ASReview is a free, open-source software tool that uses active machine learning to reorder title/abstract screening in a systematic review so that the records most likely to be relevant to your inclusion criteria surface earlier in the screening queue. It is developed and maintained by Utrecht University’s AI-aided Knowledge Discovery (AIKD) lab, and the underlying method was validated and published in Nature Machine Intelligence in 2021. Unlike a review-management workbench, ASReview is a screening-prioritization tool: it does not replace a pre-registered protocol, a documented search strategy, or dual independent screening — it changes the order in which records are presented to a human reviewer so that, on average, relevant records are found sooner in a large, unscreened set.
What ASReview actually does
A conventional systematic-review search can return anywhere from a few hundred to tens of thousands of candidate records after deduplication. Screening them in whatever order a database export happens to list them means a reviewer typically has to read through the full set, including a long tail of records that are almost certainly irrelevant, before screening is complete.
ASReview changes the order, not the workload rule. It works like this:
- You seed the model. A reviewer labels a small number of records as relevant or irrelevant, often starting from one or a few records already known to be relevant (a "prior knowledge" seed) plus a handful of clearly irrelevant ones.
- The model ranks the rest. Using a text classifier trained on those labels (title and abstract text, with a configurable feature representation and classifier such as naive Bayes or SVM in ASReview LAB), the software predicts which unscreened records are most likely to also be relevant and presents the single most-likely record next.
- The model retrains after every decision. As the reviewer labels each new record, the ranking of everything still unscreened is recalculated, so the model keeps refining its sense of what "relevant" looks like as screening continues.
- Screening still ends at a decision, not a preset cutoff. ASReview does not tell you how many records are "really" relevant; teams typically keep screening until they hit a pre-specified stopping rule (for example, a run of consecutive irrelevant records) and report that rule in their methods, consistent with PRISMA reporting expectations for how screening was conducted.
This is the same underlying idea, screening prioritization via active learning, used by several other systematic-review tools; what’s specific to ASReview is that it’s free, open-source, and built as an academic research tool with a published, peer-reviewed evaluation of the method behind it, rather than a proprietary black-box ranking algorithm inside a commercial product.
Who built it, and what’s been validated
ASReview is a project of Utrecht University’s AI-aided Knowledge Discovery lab. The core method was described in van de Schoot et al., "An open source machine learning framework for efficient and transparent systematic reviews," published in Nature Machine Intelligence, volume 3, in 2021, a peer-reviewed methods paper, not vendor marketing copy, which is unusual in this software category and is a large part of why the tool is trusted inside academic and information-specialist workflows. The software itself (ASReview LAB, now on its version-3 release line) is distributed on GitHub under an Apache 2.0 open-source license, meaning there is no purchase price, subscription fee, or per-review-team licensing cost for the software itself; the only real cost is whatever compute you run it on, since it’s self-hosted (a local install, an institutional server, or your own cloud instance) rather than sold as a managed SaaS product.
How ASReview fits into a review workflow
ASReview is a screening tool, not a full review-management platform. It sits after your search strategy is finalized and your records are deduplicated, and before full-text screening and data extraction. Concretely:
- What it replaces: the arbitrary or alphabetical/chronological order a reference-manager export would otherwise put your candidate records in.
- What it doesn’t replace: a documented eligibility criteria set, a pre-registered protocol, dual-reviewer conflict resolution, full-text screening, data extraction forms, risk-of-bias assessment, or PRISMA flow-diagram reporting, all of which still need to happen, whether inside ASReview’s own interface (which supports an "Oracle" mode for real reviews, plus "Exploration" and "Simulation" modes for training and methods testing) or in whatever tool you use downstream.
- Where it’s strongest: large candidate sets (many hundreds to many thousands of records) where reading everything in an unordered list would otherwise take substantially longer than reading the same records in a relevance-prioritized order. On smaller record sets, the time saved by prioritization is proportionally smaller.
Reviewers who use active-learning prioritization and stop screening before reaching every record still need to disclose that in their methods section: how many records were screened, what stopping rule was used, and ideally what proportion of known-relevant records the stopping point captured, the same transparency expectation that applies to any deviation from exhaustively screening a full candidate set.
ASReview vs. other systematic-review screening tools
ASReview is one of several tools research teams use at the screening stage of a review; the main practical differences are cost model and scope:
- Rayyan is a free-to-start, hosted screening platform that also uses machine-learning relevance ranking, plus built-in duplicate detection and blinded dual-reviewer workflows.
- Covidence is a paid, hosted review workbench covering screening, full-text review, data extraction, and risk-of-bias assessment in one connected pipeline, and is the tool Cochrane recommends for authors preparing new Cochrane reviews.
- DistillerSR is an enterprise-grade, paid platform built for regulated, auditable review workflows (pharma, HTA/HEOR, guideline development), with its own AI-assisted prioritization features layered on top of a full review-management pipeline.
Against that set, ASReview’s distinguishing features are: no license cost, a peer-reviewed and published method, and a narrower scope (screening prioritization specifically, rather than a full extraction-and-reporting workbench), which makes it a common choice for teams who already have a data-extraction and reporting process and specifically want a transparent, free, well-documented prioritization engine to sit in front of it. See CASRAI’s broader systematic literature review tools comparison and AI tools for systematic literature reviews guide for how these and other tools compare across the full screening-to-extraction pipeline.
Frequently asked questions
Is ASReview free?
Yes. ASReview LAB is open-source software released under an Apache 2.0 license, with no purchase price or subscription fee for the software itself. Because it’s self-hosted rather than offered as managed SaaS, the only real cost is the compute you run it on, a laptop, an institutional server, or your own cloud instance.
Does ASReview replace dual-reviewer screening?
No. ASReview changes the order records are presented in; it doesn’t remove the need for the screening approach your protocol specifies. Many teams still use two independent reviewers and resolve disagreements the same way they would with any other screening tool or method.
Is ASReview accepted for Cochrane or PRISMA-reported reviews?
Active-learning screening prioritization is increasingly reported in published systematic reviews, including some Cochrane reviews, but any deviation from screening every candidate record (i.e., stopping before 100% of records are screened) needs to be disclosed and justified in the methods section, with the stopping rule stated, consistent with PRISMA 2020 reporting guidance on how the study selection process was conducted.
What’s the difference between ASReview LAB, Oracle mode, and Simulation mode?
ASReview LAB is the software itself. Within it, Oracle mode is the mode used for a real, live screening review where a human reviewer labels records as the model actively re-ranks the remaining set. Exploration and Simulation modes are used for training and for methods testing, for example, evaluating how well active-learning prioritization would have performed against an already-completed, fully-labeled review.
This page provides general, vendor-neutral information about a third-party open-source tool for CASRAI’s research-audience readers. It is not sponsored by, affiliated with, or reviewed by Utrecht University or the ASReview project.
