Improving Prediction Accuracy of Breast Cancer Survivability and Diabetes Diagnosis via RBF Networks trained with EKF models

The continued reliance on machine learning algorithms and robotic devices in the medical and engineering practices has prompted the need for the accuracy prediction of such devices. It has attracted many researchers in recent years and has led to the development of various ensembles and standalone models to address prediction accuracy issues. This study was carried out to investigate the integration of EKF, RBF
networks and Ada Boost as an ensemble model to improve prediction accuracy. In this study we proposed a model termed EKF-RBFN-ADABOOST.

It uses EKF to enhance the slow training speed and to improve the effectiveness of the RBF network training parameters. AdaBoost was then applied as an ensemble meta-algorithm to generate and combine several RBFN-EKF weak classifiers to form a final strong predictor of the model. Breast cancer survivability, diabetes diagnostic, credit card payment defaults and staff absenteeism datasets used in the study were obtained from the UCI repository. The prediction accuracy of the proposed model was explored using various statistical analysis methods. During the study we also proposed and developed an ensemble logistic regression model using the breast cancer dataset.

Results are presented on the proposed model EKF-RBFN-ADABOOST, as applied to breast cancer survivability, diabetes diagnostic, credit card payment defaults and staff absenteeism predictive problems. The model outputs an accuracy of 96% when EKF-RBFN was applied as a base classifier compared to 94% when Decision Stump was applied and AdaBoost as an ensemble technique in both cases. Also, a significant performance was observed for staff absenteeism at 96 % compared with credit card payment defaults that had a performance accuracy of 85%. The ensemble logistic model outputs an accuracy of 94% when we used 70% and 30% as training and testing datasets respectively compared with accuracy of 95% prediction when we used 60% of the data for training and 40% for testing respectively.