On the performance and interpretability of Mamdani and Takagi-Sugeno-Kang based neuro-fuzzy systems for medical diagnosis

Abstract

PurposeNeuro-fuzzy systems aim to combine the benefits of artificial neural networks and fuzzy inference systems: a neural network can learn patterns from data and achieves high performance, whereas a fuzzy system matches inputs and outputs using linguistic and interpretable rules. The combination of these two techniques yields models that can both perform well and provide interpretability in a fuzzy linguistic manner. DesignIn this paper, the performance and interpretability of five neuro-fuzzy classifiers were evaluated (three Takagi-Sugeno-Kang (TSK) classifiers: adaptive neuro-fuzzy inference system (ANFIS), dynamic evolving neuro-fuzzy system (DENFIS), self-organizing fuzzy neural network (SOFNN), and two Mamdani classifiers: hybrid fuzzy inference system (HyFIS) and neuro-fuzzy classifier (NEFCLASS)). All the empirical evaluations were over four benchmark medical datasets (Wisconsin breast cancer dataset, SPECTF heart dataset, Parkinsons dataset, and diabetic retinopathy Debrecen dataset), and used five performance criteria (accuracy, precision, recall, f score, and training time) and two interpretability criteria (number of rules and number of membership functions). FindingsResults showed that the TSK-based self-organizing fuzzy neural network classifier, in general, outperformed the others. In terms of interpretability, DENFIS and NEFCLASS were the best Takagi-Sugeno-Kang and Mamdani classifiers respectively. The findings also suggested that three classifiers: DENFIS, SOFNN, and NEFCLASS achieved a good performance-interpretability tradeoff. OriginalityTo the best of our knowledge, no study has compared the neuro-fuzzy techniques presented in this paper in terms of performance and interpretability in the medical domain.