PubMed-indexed articles published between 2016 and 2021 were identified with the Medical Subject Heading (MeSH) terms “ophthalmology,” “radiology,” and “neurology.” These fields were selected to evaluate journal recommendations in a clinical domain compared to computer science or biomedicine. These 3 fields are often related and provide a broader scope compared to a single discipline. Titles, abstracts, author lists, and MeSH terms were collected. The journal’s ISSN for each article was identified, along with the associated journal impact factor. Articles without a specified journal ISSN or articles that could not be linked to an impact factor were excluded from the study. Journal impact factor and Eigenfactor scores were listed by the 2020 Clarivate Journal Citation Report. The journals included in the study were allocated percentile ranks based on the impact factor and Eigenfactor scores, compared with other journals that released publications in the same year. The impact factor is calculated by dividing the current year citations by the source items published in that journal during the previous 2 years. The Eigenfactor score is a weighted measure of the number of times articles from the journal published in the past 5 years has been cited in the Journal Citation Report year, eliminating self-citations. The calculation algorithm is freely available [24]. These complementary measures were used, as the impact factor reflects popularity as opposed to the Eigenfactor score that reflects prestige or trustworthiness [25].

All abstracts underwent a process involving the removal of abstract structure (eg, line breaks) and headings (eg, “background” and “objective”). Titles, abstracts, authors, and MeSH terms were combined and used as a single input. The input data underwent preprocessing with the inbuilt ktrain Bidirectional Encoder Representations from Transformers (BERT) preprocessing library before analysis with BERT [26]. Ktrain is a wrapper of the TensorFlow Keras library, which was used to conduct these BERT analyses. With this inbuilt BERT preprocessing model, a maximum length of 400 words was specified, which was selected due to the length of included texts and the processing requirements of the analysis.

Before use for logistic regression and XGBoost models, the input data underwent punctuation removal, negation detection, as well as stemming and conversion into a term frequency-inverse document frequency (TF-IDF) array. Punctuation removal involved the removal of all nonletter and nonnumerical characters with regular expressions. Negation detection was employed using the 19 libraries [27]. This negation detection uses a prespecified set of negating terms, following which subsequent terms will be flagged as negated. Word stemming was conducted with the Natural Language Toolkit Porter stemmer function. The Porter stemming approach involves the removal of common word suffixes and inflections. Finally, the text was converted into TF-IDF arrays using scikit-learn [28]. The n-gram range was allowed to vary from n-grams of 1 segment in length to 3 segments during training. Ultimately, the used n-gram range was 1 to 3. A similar approach was employed for the maximum number of features, which was set at 10,000.

Following this preprocessing, data were split into training and testing data sets randomly. This separation was conducted with a 3:1 train:test ratio. This train-test split was performed once. These preprocessing methods and the models used below were selected, as similar methods had previously proved effective in analyzing scientific abstracts [29].

Models were developed to predict whether a given article would be published in a first, second, or third tertile journal (0-33rd centile, 34th-66th centile, or 67th-100th centile), as ranked either by impact factor or Eigenfactor score. BERT, XGBoost, and logistic regression models were developed on the training data set before evaluation on the hold-out test data set. The logistic regression model was selected to provide a baseline for classification accuracy. This model was trained on the training data set and then evaluated on the test data set using the default values for all hyperparameters as in the scikit-learn library. The XGBoost and BERT algorithms were developed on the training set using 5-fold cross-validation. During this process, multiple XGBoost hyperparameters were varied, including the number of estimators, maximum depth, and learning rate. Following experimentation, there were no significant improvements in model performance on the training data set; therefore, library defaults were also used for this model. The BERT model was developed using ktrain. During the development of the BERT model, learning rate, epochs and batch size were allowed to vary. The hyperparameters used were a batch size of 3, a learning rate of 2×10^–5, and 3 epochs. Following training, these models were evaluated on a hold-out test data set (to minimize the risk of data leakage).

The primary outcome was overall classification accuracy for the best-performing model in the prediction of accepting journal impact factor tertile. In addition, 2 random examples were selected to demonstrate instances of correct and incorrect classification to improve interpretability. Open-source Python libraries were used for analysis, including XGBoost, scikit-learn, TensorFlow, Natural Language Toolkit, and ktrain [26-28,30,31].

There were 10,813 articles included in the study. These articles were from 382 unique journals. The median impact factor for the journals was 2.117 (IQR 1.102-2.622). The median Eigenfactor score was 0.00247 (IQR 0.00105-0.03).

The BERT model achieved the highest classification accuracy of 75.0%. XGBoost returned a classification accuracy of 71.6%. Logistic regression, which served as a baseline classification accuracy, achieved a classification accuracy of 65.4%.

Performance in Eigenfactor prediction was similar to impact factor prediction. The BERT model achieved a classification accuracy of 73.6%. XGBoost and logistic regression returned 71.8% and 65.3% classification accuracies, respectively. In this instance, logistic regression again provided a baseline classification accuracy.

An article by Kubak et al [32] was selected as a random example of a correct classification. When provided with the title, abstract, authors, and MeSH terms, as may have been provided at the time of submission, the BERT model correctly predicted that the article would be published in a highest tertile journal (Annals of Translational Medicine, impact factor of 3.932; Figure 1A). Conversely, the article by Gao et al [33] was predicted to be published in a highest tertile journal and was published in a middle tertile journal (Journal of Ophthalmology, impact factor 1.909; Figure 1B) [33].

This study demonstrates the ability of machine learning classifiers to predict the impact factor and Eigenfactor score of journals that will accept a given peer-reviewed article based on an abstract in the often-related fields of ophthalmology, neurology, and radiology. The BERT model was most effective in this study, achieving an impact factor and Eigenfactor score tertile classification accuracy of 75.0% and 73.6%, respectively. Artificial intelligence may assist researchers in selecting journals with a higher probability of acceptance, with the potential for increasing rates of full publication or reducing time to publication. Further studies are required to assess this.

The algorithms involved in this project have similar limitations to existing recommender systems. Since the models have learned from previous publications, they may perpetuate existing research trends regarding which studies are or are not considered a high priority. Similarly, the models may not have been exposed to new terminology associated with novel research, with resultant reduced prediction accuracy until the model is trained with these terms.

Ongoing research in this area may develop similar models that encompass additional fields. Research examining the use of such models to ascertain whether their use increases submission efficiency is required. In addition, future studies examining the development and use of other scientific writing recommender systems may be beneficial.

Content-based journal recommender systems have evolved with increasing accuracy with the introduction of artificial intelligence and deep learning algorithms. Compared to the major prior algorithms that recommend specific journals, our approach recommends a journal impact factor tertile. This may be beneficial in smaller domains, such as a clinical speciality, where the journals and their hierarchy are well known to researchers. Deciding which impact factor or Eiegenfactor score tertile to target may be more informative than a single journal. Other available major prior algorithms vary in their approaches. As mentioned, the Publication Recommender System uses TF-IDF and chi-squared feature selection and logistic regression as a classifier [13]. The accuracy of the classifier model for correctly recommending a journal within the top 3 suggestions was 0.6137. This algorithm was extended by Huynh et al [14] in their Scientific Submission Recommendation System for Computer Science (S2RSCS) algorithm using multilayer perceptrons as classifiers. Evaluated on the Wang et al’s [13] computer science data set, they achieved an accuracy of 0.8907 when using the title, abstract, and keywords as input to predict the top 3 computer science journals. Later, Nguyen et al [15] introduced S2CFT, a variation on this technique using deep learning that demonstrated improved accuracy. FastText embeddings were combined with a 1-dimensional convolutional neural network to produce a matching score model. The output vector from this model was averaged with the S2RSCS model output vector to create a new vector to calculate a matching score. The top-3 accuracy achieved on the Wang et al [13] computer science data set using title, abstract, and keywords was 0.9021 [15]. Deep learning is also used in Feng et al [34]’s Pubmender algorithm, which uses pretrained word2vec to construct the initial feature space, with a subsequent deep convolutional neural network to achieve a high-level representation of article abstracts [34]. A softmax regression model was used to recommend the top 3 biomedical journals. The accuracy of the Pubmender model for the top 3 biomedical journals was 0.71. Nguyen et al’s [16] “paper submission recommendation using mixtures of transformer encoders” (PSRMTE) algorithm is another framework that has shown further increased performance. The PSRMTE method uses different bidirectional transformer encoders and combines them using the mixtures of transformer encoders (MTE) technique to train a classification model to estimate the conditional probability of a given paper submission belonging to a journal. Matching scores between the paper submitted and each journal are used to return the top journals. PSRMTE performance was evaluated using combinations of title, abstract, and the list of keywords. The top-3 accuracy of the MTE algorithm on the Wang et al [13] computer science data set using the abstract and keywords was 0.9225. The excellent results of the MTE algorithm demonstrate the power of transformer encoders in this task. The extensive benefits of transformer encoders [16] and their demonstrated effectiveness explain why the BERT model performed best among our specified algorithms. Another algorithm, Content and Bipartite Graph to Recommend Journals for Submission (CBGJRS), is a hybrid algorithm that also uses transformer encoders [17]. CBGJRS uses a “BERT pre-training model to obtain a vector representation of articles by feature learning from abstracts at the text level and a self-coding network to obtain a vector representation of journals, followed by a scoring function and a softmax classifier to achieve journal recommendations.” [17] The accuracy of the CBGJRS approach to recommend a set of top 3 journals was 0.9536 using a closed data set and 0.7206 on a larger validation data set. The authors conclude that adding journal features in their hybrid approach improves performance. Our results suggest that inputs such as citation metrics could be incorporated into future algorithms. Further work on journal recommendations for clinical journals is required to show the cross-domain potential of these algorithms. This would require standardized data sets for clinical journals to allow for algorithm comparison, as has been used in evaluating algorithms in other domains.

In conclusion, we demonstrated the pilot ability of open-source artificial intelligence classifiers to predict the accepting-journal impact factor and Eigenfactor score of an article abstract in the fields of ophthalmology, neurology, and radiology. Among BERT, XGBoost, and logistic regression classifiers, the BERT model was the most accurate. Citation metrics may potentially be incorporated into future recommendation system algorithms. Further, artificial intelligence approaches to journal recommendation are evolving with the potential to increase the efficiency of the publication process for researchers.