OBJECTIVES: Literature review and synthesis of clinical evidence (ES) require increasing time and resources as the amount of available clinical literature increases. To limit this burden, artificial intelligence (AI) methods may be used at different stages of ES. This review aims to summarize current state of AI-based solutions used in ES.

METHODS: A targeted literature was performed, using EMBASE, MEDLINE and grey literature, limited to 2020-2021. A previously developed search filter for AI was used. Topics of interest included different steps of the ES process: keywords selection, literature selection and information classification, data extraction, and analysis.

RESULTS: Out of 474 hits, 48 publications were selected. We identified 37 commercial and non-commercial custom tools and 4 solutions using standard Python or R libraries. Most solutions are based on machine learning or natural language processing. They mostly cover literature selection and information classification (n=26/70%), data extraction (n=8/22%), keywords selection (n=6/16%), and analysis (n=1/3%). 17 tools (none commercial) share their source code, often available through GitHub. 13 publications reported on integrated automatic ES approaches, including 5 on real-time systematic literature review (SLR) or meta-analysis (MA). Good performance vs. human experts was generally reported, but quality of reporting was heterogeneous. Studies emphasized savings in man-hours. No tool allowed for complete elimination of human involvement , mostly involving preparing training datasets or human supervision of AI, at all ES stages.

CONCLUSIONS: Recent advances in AI-based solutions allow for automatic ES with reported high concordance with human expert and a substantial reduction in time. However, published examples of AI applications may be limited to the most successful ones and the generalisability of performance measures is questionable.