The challenges of AI-supported literature search in pharmacovigilance

The continuous monitoring of scientific literature is an essential part of pharmacovigilance (PV). It has the purpose of identifying scientific publications with potential side effects, interactions, and other safety-relevant aspects of medicinal products. There it serves to obtain new information for the assessment of benefit-risk profiles. However, manual searches are reaching their limits due to the rapidly increasing number of scientific publications and the complexity of the data. The use of machine learning (ML) and statistical predictions – i.e., artificial intelligence (AI) – in combination with process automation offers promising solutions for the optimisation of literature search, but these also come with a number of challenges.

Potential use cases of automated and AI-supported literature searches

• Automated database searches: Automated systems can continuously search scientific databases, such as PubMed, for relevant articles. They can apply complex search strategies to ensure that no crucial articles are overlooked.
• Text mining and information extraction: Using natural language processing (NLP), smart systems can automatically extract relevant information from scientific articles and identify adverse drug reactions, drug interactions and other safety-relevant data, such as dosage information.
• Automatic classification and prioritisation: ML algorithms can be used to classify articles and assess their relevance. This allows articles to be prioritised, which draws attention to crucial information when evaluating the search results. Intelligent pattern recognition can also assist in identifying and analysing case reports.
• Signal detection and trend analysis: By analysing data, statistical methods can help identify new or unexpected interactions and side effects and thereby track trends in drug safety.

Advantages of automated and AI-supported literature searches

• Increased efficiency and accuracy: Automated systems can automatically search and analyse large volumes of literature in a short time. Through ML and NLP, an AI can extract relevant information with high precision.
• Consistency: Automated systems work according to clearly defined algorithms, which leads to a consistent quality of search results.
• More comprehensive analyses: Learning algorithms enable the analysis of a larger amount of data, including data that may be overlooked in traditional searches. This can help recognise signals for drug risks more quickly.

Challenges of automated and AI-supported literature searches

Regulatory requirements and quality assurance: In a highly regulated area such as pharmacovigilance, automated systems and AI must be carefully developed and monitored. All regulatory requirements, e.g., Good Pharmacovigilance Practices (GVP), GAMP 5 or the AI act, must be met, and patient safety must be always ensured. Quality assurance measures, such as regular checks and updates of the software, must be implemented to assure that all relevant results are always delivered accurately.

Validation of automated and AI-assisted systems: Validation is a crucial step in ensuring reliable literature search results. Validation processes include the checking of software for accuracy, reliability, and consistency in its results. This is particularly important since decisions based on these searches can have an impact on patient safety.

Data quality and availability: The quality of the search results strongly depends on the quality of the underlying data. Learning systems should therefore have access to high-quality and current data. Incomplete or incorrect data can lead to inaccurate results. AI systems must be continuously updated and trained to keep pace with the latest developments and to improve their accuracy. For this, the quality of the training data plays an important role. If the data is not representative and comprehensive, the predictions and pattern recognition of the automated literature search can be similarly skewed and thus cannot sufficiently support the human experts. Seeing as incorrect information can jeopardise patient safety, this can have serious consequences in PV.

Interpretability: Decisions made based on AI-supported analyses must be comprehensible and verifiable. However, AI models often seem like "black boxes" where it is difficult to understand how they come to their conclusions.

Linguistic and semantic diversity: Scientific texts with PV data are often written in a highly specialised terminology and available in different languages. Systems must therefore not only be able to process texts in multiple languages, but also correctly interpret specialised terminology and context.

Integration into existing systems: The integration of AI-supported research tools into existing pharmacovigilance systems requires significant technological and organisational adjustments that involve effort and costs, including the modernisation of existing IT infrastructures, training staff, and the adjustment of workflows.

Conclusion:

Automated and AI-assisted systems can support and facilitate literature searches in pharmacovigilance in various aspects. However, the challenges associated with their implementation and use of these tools require careful planning, development, monitoring, and adaptation. This involves both technological innovation and close collaboration between drug safety professionals and computer scientists.