Bagged Fuzzy-Rough Nearest Neighbors (BFRNN): A Novel Ensemble Learning Algorithm for Disease Diagnosis and Prognosis Prediction

Purpose of the study is to develop a novel machine learning (ML) algorithm that can accurately predict malignant versus benign tumors. A novel ML hybrid ensemble model called "Bagged Fuzzy-Rough k-Nearest Neighbors" (BFRNN) was developed. BFRNN is an improvement over the widely used k-Nearest Neighbors algorithm due to its use of fuzzy-rough logic and an unique ensemble voting algorithm. Initially, graphical libraries were used to visualize the Wisconsin Breast Cancer biomarker dataset (WBCBD) to capture useful insights about the data. Following preprocessing of the data (e.g. encoding categorical data snd removing outliers), a small subset of the most important breast cancer biomarkers were chosen based on feature selection technique and applying breast cancer domain knowledge. The performance of BFRNN was compared with a sample of five commonly used ML classification algorithms. The criteria for the evaluation the performance of ML was based on accuracy, area under the Receiver Operating Characteristic curve, and the ability to overcome overfitting. Discussion: Among the algorithms evaluated, BFRNN was the best classifier of WBCBD achieving an average training score of 98.47% and an average testing score of 99.09%. Among the other common ML algorithms evaluated, the highest test accuracy observed was 95.1% for Random Forest, with significant overfitting. In addition, outlier removal from the dataset and Pearsons Correlation evaluation steps can be avoided for the implementation of the BFRNN algorithm. BFRNN has shown high accuracy in classifying the malignant versus benign characteristics and this algorithm could be a useful tool in disease diagnosis.