Analysis of 19.9 million publications from the PubMed/MEDLINE database using artificial intelligence methods: approaches to the generalizations of accumulated data and the phenomenon of “fake news

Introduction. The English-language databases PubMed/MEDLINE and Embase are valuable information resources for finding original publications in basic and clinical medicine. Currently, there are no artificial intelligence systems to evaluate the quality of these publications. Aim. Development and testing of a system for sentiment analysis (i.e. analysis of emotional modality) of biomedical publications. Materials and methods. The technique of analysis of the “Big data” of biomedical publications was formulated on the basis of the topological theory of sentiment analysis. Algorithms have been developed that allow for the classification of texts from 16 sentiment classes with 90% accuracy (manipulative speech, research without positive results, propaganda, falsification of results, negative personal attitude, aggressive text, negative emotional background, etc.). Based on the algorithms, a scale for assessing the sentiment quality of research ( β -score) is proposed. Results. Abstracts of 19.9 million publications registered in PubMed/MEDLINE over the past 50 years (1970–2019) were analyzed. It was shown that publications with low sentiment quality (the value of the β -score of the text is less than zero, which corresponds to the prevalence of manipulative and negative sentiments in the text) comprise only 18.5% (3.68 out of 19.9 million). The greatest values of the β -score were characterized by publications on sports medicine, systems biology, nutrition, on the use of applied mathematics and data mining in medicine. The rubrication of the entire array of publications by 27,840 headings (MESH-system of PubMed/MEDLINE) indicated an increase in the β -score by years (i.e., the positive dynamics of sentiment quality of the texts of publications) for 27,090 of the studied headings. The most intense positive dynamics was found for research in genetics, physiology, pharmacology, and gerontology. 249 headings with sharply negative dynamics of sentiment quality and with a pronounced increase in the manipulative sentiments characteristic of the tabloid press were highlighted. Separate assessments of international experts are presented that confirm the patterns identified. Conclusion. The proposed artificial intelligence system allows a researcher to make an effective assessment of the sentiment quality of biomedical research papers, filtering out potentially inappropriate publications disguised as “evidence-based”.