Explainable AI (XAI) for Neonatal Pain Assessment via Influence Function Modification.

Area of Science:

• Medical informatics
• Machine learning explainability
• Computational methods

Background:

• Machine learning (ML) models are vital in healthcare, but their complexity necessitates explainability.
• Influence functions are key for understanding black-box ML models by assessing training data impact.
• Existing methods show influential data points are semantically similar to test data.

Purpose of the Study:

• To propose a novel, modified influence function for more precise ML model explainability.
• To enhance the identification of influential training samples in medical applications.

Main Methods:

• Developed a modified influence function incorporating a data similarity weighting factor.
• Utilized cosine similarity, Euclidean distance, and structural similarity index for weighting.
• Evaluated the method on a neonatal pain assessment ML model.

Main Results:

• The modified influence function demonstrated superior performance in identifying influential training points.
• Achieved excellent accuracy in pinpointing key data samples compared to baseline methods.
• Successfully elucidated the decision-making process of the neonatal pain assessment model.

Conclusions:

• The proposed modified influence function significantly improves the explainability of complex ML models.
• This approach offers a more precise way to identify critical training data influencing predictions.
• Effective for enhancing transparency in medical ML applications.