A Positive-Confidence based approach to Classifying Imbalanced data: A Case Study on Hepatitis |Biomedgrid

Mining Imbalanced data is a challenge especially with clinical datasets where it is essential to obtain maximum possible prediction accuracy on the minority (diseased class which is the category of greatest interest) without impacting negatively on predictions on the majority class. A false positive that wrongly signals the presence of disease has as much importance as a false negative diagnosis in such scenarios. Apart from high classification accuracy, models need to be interpretable as well in order to enhance clinician's confidence in them. We propose a positive confidence-based modelling approach which uses data features to assign a confidence value to each sample that signifies the probability that a given data sample belongs to the positive class. We aim to predict the outcome using the minimum number of data features in order to maximize model interpretability. We use the publicly available Hepatitis dataset to test our approach and the results are promising. We obtain an F2 Score of 0.84 which is significantly higher than that obtained from widely used classifiers.