A PSO-based rule extractor for medical diagnosis.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Computational Neuroscience

Background:

• Conventional neural networks pose interpretability challenges in medicine due to opaque, numerically encoded knowledge.
• Previous Hyper-Rectangular Composite Neural Networks (HRCNNs) offered rule-based interpretability but suffered from ineffective rules and poor generalization.
• A need exists for enhanced AI models that provide interpretable insights while maintaining high diagnostic accuracy in medical applications.

Purpose of the Study:

• To propose a Particle Swarm Optimization (PSO)-based Fuzzy Hyper-Rectangular Composite Neural Network (PFHRCNN).
• To refine HRCNNs by optimizing and trimming ineffective If-Then rules to improve generalization.
• To enhance the interpretability and recognition performance of AI models in medical diagnostics.

Main Methods:

• Development of the PFHRCNN model integrating PSO for rule optimization.
• Application of PSO to prune and refine the If-Then rules generated by a trained HRCNN.
• Validation of the PFHRCNN model on benchmark medical datasets.

Main Results:

• The PFHRCNN effectively trims ineffective rules from HRCNNs, enhancing generalization.
• The proposed model demonstrates maintained or improved recognition performance compared to standard HRCNNs.
• Successful application across diverse medical datasets including liver disorders, breast cancer, and Parkinson's disease.

Conclusions:

• The PFHRCNN offers a significant advancement in interpretable AI for medical applications.
• This approach addresses the generalization limitations of previous HRCNN models.
• PFHRCNN shows promise for reliable and interpretable AI-driven diagnostics in healthcare.