Multi-Tagger File Download

We have processed all PubMed-indexed articles (written in English, containing an abstract, with at least 25 words in title+abstract). For each article, we have estimated the probability that it belongs to each of the 50 most important and the most clinically relevant publication types and study designs. Here, we provide a comprehensive dataset of PubMed articles and their predicted publication type/study design scores. (The Anne O'Tate tool's Publication Types button also allows users to view predicted publication types for articles retrieved by any query in PubMed.)

There are large files for all articles up through December 31, 2020, and yearly files going forward that are updated weekly.

All files are tab-delimited with a header row. Each row contains 50 columns, separated by tab ('\t', ASCII 9). The first column is the PubMed article ID (PMID), followed by 49 columns for each Publication Type and Study Design processed by our models.
Click here to download a column reference that includes the multi-tagger model names as well as their optimal F1 threshholds.

Model Score Files: These files contain scores between 0 and 1 for each Publication Type and Study Design for each article. Scores represent estimated probabilities that the article belongs to each of the Publication Types or Study Designs. Scores are displayed as floating point numbers.

mt_modelscores_thru2020.tsv.gz
(all articles up to 12/31/2020, 1.4G compressed, 7.3G uncompressed)
mt_modelscores_2021-2023.tsv.gz
(full years 2021-2023, 295M compressed, 1.5G uncompressed)
mt_modelscores_2024.tsv.gz
(all articles currently available for 2024, 66M compressed, 333M uncompressed)

Model Indicator Files: These files are similar to those above, except that instead of showing estimated probability scores, we display 0 or 1 values. A value of 1 indicates that the article's probability score was above a threshold for considering the article to be each Publication Type or Study Design. The threshold values were chosen separately for each Publication Type or Study Design, by fixing the value that exhibited the highest F1 (balance of recall and precision) on our training data (see Column Reference above). These predictions are employed in the Publication Types button in our Anne O'Tate search engine.

mt_modelindicators_thru2020.tsv.gz
(all articles up to 12/31/2020, 100M compressed, 2.3G uncompressed)
mt_modelindicators_2021-2023.tsv.gz
(full years 2021-2023, 20M compressed, 457M uncompressed)
mt_modelindicators_2024.tsv.gz
(all articles currently available for 2024, 4M compressed, 102M uncompressed)

Reference:

Cohen AM, Schneider J, Fu Y, McDonagh MS, Das P, Holt AW, Smalheiser NR. (2021) Fifty Ways to Tag your Pubtypes: Multi-Tagger, a Set of Probabilistic Publication Type and Study Design Taggers to Support Biomedical Indexing and Evidence-Based Medicine. medRxiv. 2021 Jan 1. 2021.07.13.21260468v1.