TY - JOUR

T1 - Анализ 19,9 млн публикаций базы данных PubMed/MEDLINE методами искусственного интеллекта

T2 - подходы к обобщению накопленных данных и феномен “fake news”

AU - Yu, Torshin I.

AU - Gromova, O. A.

AU - Stakhovskaya, L. V.

AU - Vanchakova, N. P.

AU - Galustyan, A. N.

AU - Kobalava, Zh D.

AU - Grishina, T. R.

AU - Gromov, A. N.

AU - Ilovaiskaya, I. A.

AU - Kodentsova, V. M.

AU - Kalacheva, A. G.

AU - Limanova, O. A.

AU - Maksimov, V. A.

AU - Malyavskaya, S. I.

AU - Mozgovaya, E. V.

AU - Tapilskaya, N. I.

AU - Rudakov, K. V.

AU - Semenov, V. A.

N1 - Funding Information: This work was supported by grants from the Russian Foundation for Basic Research 19-07-00356 17-07-00935 17-07-01419 18-07-01022 18-07-00944 18-07-00929 16-07-01133. Publisher Copyright: Copyright © 2020, Farmakoekonomika. All rights reserved.

PY - 2020

Y1 - 2020

N2 - 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”.

AB - 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”.

KW - Artificial intelligence

KW - Big data analysis

KW - Evidence-based medicine

KW - Machine learning

KW - Pharmacoinformatics

KW - Publication quality assessment algorithms thematic modeling

UR - http://www.scopus.com/inward/record.url?scp=85091110828&partnerID=8YFLogxK

U2 - 10.17749/2070-4909/FARMAKOEKONOMIKA.2020.021

DO - 10.17749/2070-4909/FARMAKOEKONOMIKA.2020.021

M3 - статья

AN - SCOPUS:85091110828

VL - 13

SP - 146

EP - 163

JO - Farmakoekonomika

JF - Farmakoekonomika

SN - 2070-4909

IS - 2

ER -