Balancing Accuracy and Interpretability through Neuro-Fuzzy Models for Cardiovascular Risk Assessment

Machine learning algorithms have been proven to be effective in supporting medical decisions. However, often the resulting models are difficult to understand by clinicians. Fuzzy Inference Systems are effective in returning interpretable models that can be easily described by IF-THEN rules expressed in natural language. Yet, conventional fuzzy systems require human expertise to design the fuzzy rule base. Neuro-fuzzy systems are data-driven algorithms, that are able to learn fuzzy rule-based models from data, thus avoiding the time consuming task of manually designing the rules. In this work we propose the use of neuro-fuzzy systems to learn predictive models from data in form of fuzzy rules that are suitable to support decision making for cardiovascular risk assessment. Our results show the effectiveness of these models in returning accurate predictions while preserving the interpretability. Furthermore, our results indicate that the models derived automatically are more accurate than the models defined manually, with the best accuracy of 0.91.