Multi-objective evolutionary algorithms for fuzzy classification in survival prediction.

摘要: Abstract Objective This paper presents a novel rule-based fuzzy classification methodology for survival/mortality prediction in severe burnt patients. Due to the ethical aspects involved this medical scenario, physicians tend not accept computer-based evaluation unless they understand why and how such recommendation is given. Therefore, any classifier model must be both accurate interpretable. Methods materials The proposed three-step process: (1) multi-objective constrained optimization of patient's data set, using Pareto-based elitist evolutionary algorithms maximize accuracy minimize complexity (number rules) classifiers, subject interpretability constraints; step produces set alternative (Pareto) classifiers; (2) linguistic labeling, which assigns label each essential (3) decision making, whereby chosen, if it satisfactory, according preferences maker. If no satisfactory maker, process starts again with different input parameter set. Results performance three algorithms, niched pre-selection algorithm, algorithm diversity reinforcement (ENORA) non-dominated sorting genetic (NSGA-II), was tested from an intensive care burn unit standard machine learning repository. results are compared hypervolume metric. Besides, have been other non-evolutionary techniques validated cross-validation technique. Our proposal improves rate obtained by (decision trees, artificial neural networks, Naive Bayes, case-based reasoning) obtaining ENORA 0.9298, specificity 0.9385, sensitivity 0.9364, 14.2 interpretable rules on average. Conclusions techniques. We also conclude that outperforms NSGA-II algorithms. Moreover, given our non-combinational based real optimization, time cost significantly reduced approaches existing literature combinational optimization.