A Novel Ensemble of Support Vector Machines for Improving Medical Data Classification

Abstract

In recent years, the increasing volume and availability of healthcare and biomedical data are opening up new opportunities for computational methods to enhance healthcare in many hospitals. Medical data classification is regarded as the challenging task to develop intelligent medical decision support systems in hospitals. In this paper, the ensemble approaches based on support vector machines are proposed for classifying medical data. This research’s key contribution is that the ensemble multiple support vector machines use the function kernel in the style of gradient boosting and bagging to produce a more accurate fusion model than the mono-modality models. Extensive experiments have been conducted on forty benchmark medical datasets from the University of California at Irvine machine learning repository. The classification results show that there is a statistically significant difference (p-values < 0.05) between the proposed approaches and the best classification models. In addition, the empirical analysis of forty medical datasets indicated that our models can predict diseases with an accuracy rate of 82.82 and 81.76 percent without feature selection in the preprocessing data stage.

IHCP: interpretable hepatitis C prediction system based on black-box machine learning models

“…To demonstrate the superiority and robustness of our proposed model, we conducted experiments on an independent testing set. Huynh et al [ 29 ] proposed an ensemble method that obtained 83.42% accuracy and 0.8418 AUC on this dataset, and Rosly et al [ 30 ] proposed a stacking technique combined with a multilayer perceptron that obtained 86.25% accuracy. In comparison, as shown in Table 6 , our IHCP has improved in both AUC and accuracy and is more suitable as a prediction model for hepatitis C.…”

“…To validate the generalization capability of the proposed model, we conducted independent dataset tests based on the second dataset. This dataset was obtained from the study by Huynh et al [ 29 ]. It contains 155 data samples with 18 input attributes and 1 output attribute.…”