Accurate Neonatal Face Detection for Improved Pain Classification in the Challenging NICU Setting.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Neonatal Care

Background:

• Object detection systems often fail in complex environments like Neonatal Intensive Care Units (NICUs).
• Existing methods struggle to accurately detect facial expressions in neonates for pain assessment.
• Standard face detectors and adult-trained models are ineffective for neonatal pain expression analysis.

Purpose of the Study:

• To develop an accurate facial expression detection system for classifying neonatal pain.
• To improve pain assessment in post-surgical neonates within the NICU setting.
• To provide a tool to aid NICU nurses in predicting and mitigating severe pain.

Main Methods:

• Training a state-of-the-art "You-Only-Look-Once" (YOLO) face detection model.
• Utilizing the USF-MNPAD-I dataset of neonate faces for model training.
• Comparing the YOLO model's performance against manual pain scoring by NICU nurses.

Main Results:

• The trained YOLO model achieved 8.6% higher mean Average Precision (mAP) in pain classification.
• The model demonstrated a 21.2% greater Area under the ROC Curve (AUC) for automatic pain classification.
• Significant improvements in accuracy were observed compared to existing methods in complex NICU scenes.

Conclusions:

• A YOLO-based facial expression detection model can accurately classify neonatal pain.
• Sharing trained model weights can accelerate the development of pain management strategies.
• This technology can help predict pain onset, enabling timely therapeutic interventions for neonates.