2024.
T.12.
№ 3.
id 1623
DOI: 10.26102/2310-6018/2024.46.3.020
The materials of the article are intended for specialists in the field of machine learning for the organization of technologies for improving the quality of information perception and interpretation of measurements in the practice of making medical decisions. The article proposes a method for selecting, tuning and testing a classifier under conditions of a deficit of a priori information in the data used. This is relevant when small samples of measurements of biological objects and their systems are formed at the initial stage of scientific research, the nonlinear properties of which often lead to the failure of statistical criteria. Nevertheless, the consistency of "inconvenient" distributions should be expressed in an adequate response of the program for assisting a medical decision. Based on this, the goal is determined - the choice of a solution method from the proposed set of methods for machine tuning of feature separation. Most tuning algorithms are heuristic, where the stop of parametric estimation is based on the criteria of entropy minimization as an indirect maximization of the received information. The main task is to determine the algorithm for learning and tuning the classification regression with an explicit predictive behavior of the similarity of the statistical convergence of the minimized errors. This guarantees an increase in the quality of risk classification with a trivial increase in training instances. The peculiarity of the case under consideration lies in the duality of the nature of chronic hepatitis C (CHC) progression in children: with HIV coinfection and CHC itself. This raises the problem of finding unified conditions for metric minimization of errors in еstimation the risk of developing CHC based on machine learning methods. Data sets were studied on small samples in order to identify significant parameters for heuristic identification of the presence of risks of developing the main and concomitant diseases. In this study, several methods of shallow machine learning of linear regressions were used, mainly using heuristic solutions for probabilistic separation of features. The article selectively describes the use of some basic learning methods taking into account their features in the technological verification of risk classifiers.
Keywords:
machine learning, chronic hepatitis C, HIV coinfection, binary classifiers, lasso regression, sum of squared errors (MSE), regularization, decision Tree Classifier, ROC curve, area Under Curve (AUC)
