Deep Learning-Based Instance Appraisable Model (EDi Pain) for Pain Estimation via Facial Videos: A Retrospective...

Area of Science:

• Medical technology
• Computer vision
• Machine learning

Background:

• Automated pain assessment systems are scarce, despite pain being critical in healthcare.
• Current methods like the visual analog scale (VAS) rely on subjective self-reporting, which can be unreliable during triage.
• Objective, automated pain estimation requires analyzing video sequences and utilizing professional assessments.

Purpose of the Study:

• To develop and validate an automated, deep learning-based model for estimating pain intensity from facial videos.
• To address the challenge of weak labels in video data by employing flexible multiple instance learning.
• To create an instance-appraisable model capable of evaluating the significance of video segments for pain inference.

Main Methods:

• Treated short video clips as instance segments with video-level ground truth (physician-rated VAS).
• Proposed flexible multiple instance learning approaches with a specialized loss function and sampling strategy.
• Developed the EDi Pain model, an instance-appraisable system for pain intensity estimation from facial videos.

Main Results:

• The EDi Pain model achieved a mean absolute error (MAE) of 1.85 and a Pearson correlation coefficient (PCC) of 0.63 on the UNBC-McMaster dataset.
• Validation on a prospective dataset of 931 patients yielded an MAE of 1.48 and a PCC of 0.22.
• The model demonstrated competitive performance in video-level pain intensity estimation.

Conclusions:

• A novel deep learning-based, instance-appraisable model (EDi Pain) was developed and validated for pain intensity estimation using facial videos.
• The EDi Pain model shows potential for real-time clinical applications, providing a more objective and dynamic pain assessment method.
• This approach offers a promising alternative to subjective pain reporting in healthcare settings.