ANALYTICAL REVIEW OF PUBLICATIONS ON MACHINE LEARNING METHODS IN ONCOLOGY AND APPROACH TO EVALUATING THEIR QUALITY

Abstract

The work includes an analytical review of publications on machine learning methods in oncology and an approach to evaluating their quality. An analysis of publications by year was conducted in the Web of Science and Scopus bibliometric databases. The highest number of authors, the number of publications among universities, the number of countries, and publication categories in the Scopus bibliometric database on machine learning methods in oncology are presented. A multifactor regression prediction model for bone tissue density in oncological pathology predicting four severity grades of the studied disease course was proposed. This model included the following factors with corresponding weights: gender (2.1), age (0.06), stage (0.9), absence/presence of B-symptoms (A/B) (0.9), international prognostic index (IPI-NCCN) ( 1.1), body mass index (BMI) (-0.2), number of chemotherapy courses (0.9), Charlson comorbidity index (СCI) (0.3), bone mineral density after completion of chemotherapy (HU C) (-0.08), β-2-microglobulin (B2M) level (0.0007), lactate dehydrogenase (LDH) (0.006), body surface area (BSA) (-3.3). To assess the level of confidence in the proposed model for predicting bone density disorders in oncological pathology, ROC analysis was performed to obtain the corresponding curves and the area under them was estimated. A conclusion was made about the quality of the classification and the sensitivity, specificity, prognostic value of positive and negative results, the ratio of the probability of positive and negative results, as well as the accuracy of the classification were determined. For each of the four degrees of severity of violations (1C, 2C, 3C, 4C), it is necessary to carry out appropriate calculations, the matrices of inconsistencies for which are given in four tables. Sensitivity was calculated for 1C (98.8%), 2C (97.5%), 3C (95.2%) and 4C (98.5%); specificity for 1C (90.4%), 2C (83.3%), 3C (90.9%) and 4C (95%); predictive value of a positive result for 1C (97.6%), 2C (95.2%), 3C (97.5%) and 4C (97%); predictive value of a negative result for 1C (95%), 2C (90.9%), 3C (83.3%) and 4C (97.4%); accuracy for 1C (97.1%), 2C (97.1%), 3C (97.1%), and 4C (97.1%). According to the results of the analysis of ROC curves, a high level of classification of 1C (AUC=0.869), 3C (AUC=0.869) and 4C (AUC=0.869) was established. The average level of classification of bone density disorders according to 2C (AUC=0.758).