Diabetic prediction using ridgelet process neural network with noise contrastive estimation

Abstract

Globally, diabetes is the most prevalent disease and a major public health concern, often leading to serious complications like kidney disease, vision loss, and coronary issues. Deep learning (DL) and data mining techniques have emerged as reliable and highly effective methods for predicting diabetes, offering significant support to the medical field in accurate diagnosis. However, challenges like missing data and difficulties in handling high-dimensional data significantly impact diabetes prediction, resulting in poor performance. This research proposes a ridgelet process neural network (RPNN) with noise contrastive estimation (NCE) to mitigate the effects of noise and remove irrelevant features, enhancing prediction accuracy. In the pre-processing, standard scaling is involved to ensure data categorization, and missing data is handled using the interquartile range (IQR) method to improve prediction performance. The feature selection using partial least angle regression (PLAR) is utilized to handle high-dimensional data and minimize multicollinearity among features efficiently. The RPNN using a ridgelet transformer maps the data into higher-dimensional space, enabling capture of complex features and effective handling of linear patterns. The proposed method achieves better accuracy of 99.89%, 98.63%, and 99.37% on datasets like diabetes, PIMA, and diabetes 130-US when compared with existing methods like guided artificial neural network (G-ANN).