OBJECTIVES : Performing systematic literature reviews (SLRs) on clinical data is a time-consuming process. A retrospective analysis was conducted based on the clinical effectiveness searches from health technology assessments (HTAs) to validate the use of artificial intelligence (AI) and machine-learning in clinical evidence gathering.

METHODS : HTAs from the National Institute for Health and Care Excellence (NICE) were selected if they included sufficient detail on the efficacy and safety SLR search criteria and provided the reference list of included studies. AI capable of extracting disease terms, interventions, and outcomes from references in PubMed was used to perform searches corresponding to the search criteria for each SLR.

RESULTS : Five HTAs detailing efficacy and safety SLRs were identified across a range of indications: type 1 diabetes (TA622), non-small-cell lung cancer (TA571), multiple myeloma (TA587), chronic lymphocytic leukaemia (TA689), and migraine (TA631). These SLRs identified 42 relevant clinical evidence publications, 40 of which were identified by the AI platform, Evid Science. Of the two publications not identified, one was from a conference not searched on the AI platform. The AI was unable to read and extract data from the second publication.

Information on the total number of studies screened was available from two SLRs. In TA689, a total of 1,691 studies were screened across all databases searched while 3,073 studies were screened in TA631. In contrast, the searches developed using AI identified 1,585 and 569 studies, respectively.

CONCLUSIONS : The use of AI enhances search precision enabling relevant studies to be identified while likely reducing screening time. Further research is required to understand the magnitude of time savings. However, there are errors from inputs to the machine learning of about 2.5% that need to be taken into consideration. The error rate can be mitigated by hand-searching and citation checking as per SLR protocol.