An Attentive Convolutional Neural-Fuzzy Classifier With Interpretability for Classification

Abstract

The integration of deep learning into clinical decision support systems offers immense potential for diagnostic accuracy, yet the ”black-box” nature of many models impedes their adoption in high-stakes medical environments where transparency and trust are paramount. Conventional approaches often fail to provide clear, actionable insights into their predictive reasoning, a critical requirement for clinical utility. To address this challenge, this paper introduces the Attentive Convolutional Neuro-Fuzzy Classifier (AConNFC), a novel deep neuro-fuzzy architecture that synergistically combines a one-dimensional convolutional neural network (CNN) with an attention mechanism and a neuro-fuzzy inference system. This hybrid model is specifically designed for tabular medical data, leveraging the convolutional attention framework to dynamically identify and weight the most salient diagnostic features. These highlighted features subsequently populate a set of human-interpretable fuzzy rules, generating a dual-layered explanation for each prediction. Validated on Fetal Cardiotocography classification tasks, AConNFC demonstrates superior predictive performance while offering unparalleled interpretability. The model’s ability to produce explicit, rule-based explanations for its decisions marks a significant advancement in building trustworthy and auditable AI for medical diagnosis. This work underscores the potential of AConNFC to serve as a robust and transparent solution, fostering greater confidence and collaboration between clinicians and intelligent systems in critical decision-making processes.