Feature Reduction for Interpretability of Neuro-Fuzzy Classifier

Abstract

The growing use of machine learning in healthcare demands models that not only predict well but clearly explain their reasoning. Neuro-fuzzy classifiers combine neural networks and fuzzy logic for accurate yet interpretable modeling. However, high-dimensional medical data challenges comprehension. This study presents techniques to distill neuro-fuzzy classifiers down to their most essential components for enhanced interpretability. Recursive feature elimination is employed, iteratively removing the least contributory features according to classifier weights. On a breast cancer prognosis task, these methods significantly reduce features to the minimal set necessary for strong accuracy. The pruned neuro-fuzzy classifier demonstrates competitive prediction with increased transparency through an intuitive rule-based structure relying on only the most critical inputs. Feature reduction thus holds promise for optimizing the balance between predictive capacity and comprehensive explanations in medicine. However, challenges remain in scaling these methods to even higher-dimensional data like medical images while retaining accuracy. Future work should explore the effects of different criteria for ranking feature importance beyond classifier weights. Overall, this study provides an initial demonstration of how feature reduction techniques can produce simplified neuro-fuzzy classifiers more amenable to clinical implementation, but continued research is needed to extend these approaches to complex real-world medical data. Distilling machine learning models down to their most essential components will be critical for trust and adoption of artificial intelligence in high-stakes healthcare applications.