Interpretation of cluster structures in pain-related phenotype data using explainable artificial intelligence (XAI)

Area of Science:

• Pain research
• Machine learning
• Data science

Background:

• Subgrouping patients by pain characteristics is common in clinical practice.
• Computer-aided clustering is frequently used for this purpose.
• There is a growing need for transparency in computer-aided decision-making.

Purpose of the Study:

• To develop a machine learning approach for understandable interpretation of cluster structures.
• To enable transparent decision-making in subject subgrouping based on pain characteristics.
• To provide clear explanations for why individuals are placed in specific clusters.

Main Methods:

• Transformed cluster interpretation into a classification problem.
• Utilized sub-symbolic algorithms to determine pain measure importance for cluster assignment.
• Employed explainable artificial intelligence (XAI) with symbolic algorithms for understandable cluster assignment rules.
• Validated the approach using 100-fold cross-validation.

Main Results:

• Variable importance for clustering varied across different scenarios.
• Sub-symbolic classifiers achieved the highest median accuracy.
• Generalized post-hoc interpretation reduced median accuracy.
• XAI models interpreted cluster structures accurately, with a minor loss of accuracy.

Conclusions:

• Assessing variable importance is crucial for understanding cluster structures in pain research.
• XAI models enhance the interpretability of cluster structures for human understanding.
• Model selection should be tailored to specific clustering problems.
• Increased comprehensibility of clustering models may involve a trade-off with accuracy.