Machine learning approaches using correlation filters for heart failure diagnosis: a comparative study of supervised techniques

This study addressed how machine learning could be used to detect factors that influenced the probability of survival of patients with heart failure, based on a database with 12 attributes collected from 299 different patients. Along with applying correlation filters, to obtain attributes that may be more important in a certain way for the disease, further assisting in new forms of treatment, and helping to reduce costs for diagnosis. In this study, we evaluated the accuracy of eight data mining algorithms for predicting heart disease using the heart failure dataset. We implement a methodology that includes 100 simulations with 10 correlation filter variations to ensure reliable and robust results. Among the eight classification algorithms, Support Vector Machine and Random Forest provided the best accuracy (84.18%). Considering the averages for all correlation filter variations, the Random Forest algorithm had the highest average (80.07%), and the Probabilistic Neural Network had the worst performance (69.43%). Analysis of other evaluation metrics revealed that our approach using a Multilayer Perceptron with a correlation filter (0.10) was the best alternative with 83.50% accuracy. Therefore, the diagnosis of cardiac insufficiency required only four attributes: creatinine phosphokinase, serum sodium, sex, and hospitalization time. This streamlined approach not only saved time and resources but also enhanced diagnostic efficiency, unlike previous works that use all base attributes for classification. Our findings suggest that data mining techniques can be a useful tool for predicting heart disease, and the proposed method.