Технологія Побудови Моделі Для Прогнозування Максимальної Корегованої Гостроти Зору

Abstract

The paper considers the technology of building a regression model for the maximum corrected visual acuity prediction based on a real medical dataset. The dataset included 120 features for 91 patients with nonproliferative diabetic retinopathy. The technology implied three stages. The first stage involved data preprocessing which included missing values handling, anomaly detecting, features normality checking, correlations between features analyzing and multicollinearity detecting. During the missing values handling, 14 features and one row were removed from the dataset. The Pearson correlation coefficients between independent and target features were analyzed in order to select important features for a regression model. Five features with a correlation greater than 0.8 were selected. The Pearson correlation coefficients between all independent features were explored to exclude highly correlated independent features. 27 independent features were deleted after the analysis. The variance inflation factor was used to detect multicollinearity. The second stage involved building a multiple linear regression model using a filter method for feature selection. The model with three features that were the most strongly correlated