An artificial neural network (ANN) model for publication bias: a machine learning-based study on PubMed meta-analyses

Abstract

PurposeNo study has investigated the effects of different parameters on publication bias in meta-analyses using a machine learning approach. Therefore, this study aims to evaluate the impact of various factors on publication bias in meta-analyses.Design/methodology/approachAn electronic questionnaire was created according to some factors extracted from the Cochrane Handbook and AMSTAR-2 tool to identify factors affecting publication bias. Twelve experts were consulted to determine their opinion on the importance of each factor. Each component was evaluated based on its content validity ratio (CVR). In total, 616 meta-analyses comprising 1893 outcomes from PubMed that assessed the presence of publication bias in their reported outcomes were randomly selected to extract their data. The multilayer perceptron (MLP) technique was used in IBM SPSS Modeler 18.0 to construct a prediction model. 70, 15 and 15% of the data were used for the model's training, testing and validation partitions.FindingsThere was a publication bias in 968 (51.14%) outcomes. The established model had an accuracy rate of 86.1%, and all pre-selected nine variables were included in the model. The results showed that the number of databases searched was the most important predictive variable (0.26), followed by the number of searches in the grey literature (0.24), search in Medline (0.17) and advanced search with numerous operators (0.13).Practical implicationsThe results of this study can help clinical researchers minimize publication bias in their studies, leading to improved evidence-based medicine.Originality/valueTo the best of the author’s knowledge, this is the first study to model publication bias using machine learning.