OBJECTIVES

Systematic literature reviews (SLRs) are foundational for evidence-based decisions-making across the pharmaceutical and medical devices product lifecycle, from early research and development to market launch, and post-market activities.

Conventional SLRs are often time- and resource-intensive, leading to delayed insights, decision making, and variable outputs. Artificial intelligence (AI)-driven approaches offer promising opportunities to streamline these processes.

This study aimed to develop and validate an AI-powered tool to accelerate and enhance the SLR process enhancing efficacy and curtailing time consumption, while preserving methodological rigor and adherence to established systematic review methodologies.

METHODS

The tool searches and automatically deduplicate articles across multiple libraries (PubMed, Semantic Scholar), and Google Scholar based on the defined research questions.

It leverages large language models to automate two primary tasks:

1) title/abstract screening, via data extraction aligned with PICOTS (Population, Intervention, Comparator and Outcome(s), Timing, Setting), study type and sample size;

2) customization of the research scope through filtering criteria to exclude irrelevant articles.

RESULTS

In pilot evaluations involving five use-cases of SLR across different therapeutic areas (including inflammatory diseases, dermatology, neurology, and orthopaedics), the tool showed good performances in data extractions (e.g., PICOTS, study type) and relevance of selected articles after filtering criteria definition were applied.

Across all use-cases, a comparison between the AI-assisted review and expert manual review showed recall rates ranging from 96% to 100%, while reducing the number of abstracts to read from 46% up to 90% for the initial title/abstract screening phase of an SLR.

CONCLUSIONS

This AI-powered SLR tool enables faster and consistent literature search to support decision-making across the drug, and medical device development lifecycle.

By reducing the manual screening workload, it improves time-to-insight while maintaining compliance with methodological standards.

Future work will focus on integrating of additional functionalities such as adding bibliographic databases and implementing relevance-based ranking.