Automatic Movement Recognition for Evaluating the Gross Motor Development of Infants.

Area of Science:

• Pediatric neurology
• Machine learning applications in healthcare
• Infant motor development assessment

Background:

• Current infant developmental delay diagnosis relies on clinical exams, which can be influenced by unfamiliar environments.
• Infants may exhibit atypical behaviors in clinical settings, potentially misrepresenting their true developmental status.

Purpose of the Study:

• To develop and validate a video-based method for early detection of gross motor abnormalities in Taiwanese infants aged 2-6 months.
• To leverage machine learning and pose recognition for objective assessment of infant motor development.

Main Methods:

• Utilized home-recorded infant videos for analysis.
• Employed two pediatric neurologists for manual annotation of gross motor delays.
• Applied transfer learning with ViTPose, HRNet, DARK, and UDP for pose recognition.
• Trained random forest, support vector machine, logistic regression, and XGBoost models for developmental status prediction.

Main Results:

• ViTPose demonstrated superior performance in pose recognition.
• 227 kinematic, motion, and posture features were extracted, with 106 identified as significant.
• The random forest model achieved the highest performance: 0.94 F1-score, 0.98 weighted average AUC, and 94% accuracy.

Conclusions:

• Video analysis combined with machine learning offers a promising approach for early and accurate detection of gross motor delays in infants.
• Home-based video assessment can provide a more reliable reflection of infant development compared to clinical visits.
• The developed model demonstrates high efficacy in identifying infants requiring further developmental evaluation.