

Background:

• Automated emotion classification can assist individuals with recognizing emotions, particularly children with developmental conditions like autism.
• Existing computer vision models for emotion recognition underperform on child faces due to training on adult data.

Purpose of the Study:

• To gamify the collection and labeling of child emotion-rich images.
• To enhance the performance of automatic child emotion recognition models for digital healthcare applications.

Main Methods:

• Utilized a therapeutic smartphone game (GuessWhat) for secure video data collection of children's emotions.
• Developed a web interface (HollywoodSquares) for gamified human labeling of emotion data.
• Created an expanded pediatric emotion-centric database and trained a convolutional neural network (CNN) classifier.

Main Results:

• The CNN classifier achieved 66.9% balanced accuracy on the CAFE dataset and 79.1% on CAFE Subset A.
• Performance on CAFE Subset A was over 10% higher than previous classifiers.
• The expanded database was over 30 times larger than existing public pediatric emotion datasets.

Conclusions:

• Gamified mobile games for pediatric therapies can generate large datasets for training advanced classifiers.
• This approach supports precision health efforts by improving emotion recognition in children.
• The developed methods enable the creation of state-of-the-art emotion recognition models for pediatric applications.