Enhancing Neonatal Pain Assessment Transparency via Explanatory Training Examples Identification

Area of Science:

• Medical Informatics
• Signal Processing
• Artificial Intelligence

Background:

• Visual data occlusion is common in neonatal pain assessment.
• Audio signals offer an alternative for pain detection when visual data is unavailable.
• Current deep learning (DL) methods for audio-based pain assessment lack transparency.

Purpose of the Study:

• To develop an explainable DL approach for neonatal pain detection from audio signals.
• To improve user trust in automated pain assessment systems.
• To identify and mitigate the impact of unhelpful training data.

Main Methods:

• Visualizing neonatal audio signals as spectrogram images for pain classification.
• Implementing an instance-based approach for explaining black-box DL model decisions.
• Utilizing an influence score to analyze and remove harmful training instances.

Main Results:

• The proposed method successfully visualizes audio signals for pain classification.
• Explainability was added to DL-based pain detection, enhancing user trust.
• Harmful training instances were detected and removed, leading to a compressed dataset and improved predictions.

Conclusions:

• The approach provides a transparent and explainable method for neonatal pain assessment using audio.
• This technique can enhance clinical trust in AI-driven pain detection systems.
• The study offers a viable strategy for improving pain assessment in real-world clinical environments.