Optimizing the early glaucoma detection from visual fields by combining preprocessing techniques and ensemble classifier with selection strategies

Abstract

Artificial Intelligence is booming and many issues of research are being explored to improve technical performance in health systems. But also making them suitable for targeted medical practices. Their cost must also be justified by real added value for medical practitioners and patients. Extracting accurate information from datasets usually comes up against the amount of data and its distribution, which greatly affect the performance of the classifiers. Unbalanced classes or insignificant data features do not provide information for classifiers. Medical data like those of visual field (VF) most suffer from these problems. These factors limit the performance of individual classifiers. However, ensemble methods such as the bagging classifier (BC) can overcome these limitations and return good performances. BC is simple to process and very favorable to the combination with dynamic/static selection strategies (BC-DS/SS) which considerably improves its performance. By remaining sensitive to the problem of data distribution, this combination requires a fusion with pre-processing techniques such as feature selection and data rebalancing to be efficient. Thus, combining pre-processing techniques with the BC-DS/SS ensemble classifiers would allow to extract more accurate information from VF datasets. The stake of this classifier combining pre-processing techniques and ensemble methods with selection strategies named C2PEMS2 (C2 relates to Classifier Combining, PEM refers to Pre-processing and Ensemble Methods and S2 refers to Selection Strategies) consists of: (1) optimizing the performances while reducing the over-fitting, (2) saving in processing time and more importantly (3) predicting more efficiently the targeted class which often is the minority in unbalanced datasets. The experiments of our approach on VF datasets allowed to predict early glaucoma with greater efficiency compared to the state of the art.