Built-in filters and machine learning classifiers
Introduction
Entering search filters manually is only one way to use search filters. Some database interfaces (for example Ovid) include an option to apply a filter to search results without the need to manually key in the search filter. Some examples of database interfaces with in-built filters are described below .
Search filters are also a growing feature or option within systematic review software. Search filters within review software may be characterised as "classifiers" that have been designed by the software provider to find specific study designs from among a set of search results. Alternatively, some systematic review software offer do-it-yourself machine learning options. In that option, screeners can train the software to recognise specific types of study or relevant topics. Examples of these newer types of approach to designing and using filters using machine learning are described below.
Database interfaces with in-built filters
Clinical Queries
PubMed, Ovid Embase and Ovid MEDLINE all have the option to limit to 'Clinical Queries'. Using this feature allows the searcher to restrict their search results to one of the following:
therapy, diagnosis, prognosis, causation (etiology), clinical prediction guides, qualitative, economics, and cost.
Within these categories it is also possible to select whether sensitivity, specificity or a balance of these two factors is required. More information about these “Clinical Queries” search filters is available.
Other filters in PubMed and Ovid
Another built in filter available in PubMed, Ovid Embase and Ovid MEDLINE is the Systematic Reviews filter. In PubMed this can be applied via Clinical Queries, Article Types or by adding systematic [sb] to the search query. More information about the coverage of this filter is available. In the Ovid databases it can be applied via selecting Additional limits and then Clinical Queries
PubMed also includes a range of Medical Genetics search filters that can be applied to search results. These include filters for diagnosis, differential diagnosis, clinical description, management and testing. More information about these and their development is available.
Dialog Search Hedges
Dialog offers hedges for all Dialog databases within its Saved Searches and Alerts Workspace:
Cochrane publications
Medical Device Adverse Events
Specific Adverse Events
Humans
Elderly
Child
Animals
Covid-19
Clinical Trials
McMaster Superfilters
The Superfilters service from McMaster University is aimed at clinicians and provides a simultaneous search of pre-appraised clinical literature from the McMaster Premium Literature Service (McMaster PLUSTM) and PubMed, incorporating a wide selection of search filters including: therapy, diagnosis, prognosis, clinical prediction guides, economics, etiology for harm, quality improvement, knowledge translation, qualitative studies, appropriateness, process assessment, outcome assessment, costs and review. The searcher can select a specific, balanced or sensitive approach to suit their requirements. Registration required but subsequently free to use.
Automated filters
Automated filters are offered as standalone tools. One example is RCT Tagger. In this resource a subject search is entered into a search line and the search is run against PubMed. The results returned are presented in order of probability of the record being a report of an RCT according to the underlying algorithm. The development of the filter is presented in a supporting paper.
Filters (classifiers) within systematic review software and machine learning classifiers
Some systematic review software contains built-in filters or pre-designed classifiers which can identify certain types of study design.
EPPi-Reviewer offers (February 2024) the Cochrane RCT classifier and a randomized trials classifier, a systematic review classifier, and an economic evaluation classifier.
RobotReviewer offers RobotSearch (May 2023). This is a classifier that identifies reports of RCTs.
Covidence offers an option (February 2024) to remove studies which are not reporting RCTs, using the Cochrane RCT classifer.
Some software, particularly systematic review software, offer machine learning facilities which you can use to design or train your own bespoke classifier (search filter). These include (but are not limited to):
McMaster University offers a machine learning classifier to identify MEDLINE studies with a "high probability of being scientifically sound and clinically relevant to clinicians and medical researchers".
Access requires free registration.
Details of the algorithms behind the classifer are provided as well as a worked example.
Researchers are also providing stand alone classifiers for specific tasks:
Al-Jaishi, A.A., Taljaard, M., Al-Jaishi, M.D. et al. Machine learning algorithms to identify cluster randomized trials from MEDLINE and EMBASE. Syst Rev 2022; 11, 229.
The software is available here.
Machine learning classifiers are being compared to published filters and also being evaluated in other ways to assess their performance.
Researchers used a deep learning method to identify rigorous clinical research and compared their method to the PubMed Clinical Query treatment filters and a McMaster filter.
Del Fiol G, Michelson M, Iorio A, Cotoi C, Haynes RB. A Deep Learning Method to Automatically Identify Reports of Scientifically Rigorous Clinical Research from the Biomedical Literature: Comparative Analytic Study. J Med Internet Res. 2018;20(6):e10281. doi: 10.2196/10281.
Researchers seeking to develop a classifier to identify diagnostic test accuracy studies for diagnostic test accuracy reviews concluded that they were unable to develop an abstract classifier with sufficient sensitivity:
Kataoka Y, Taito S, Yamamoto N, et al. An open competition involving thousands of competitors failed to construct useful abstract classifiers for new diagnostic test accuracy systematic reviews. Res Synth Methods. 2023;14(5):707-717.