Pseudo-labeling based adaptations of pain domain classifiers

Area of Science:

• Biomedical Engineering
• Pain Medicine
• Machine Learning

Background:

• Pain perception is subjective and data for long-term pain assessment is scarce compared to short-term pain.
• Existing pain recognition techniques perform well on short-term pain but struggle with long-term pain episodes.
• Accurate pain assessment requires representative data, necessitating methods for long-term pain analysis.

Purpose of the Study:

• To adapt short-term pain classifiers to the long-term pain domain.
• To address the challenge of limited labeled data in long-term pain assessment.
• To improve the accuracy of pain assessment systems for prolonged pain episodes.

Main Methods:

• Utilized pseudo-labeling techniques to transfer knowledge from short-term to long-term pain domains.
• Analyzed physiological signals from both short-term and long-term pain recordings.
• Combined physiological signals with electric and thermal pain stimulation for analysis.

Main Results:

• Augmenting training data with pseudo-labeled long-term samples proved beneficial.
• Achieved a 2.4% improvement in classification performance for electric pain (to 80.4%) using an early fusion approach.
• Achieved a 2.8% improvement in classification performance for thermal pain (to 70.0%) using an early fusion approach.

Conclusions:

• Knowledge transfer from short-term to long-term pain domains is crucial for accurate pain assessment.
• Pseudo-labeling is an effective strategy for adapting classifiers to the long-term pain domain.
• The proposed methods demonstrate significant improvements in classifying long-term pain episodes.