An approach towards the response surface linearization via ANN-based cascade scheme for regression modeling in Healthcare
Ivan Izonin, Roman Tkachenko · Procedia Computer Science · 2022
The digitalization of medicine, which has become widespread in recent years, opens up new opportunities for diagnosing and monitoring the patient’s condition. A large amount of digital data allows artificial intelligence tools to solve various problems in this area. In the case of solving significantly nonlinear tasks, the accuracy of its processing has priority. Existing methods, single models, and sometimes ensemble methods do not provide sufficient prediction accuracy. Response surface linearization procedures provide the ability to improve prediction accuracy when a nonlinear machine learning algorithm or an artificial neural network does not give satisfactory results. The disadvantage of this approach is a significant increase in the duration of procedures for both training and application of the method for large data sets. However, this shortcoming can be compensated by applying the high-speed machine learning method. That is why in this paper, we propose a new cascade scheme based on the use of hybrid, non-iterative ANN, which provides both high accuracy and high speed of the training procedure. The proposed scheme is based on the algorithm developed by the authors, which provides the possibility of efficient processing of data sets of different volumes. The experimental modeling was based on a real data set from the Healthcare domain. The optimal parameters of the proposed scheme are experimentally established. High accuracy of operation of the cascade scheme based on various performance indicators is received. By comparison with other methods of this class, the high efficiency of the application of the cascade scheme is established.