Systematic review software​

Reviewers can use machine learning to help assist in screening and data extraction. Different machine learning algorithms are implemented for systematic review (semi)automation, for example, a convolutional neural network, long short-term memory, vector machines, or a combination of these algorithms. There are roughly two deployment approaches, generic machine learning, and review-specific machine learning. In the generic machine learning models approach, an ML model is trained to do a specific task applicable across various systematic reviews. Generic machine learning models are often trained to apply specific exclusion criteria like detecting randomized controlled trials. With the review-specific machine learning approach, a model is trained to make in- and exclusion decisions based on a human’s previous decisions with this particular review. Active learning is used to update the ranking of the studies likely needing to be included. These prioritization algorithms help reviewers screen quicker by cueing the papers likely needing to be included to be screened first until the next following article is almost certainly not relevant.

One of the current constrained in automating literature screening is the availability of training data. We are working on enabling easy sharing of datasets under different licensing. Users can configure the settings for the machine learning models that can be used. Users can use the ML evidence dashboard to evaluate and validate these machine learning models.