Identifying Infant Body Position from Inertial Sensors with Machine Learning: Which Parameters Matter?

Area of Science:

• Biomedical Engineering
• Infant Motor Development
• Machine Learning Applications

Background:

• Accurate infant body position classification is vital for monitoring motor development and enabling early intervention.
• Manual video analysis is labor-intensive, prompting the use of Inertial Motion Unit (IMU) sensors for automated classification.
• Supervised machine learning with hand-crafted features is a common approach for IMU data classification.

Purpose of the Study:

• To compare the performance of CatBoost Classifier against Random Forest Classifier for classifying infant body positions using IMU data.
• To identify the most important sensor data features for accurate body position classification in infants aged 4-12 months.

Main Methods:

• Utilized a longitudinal dataset of IMU recordings from infants (4-12 months) during three play activities.
• Employed supervised machine learning, specifically CatBoost and Random Forest classifiers.
• Conducted data ablation experiments and SHAP value analysis to assess feature importance.

Main Results:

• The CatBoost Classifier significantly outperformed the Random Forest Classifier.
• Achieved high classification accuracies: Supine (97.7%), Sitting (93.5%), and Prone (89.9%).
• Accelerometer and magnetometer data, particularly their statistical features, were identified as critical for classification accuracy.

Conclusions:

• The CatBoost model offers a superior, automated method for classifying infant body positions from IMU data.
• This approach can aid in the early detection of developmental issues and facilitate timely interventions.
• Feature importance analysis highlights the value of specific sensor data characteristics for robust classification.