Extraction of Key Features and Enhanced Prediction Framework of Breast Cancer Occurrence

In this work, feature selection is implemented in combination with various classification models to predict the occurrence of breast cancer from morphological characteristics of the cell nuclei from Breast Cancer Wisconsin (Diagnostic). This proposed method enables recognizing the critical features by eliminating the redundant ones. By doing this, issues such as overfitting and reducing multi-collinearity leading are mitigated to confounding effects induced by having too many features-consequently, more certainty in estimating the parameters of the machine learning model. The Breast Cancer Wisconsin (Diagnostic) data contains 30 features from 10 traits of the cell nuclei, acquired from a digitized image utilizing a fine needle aspirate of breast tissue. Here, it is found that using the L1 regularization (L1) features selection method combined with the Support Vector Classifier (SVC) delivers the best prediction model in terms of both recall rate and accuracy compared to all our other proposed combinations of feature selection and classifier. The high overall accuracy of 97% has been achieved. Moreover, our proposed scheme has a 97% recall rate for the malignant class of cancer using only 12 key features out of 30 features. Thus, this technique could assist the pathologist in rapidly determining malignant cases as only 13 features need to be evaluated instead of 30 with a minimal false-negative rate.