Main

With the emergence of online publishing, the number of scientific manuscripts on many topics is skyrocketing1. All of these textual data present opportunities to scholars and practitioners while simultaneously confronting them with new challenges. Scholars often develop systematic reviews and meta-analyses to develop comprehensive overviews of the relevant topics2. The process entails several explicit and, ideally, reproducible steps, including identifying all likely relevant publications in a standardized way, extracting data from eligible studies and synthesizing the results. Systematic reviews differ from traditional literature reviews in that they are more replicable and transparent3,4. Such systematic overviews of literature on a specific topic are pivotal not only for scholars, but also for clinicians, policy-makers, journalists and, ultimately, the general public5,6,7.

Given that screening the entire research literature on a given topic is too labour intensive, scholars often develop quite narrow searches. Developing a search strategy for a systematic review is an iterative process aimed at balancing recall and precision8,9; that is, including as many potentially relevant studies as possible while simultaneously limiting the total number of studies retrieved. The vast number of publications in the field of study often leads to a relatively precise search, with the risk of missing relevant studies. The process of systematic reviewing is error prone and extremely time intensive10. In fact, if the literature of a field is growing faster than the amount of time available for systematic reviews, adequate manual review of this field then becomes impossible11.

The rapidly evolving field of machine learning has aided researchers by allowing the development of software tools that assist in developing systematic reviews11,12,13,14. Machine learning offers approaches to overcome the manual and time-consuming screening of large numbers of studies by prioritizing relevant studies via active learning15. Active learning is a type of machine learning in which a model can choose the data points (for example, records obtained from a systematic search) it would like to learn from and thereby drastically reduce the total number of records that require manual screening16,17,18. In most so-called human-in-the-loop19 machine-learning applications, the interaction between the machine-learning algorithm and the human is used to train a model with a minimum number of labelling tasks. Unique to systematic reviewing is that not only do all relevant records (that is, titles and abstracts) need to seen by a researcher, but an extremely diverse range of concepts also need to be learned, thereby requiring flexibility in the modelling approach as well as careful error evaluation11. In the case of systematic reviewing, the algorithm(s) are interactively optimized for finding the most relevant records, instead of finding the most accurate model. The term researcher-in-the-loop was introduced20 as a special case of human-in-the-loop with three unique components: (1) the primary output of the process is a selection of the records, not a trained machine learning model; (2) all records in the relevant selection are seen by a human at the end of the process21; (3) the use-case requires a reproducible workflow and complete transparency is required22.

Existing tools that implement such an active learning cycle for systematic reviewing are described in Table 1; see the Supplementary Information for an overview of all of the software that we considered (note that this list was based on a review of software tools12). However, existing tools have two main drawbacks. First, many are closed source applications with black box algorithms, which is problematic as transparency and data ownership are essential in the era of open science22. Second, to our knowledge, existing tools lack the necessary flexibility to deal with the large range of possible concepts to be learned by a screening machine. For example, in systematic reviews, the optimal type of classifier will depend on variable parameters, such as the proportion of relevant publications in the initial search and the complexity of the inclusion criteria used by the researcher23. For this reason, any successful system must allow for a wide range of classifier types. Benchmark testing is crucial to understand the real-world performance of any machine learning-aided system, but such benchmark options are currently mostly lacking.

Table 1 Existing tools that implement an RITL-based active learning cycle for systematic reviewing
Full size table

In this paper we present an open source machine learning-aided pipeline with active learning for systematic reviews called ASReview. The goal of ASReview is to help scholars and practitioners to get an overview of the most relevant records for their work as efficiently as possible while being transparent in the process. The open, free and ready-to-use software ASReview addresses all concerns mentioned above: it is open source, uses active learning, allows multiple machine learning models. It also has a benchmark mode, which is especially useful for comparing and designing algorithms. Furthermore, it is intended to be easily extensible, allowing third parties to add modules that enhance the pipeline. Although we focus this paper on systematic reviews, ASReview can handle any text source.

In what follows, we first present the pipeline for manual versus machine learning-aided systematic reviews. We then show how ASReview has been set up and how ASReview can be used in different workflows by presenting several real-world use cases. We subsequently demonstrate the results of simulations that benchmark performance and present the results of a series of user-experience tests. Finally, we discuss future directions.

Pipeline for manual and machine learning-aided systematic reviews

The pipeline of a systematic review without active learning traditionally starts with researchers doing a comprehensive search in multiple databases24, using free text words as well as controlled vocabulary to retrieve potentially relevant references. The researcher then typically verifies that the key papers they expect to find are indeed included in the search results. The researcher downloads a file with records containing the text to be screened. In the case of systematic reviewing it contains the titles and abstracts (and potentially other metadata such as the authors’s names, journal name, DOI) of potentially relevant references into a reference manager. Ideally, two or more researchers then screen the records’s titles and abstracts on the basis of the eligibility criteria established beforehand4. After all records have been screened, the full texts of the potentially relevant records are read to determine which of them will be ultimately included in the review. Most records are excluded in the title and abstract phase. Typically, only a small fraction of the records belong to the relevant class, making title and abstract screening an important bottleneck in systematic reviewing process25. For instance, a recent study analysed 10,115 records and excluded 9,847 after title and abstract screening, a drop of more than 95%26. ASReview therefore focuses on this labour-intensive step.

The research pipeline of ASReview is depicted in Fig. 1. The researcher starts with a search exactly as described above and subsequently uploads a file containing the records (that is, metadata containing the text of the titles and abstracts) into the software. Prior knowledge is then selected, which is used for training of the first model and presenting the first record to the researcher. As screening is a binary classification problem, the reviewer must select at least one key record to include and exclude on the basis of background knowledge. More prior knowledge may result in improved efficiency of the active learning process.

Fig. 1: Machine learning-aided pipeline for ASReview.
figure 1

The symbols indicate whether the action is taken by a human, a computer, or whether both options are available.

Full size image

A machine learning classifier is trained to predict study relevance (labels) from a representation of the record-containing text (feature space) on the basis of prior knowledge. We have purposefully chosen not to include an author name or citation network representation in the feature space to prevent authority bias in the inclusions. In the active learning cycle, the software presents one new record to be screened and labelled by the user. The user’s binary label (1 for relevant versus 0 for irrelevant) is subsequently used to train a new model, after which a new record is presented to the user. This cycle continues up to a certain user-specified stopping criterion has been reached. The user now has a file with (1) records labelled as either relevant or irrelevant and (2) unlabelled records ordered from most to least probable to be relevant as predicted by the current model. This set-up helps to move through a large database much quicker than in the manual process, while the decision process simultaneously remains transparent.

Software implementation for ASReview

The source code27 of ASReview is available open source under an Apache 2.0 license, including documentation28. Compiled and packaged versions of the software are available on the Python Package Index29 or Docker Hub30. The free and ready-to-use software ASReview implements oracle, simulation and exploration modes. The oracle mode is used to perform a systematic review with interaction by the user, the simulation mode is used for simulation of the ASReview performance on existing datasets, and the exploration mode can be used for teaching purposes and includes several preloaded labelled datasets.

The oracle mode presents records to the researcher and the researcher classifies these. Multiple file formats are supported: (1) RIS files are used by digital libraries such as IEEE Xplore, Scopus and ScienceDirect; the citation managers Mendeley, RefWorks, Zotero and EndNote support the RIS format too. (2) Tabular datasets with the .csv, .xlsx and .xls file extensions. CSV files should be comma separated and UTF-8 encoded; the software for CSV files accepts a set of predetermined labels in line with the ones used in RIS files. Each record in the dataset should hold the metadata on, for example, a scientific publication. Mandatory metadata is text and can, for example, be titles or abstracts from scientific papers. If available, both are used to train the model, but at least one is needed. An advanced option is available that splits the title and abstracts in the feature-extraction step and weights the two feature matrices independently (for TF–IDF only). Other metadata such as author, date, DOI and keywords are optional but not used for training the models. When using ASReview in the simulation or exploration mode, an additional binary variable is required to indicate historical labelling decisions. This column, which is automatically detected, can also be used in the oracle mode as background knowledge for previous selection of relevant papers before entering the active learning cycle. If unavailable, the user has to select at least one relevant record that can be identified by searching the pool of records. At least one irrelevant record should also be identified; the software allows to search for specific records or presents random records that are most likely to be irrelevant due to the extremely imbalanced data.

The software has a simple yet extensible default model: a naive Bayes classifier, TF–IDF feature extraction, a dynamic resampling balance strategy31 and certainty-based sampling17,32 for the query strategy. These defaults were chosen on the basis of their consistently high performance in benchmark experiments across several datasets31. Moreover, the low computation time of these default settings makes them attractive in applications, given that the software should be able to run locally. Users can change the settings, shown in Table 2, and technical details are described in our documentation28. Users can also add their own classifiers, feature extraction techniques, query strategies and balance strategies.

Table 2 Implemented classifiers, feature extraction techniques, query strategies and balance strategies available in ASReview
Full size table

ASReview has a number of implemented features (see Table 2). First, there are several classifiers available: (1) naive Bayes; (2) support vector machines; (3) logistic regression; (4) neural networks; (5) random forests; (6) LSTM-base, which consists of an embedding layer, an LSTM layer with one output, a dense layer and a single sigmoid output node; and (7) LSTM-pool, which consists of an embedding layer, an LSTM layer with many outputs, a max pooling layer and a single sigmoid output node. The feature extraction techniques available are Doc2Vec33, embedding LSTM, embedding with IDF or TF–IDF34 (the default is unigram, with the option to run n-grams while other parameters are set to the defaults of Scikit-learn35) and sBERT36. The available query strategies for the active learning part are (1) random selection, ignoring model-assigned probabilities; (2) uncertainty-based sampling, which chooses the most uncertain record according to the model (that is, closest to 0.5 probability); (3) certainty-based sampling (max in ASReview), which chooses the record most likely to be included according to the model; and (4) mixed sampling, which uses a combination of random and certainty-based sampling.

There are several balance strategies that rebalance and reorder the training data. This is necessary, because the data is typically extremely imbalanced and therefore we have implemented the following balance strategies: (1) full sampling, which uses all of the labelled records; (2) undersampling the irrelevant records so that the included and excluded records are in some particular ratio (closer to one); and (3) dynamic resampling, a novel method similar to undersampling in that it decreases the imbalance of the training data31. However, in dynamic resampling, the number of irrelevant records is decreased, whereas the number of relevant records is increased by duplication such that the total number of records in the training data remains the same. The ratio between relevant and irrelevant records is not fixed over interactions, but dynamically updated depending on the number of labelled records, the total number of records and the ratio between relevant and irrelevant records. Details on all of the described algorithms can be found in the code and documentation referred to above.

By default, ASReview converts the records’s texts into a document-term matrix, terms are converted to lowercase and no stop words are removed by default (but this can be changed). As the document-term matrix is identical in each iteration of the active learning cycle, it is generated in advance of model training and stored in the (active learning) state file. Each row of the document-term matrix can easily be requested from the state-file. Records are internally identified by their row number in the input dataset. In oracle mode, the record that is selected to be classified is retrieved from the state file and the record text and other metadata (such as title and abstract) are retrieved from the original dataset (from the file or the computer’s memory). ASReview can run on your local computer, or on a (self-hosted) local or remote server. Data (all records and their labels) remain on the users’s computer. Data ownership and confidentiality are crucial and no data are processed or used in any way by third parties. This is unique by comparison with some of the existing systems, as shown in the last column of Table 1.

Real-world use cases and high-level function descriptions

Below we highlight a number of real-world use cases and high-level function descriptions for using the pipeline of ASReview.

ASReview can be integrated in classic systematic reviews or meta-analyses. Such reviews or meta-analyses entail several explicit and reproducible steps, as outlined in the PRISMA guidelines4. Scholars identify all likely relevant publications in a standardized way, screen retrieved publications to select eligible studies on the basis of defined eligibility criteria, extract data from eligible studies and synthesize the results. ASReview fits into this process, particularly in the abstract screening phase. ASReview does not replace the initial step of collecting all potentially relevant studies. As such, results from ASReview depend on the quality of the initial search process, including selection of databases24 and construction of comprehensive searches using keywords and controlled vocabulary. However, ASReview can be used to broaden the scope of the search (by keyword expansion or omitting limitation in the search query), resulting in a higher number of initial papers to limit the risk of missing relevant papers during the search part (that is, more focus on recall instead of precision).

Furthermore, many reviewers nowadays move towards meta-reviews when analysing very large literature streams, that is, systematic reviews of systematic reviews37. This can be problematic as the various reviews included could use different eligibility criteria and are therefore not always directly comparable. Due to the efficiency of ASReview, scholars using the tool could conduct the study by analysing the papers directly instead of using the systematic reviews. Furthermore, ASReview supports the rapid update of a systematic review. The included papers from the initial review are used to train the machine learning model before screening of the updated set of papers starts. This allows the researcher to quickly screen the updated set of papers on the basis of decisions made in the initial run.

As an example case, let us look at the current literature on COVID-19 and the coronavirus. An enormous number of papers are being published on COVID-19. It is very time consuming to manually find relevant papers (for example, to develop treatment guidelines). This is especially problematic as urgent overviews are required. Medical guidelines rely on comprehensive systematic reviews, but the medical literature is growing at breakneck pace and the quality of the research is not universally adequate for summarization into policy38. Such reviews must entail adequate protocols with explicit and reproducible steps, including identifying all potentially relevant papers, extracting data from eligible studies, assessing potential for bias and synthesizing the results into medical guidelines. Researchers need to screen (tens of) thousands of COVID-19-related studies by hand to find relevant papers to include in their overview. Using ASReview, this can be done far more efficiently by selecting key papers that match their (COVID-19) research question in the first step; this should start the active learning cycle and lead to the most relevant COVID-19 papers for their research question being presented next. A plug-in was therefore developed for ASReview39, which contained three databases that are updated automatically whenever a new version is released by the owners of the data: (1) the Cord19 database, developed by the Allen Institute for AI, with over all publications on COVID-19 and other coronavirus research (for example SARS, MERS and so on) from PubMed Central, the WHO COVID-19 database of publications, the preprint servers bioRxiv and medRxiv and papers contributed by specific publishers40. The CORD-19 dataset is updated daily by the Allen Institute for AI and updated also daily in the plugin. (2) In addition to the full dataset, we automatically construct a daily subset of the database with studies published after December 1st, 2019 to search for relevant papers published during the COVID-19 crisis. (3) A separate dataset of COVID-19 related preprints, containing metadata of preprints from over 15 preprints servers across disciplines, published since January 1st, 202041. The preprint dataset is updated weekly by the maintainers and then automatically updated in ASReview as well. As this dataset is not readily available to researchers through regular search engines (for example, PubMed), its inclusion in ASReview provided added value to researchers interested in COVID-19 research, especially if they want a quick way to screen preprints specifically.

Simulation study

To evaluate the performance of ASReview on a labelled dataset, users can employ the simulation mode. As an example, we ran simulations based on four labelled datasets with version 0.7.2 of ASReview. All scripts to reproduce the results in this paper can be found on Zenodo (https://doi.org/10.5281/zenodo.4024122)42, whereas the results are available at OSF (https://doi.org/10.17605/OSF.IO/2JKD6)43.

Datasets

First, we analysed the performance for a study systematically describing studies that performed viral metagenomic next-generation sequencing in common livestock such as cattle, small ruminants, poultry and pigs44. Studies were retrieved from Embase (n = 1,806), Medline (n = 1,384), Cochrane Central (n = 1), Web of Science (n = 977) and Google Scholar (n = 200, the top relevant references). After deduplication this led to 2,481 studies obtained in the initial search, of which 120 were inclusions (4.84%).

A second simulation study was performed on the results for a systematic review of studies on fault prediction in software engineering45. Studies were obtained from ACM Digital Library, IEEExplore and the ISI Web of Science. Furthermore, a snowballing strategy and a manual search were conducted, accumulating to 8,911 publications of which 104 were included in the systematic review (1.2%).

A third simulation study was performed on a review of longitudinal studies that applied unsupervised machine learning techniques to longitudinal data of self-reported symptoms of the post-traumatic stress assessed after trauma exposure46,47; 5,782 studies were obtained by searching Pubmed, Embase, PsychInfo and Scopus and through a snowballing strategy in which both the references and the citation of the included papers were screened. Thirty-eight studies were included in the review (0.66%).

A fourth simulation study was performed on the results for a systematic review on the efficacy of angiotensin-converting enzyme inhibitors, from a study collecting various systematic review datasets from the medical sciences15. The collection is a subset of 2,544 publications from the TREC 2004 Genomics Track document corpus48. This is a static subset from all MEDLINE records from 1994 through 2003, which allows for replicability of results. Forty-one publications were included in the review (1.6%).

Performance metrics

We evaluated the four datasets using three performance metrics. We first assess the work saved over sampling (WSS), which is the percentage reduction in the number of records needed to screen achieved by using active learning instead of screening records at random; WSS is measured at a given level of recall of relevant records, for example 95%, indicating the work reduction in screening effort at the cost of failing to detect 5% of the relevant records. For some researchers it is essential that all relevant literature on the topic is retrieved; this entails that the recall should be 100% (that is, WSS@100%). We also propose the amount of relevant references found after having screened the first 10% of the records (RRF10%). This is a useful metric for getting a quick overview of the relevant literature.

Results

For every dataset, 15 runs were performed with one random inclusion and one random exclusion (see Fig. 2). The classical review performance with randomly found inclusions is shown by the dashed line. The average work saved over sampling at 95% recall for ASReview is 83% and ranges from 67% to 92%. Hence, 95% of the eligible studies will be found after screening between only 8% to 33% of the studies. Furthermore, the number of relevant abstracts found after reading 10% of the abstracts ranges from 70% to 100%. In short, our software would have saved many hours of work.

Fig. 2: Simulation results.
figure 2

a–d, Results of the simulation study for the results for a study systematically review studies that performed viral metagenomic next-generation sequencing in common livestock (a), results for a systematic review of studies on fault prediction in software engineering (b), results for longitudinal studies that applied unsupervised machine learning techniques on longitudinal data of self-reported symptoms of posttraumatic stress assessed after trauma exposure (c), and results for a systematic review on the efficacy of angiotensin-converting enzyme inhibitors (d). Fiteen runs (shown with separate lines) were performed for every dataset, with only one random inclusion and one random exclusion. The classical review performances with randomly found inclusions are shown by the dashed lines.

Full size image

Usability testing (user experience testing)

We conducted a series of user experience tests to learn from end users how they experience the software and implement it in their workflow. The study was approved by the Ethics Committee of the Faculty of Social and Behavioral Sciences of Utrecht University (ID 20-104).

Unstructured interviews

The first user experience (UX) test—carried out in December 2019—was conducted with an academic research team in a substantive research field (public administration and organizational science) that has conducted various systematic reviews and meta-analyses. It was composed of three university professors (ranging from assistant to full) and three PhD candidates. In one 3.5 h session, the participants used the software and provided feedback via unstructured interviews and group discussions. The goal was to provide feedback on installing the software and testing the performance on their own data. After these sessions we prioritized the feedback in a meeting with the ASReview team, which resulted in the release of v.0.4 and v.0.6. An overview of all releases can be found on GitHub27.

A second UX test was conducted with four experienced researchers developing medical guidelines based on classical systematic reviews, and two experienced reviewers working at a pharmaceutical non-profit organization who work on updating reviews with new data. In four sessions, held in February to March 2020, these users tested the software following our testing protocol. After each session we implemented the feedback provided by the experts and asked them to review the software again. The main feedback was about how to upload datasets and select prior papers. Their feedback resulted in the release of v.0.7 and v.0.9.

Systematic UX test

In May 2020 we conducted a systematic UX test. Two groups of users were distinguished: an unexperienced group and an experienced user who already used ASReview. Due to the COVID-19 lockdown the usability tests were conducted via video calling where one person gave instructions to the participant and one person observed, called human-moderated remote testing49. During the tests, one person (SH) asked the questions and helped the participant with the tasks, the other person observed and made notes, a user experience professional at the IT department of Utrecht University (MH).

To analyse the notes, thematic analysis was used, which is a method to analyse data by dividing the information in subjects that all have a different meaning50 using the Nvivo 12 software51. When something went wrong the text was coded as showstopper, when something did not go smoothly the text was coded as doubtful, and when something went well the subject was coded as superb. The features the participants requested for future versions of the ASReview tool were discussed with the lead engineer of the ASReview team and were submitted to GitHub as issues or feature requests.

The answers to the quantitative questions can be found at the Open Science Framework52. The participants (N = 11) rated the tool with a grade of 7.9 (s.d. = 0.9) on a scale from one to ten (Table 2). The unexperienced users on average rated the tool with an 8.0 (s.d. = 1.1, N = 6). The experienced user on average rated the tool with a 7.8 (s.d. = 0.9, N = 5). The participants described the usability test with words such as helpful, accessible, fun, clear and obvious.

The UX tests resulted in the new release v0.10, v0.10.1 and the major release v0.11, which is a major revision of the graphical user interface. The documentation has been upgraded to make installing and launching ASReview more straightforward. We made setting up the project, selecting a dataset and finding past knowledge is more intuitive and flexible. We also added a project dashboard with information on your progress and advanced settings.

Continuous input via the open source community

Finally, the ASReview development team receives continuous feedback from the open science community about, among other things, the user experience. In every new release we implement features listed by our users. Recurring UX tests are performed to keep up with the needs of users and improve the value of the tool.

Conclusion

We designed a system to accelerate the step of screening titles and abstracts to help researchers conduct a systematic review or meta-analysis as efficiently and transparently as possible. Our system uses active learning to train a machine learning model that predicts relevance from texts using a limited number of labelled examples. The classifier, feature extraction technique, balance strategy and active learning query strategy are flexible. We provide an open source software implementation, ASReview with state-of-the-art systems across a wide range of real-world systematic reviewing applications. Based on our experiments, ASReview provides defaults on its parameters, which exhibited good performance on average across the applications we examined. However, we stress that in practical applications, these defaults should be carefully examined; for this purpose, the software provides a simulation mode to users. We encourage users and developers to perform further evaluation of the proposed approach in their application, and to take advantage of the open source nature of the project by contributing further developments.

Drawbacks of machine learning-based screening systems, including our own, remain. First, although the active learning step greatly reduces the number of manuscripts that must be screened, it also prevents a straightforward evaluation of the system’s error rates without further onerous labelling. Providing users with an accurate estimate of the system’s error rate in the application at hand is therefore a pressing open problem. Second, although, as argued above, the use of such systems is not limited in principle to reviewing, no empirical benchmarks of actual performance in these other situations yet exist to our knowledge. Third, machine learning-based screening systems automate the screening step only; although the screening step is time-consuming and a good target for automation, it is just one part of a much larger process, including the initial search, data extraction, coding for risk of bias, summarizing results and so on. Although some other works, similar to our own, have looked at (semi-)automating some of these steps in isolation53,54, to our knowledge the field is still far removed from an integrated system that would truly automate the review process while guaranteeing the quality of the produced evidence synthesis. Integrating the various tools that are currently under development to aid the systematic reviewing pipeline is therefore a worthwhile topic for future development.

Possible future research could also focus on the performance of identifying full text articles with different document length and domain-specific terminologies or even other types of text, such as newspaper articles and court cases. When the selection of past knowledge is not possible based on expert knowledge, alternative methods could be explored. For example, unsupervised learning or pseudolabelling algorithms could be used to improve training55,56. In addition, as the NLP community pushes forward the state of the art in feature extraction methods, these are easily added to our system as well. In all cases, performance benefits should be carefully evaluated using benchmarks for the task at hand. To this end, common benchmark challenges should be constructed that allow for an even comparison of the various tools now available. To facilitate such a benchmark, we have constructed a repository of publicly available systematic reviewing datasets57.

The future of systematic reviewing will be an interaction with machine learning algorithms to deal with the enormous increase of available text. We invite the community to contribute to open source projects such as our own, as well as to common benchmark challenges, so that we can provide measurable and reproducible improvement over current practice.