Analysis of Stacked and Regular Supervised Machine Learning Algorithms to Identify Corneal Astigmatism

Abstract

According to the National Library of Medicine [2], astigmatism is an eye condition that affects more than 13 percent of the global population with refractive errors, or around 195 million people. Detecting astigmatism early, especially in children, can have a positive effect on their quality of life and vision strength. In this research, our goal is to create a machine learning (ML) model with multiple algorithms in a level-1 (meta-model) stacked [3] manner using logistic regression. Individual machine learning algorithms can sometimes be accurate; however, when multiple algorithms’ capabilities are combined to create a stacked algorithm, the overall accuracy is increased to a value higher than that of any individual algorithm. First, because of there being so many different ML algorithms, we need to narrow down which ones to use in our stacked model. In this research, we use a decision tree classifier [4], a random forest classifier [5], the XGBoost algorithm [6], and the Gradient Boosting classifier [7] in our stacked model. These algorithms were picked because of their individual accuracies being higher than other supervised ML algorithms tested, like the SVM [16], the Naïve Bayes Classifier [17], the ANN [15] (artificial neural network), and more. Finding the differences between these algorithms’ scores can help us identify which algorithms are best suited to identify astigmatism and the reason behind them being successful in this specific task. Finding similarities between these successful algorithms can help us develop an even stronger stacked model. Overall, the central goal of this research is to develop a machine-learning model capable of accurately and efficiently identifying astigmatism in a medical scenario.