Articolo
Improving literature search quality with AI to reduce missed evidence and manual rework
-
Shilpa Patil, PhD
-
Angela Vogt-Eisele, PhD
Improving literature search quality with AI can help to reduce the risk of missing critical publications while filtering the noise
Reviewing scientific literature available in the public domain is a central requirement for preparation of numerous regulatory documents, including clinical and nonclinical overviews, orphan drug designations or pediatric investigational plans. These documents demand a comprehensive and up-to-date summary of information available in scientific literature.
Regulators may ask for a reproducible systematic literature search, meaning the applicant must be able to provide information on the exact databases/platforms searched, the full search strings (all keywords/controlled terms), the date ranges and other limits/filters applied, the date(s) the searches were run, and the record counts/results so the search can be repeated and verified.1,2
Depending on the subject, the amount of information retrieved in the primary search can be extensive and demands tedious and time-consuming manual searches. Attempts to use chatbot questions and answers have proven to be unreliable, while filter functions available in public databases are not sufficiently specific and present a risk of either returning too much noise (i.e. not efficiently narrowing down the information to the relevant articles) or missing key information.
Regulators may ask for a reproducible systematic literature search, meaning the applicant must be able to provide information on the exact databases/platforms searched, the full search strings (all keywords/controlled terms), the date ranges and other limits/filters applied, the date(s) the searches were run, and the record counts/results so the search can be repeated and verified.1,2
Depending on the subject, the amount of information retrieved in the primary search can be extensive and demands tedious and time-consuming manual searches. Attempts to use chatbot questions and answers have proven to be unreliable, while filter functions available in public databases are not sufficiently specific and present a risk of either returning too much noise (i.e. not efficiently narrowing down the information to the relevant articles) or missing key information.
Furthermore, chatbots do not provide verifiable answers, cannot ensure completeness, and also present challenges with traceability at a later stage. Artificial intelligence literature mining tools have the potential to address these issues. Being able to establish logical connections between terms, rather than filtering for articles merely mentioning a combination of certain terms, allows for specific, targeted searches on defined scientific questions and reduces the risk of missing relevant information.
Focused and thorough sets of literature allow regulatory teams to build well-informed, nuanced documents supporting the respective topic – such as indications and contraindications of a medicinal product, discussion of drug interactions or the safety margin of the product in defined patient populations.
How AI literature mining aids researchers in improving efficiency and sensitivity
1. Finding key articles among large volumes of literature
Implementing a semantic search tool can identify articles relevant to the desired topic with great accuracy and efficiency, which is crucial especially for searches with large numbers of primary hits.
In a pilot study, Cencora leveraged an internal and validated literature mining tool to search for drug–drug interactions involving active substance A and several co-administered compounds (B–E). The conventional keyword-based primary search retrieved more than 7,000 articles, largely driven by extensive publication on the well-described interaction of A with B.
By contrast, identifying interactions between A and C–E was difficult because the standard search engine primarily surfaced papers describing the interaction of A and B in which C–E appeared only as secondary mentions, not in the context of interaction with A.
Using the semantic search tool enabled a focused search for articles specifically describing drug-drug interactions of A with C, D and E separately. With this approach, it was possible to greatly narrow down the number of primary hits for manual verification – from more than 7,000 in an initial PubMed query to around 950. More importantly, the tool identified publications on sparsely studied and rare interactions that had been missed in a previous manual search.
Overall, this pilot illustrates the significant role AI can play in improving literature searches: it enables high throughput screening of very large sets of primary hits without relying on overly restrictive filters, and it improves precision by ranking records based on a logical relationship between concepts (e.g., an actual interaction between A and C/D/E) rather than simple co-occurrence of terms.
In a pilot study, Cencora leveraged an internal and validated literature mining tool to search for drug–drug interactions involving active substance A and several co-administered compounds (B–E). The conventional keyword-based primary search retrieved more than 7,000 articles, largely driven by extensive publication on the well-described interaction of A with B.
By contrast, identifying interactions between A and C–E was difficult because the standard search engine primarily surfaced papers describing the interaction of A and B in which C–E appeared only as secondary mentions, not in the context of interaction with A.
Using the semantic search tool enabled a focused search for articles specifically describing drug-drug interactions of A with C, D and E separately. With this approach, it was possible to greatly narrow down the number of primary hits for manual verification – from more than 7,000 in an initial PubMed query to around 950. More importantly, the tool identified publications on sparsely studied and rare interactions that had been missed in a previous manual search.
Overall, this pilot illustrates the significant role AI can play in improving literature searches: it enables high throughput screening of very large sets of primary hits without relying on overly restrictive filters, and it improves precision by ranking records based on a logical relationship between concepts (e.g., an actual interaction between A and C/D/E) rather than simple co-occurrence of terms.
Once the search has been narrowed, the identified material needs to be manually validated. AI greatly enhances efficiency of this task while additionally generating a concise inclusion or exclusion rationale for each article. Human reviewers can then confirm the decision – or identify false positives or false negatives – by checking the rationale against the relevant passages in the source text. Having a human in the loop is key to checking and refining the model.
2. Being able to query complex logical relationships
As outlined above, semantic understanding (natural language processing or NLP) is used for the literature search rather than simple keyword matching, allowing assessment of contextual relevance to eliminate the false positive results.
Moreover, the obtained pool of primary hits can further be searched for defined aspects – e.g. pharmacokinetics in different patient populations, specific safety aspects or the investigation of certain endpoints. This provides a structured output on the information available on the respective topic, linked to the original publication providing this information.
Moreover, the obtained pool of primary hits can further be searched for defined aspects – e.g. pharmacokinetics in different patient populations, specific safety aspects or the investigation of certain endpoints. This provides a structured output on the information available on the respective topic, linked to the original publication providing this information.
3. Auditability and traceability
AI also provides a full list of keywords and filters applied and the numbers of articles identified in each step as demanded for systematic literature searches for regulatory submissions. The inclusion and exclusion criteria as well as relevance criteria, mapped to the resulting articles, allows traceability for any future analysis and review. For example, it is possible to independently select articles studying “drug-drug interactions impacting absorption”, “drug-drug interactions impacting efficacy in indication X”, “drug-drug interactions associated with adverse events” etc.
Adding value to the process, the AI tool enabled an objective, repeatable justification for inclusion and exclusion, which can be confirmed or rejected by the human reviewer. Moreover, the condensed description of the reasoning for inclusion or exclusion of an article allows a fast and efficient manual verification of the selection by the expert and ensures the reasoning behind the inclusion and exclusion criteria are recorded for traceability and auditability.
The AI -generated audit trail into the search criteria and which articles were selected or excluded demonstrates a key value of AI: freeing experts to ensure proper oversight and manage full responsibility of the final literature selection.
Adding value to the process, the AI tool enabled an objective, repeatable justification for inclusion and exclusion, which can be confirmed or rejected by the human reviewer. Moreover, the condensed description of the reasoning for inclusion or exclusion of an article allows a fast and efficient manual verification of the selection by the expert and ensures the reasoning behind the inclusion and exclusion criteria are recorded for traceability and auditability.
The AI -generated audit trail into the search criteria and which articles were selected or excluded demonstrates a key value of AI: freeing experts to ensure proper oversight and manage full responsibility of the final literature selection.
4. Data structuring for predictive analysis and insights
AI also generates a systematic and structured analysis of key data elements from relevant articles, such as safety data by age group or a summary of pharmacokinetic drug-drug interactions across available studies. These structured data can be used to support e.g. the discussion of a medicinal product’s formulation or its safety in specific patient populations. Also, the level of evidence can be mapped based on article type and study size or design, etc.
Such structured, in-depth analysis of literature data can improve response readiness, particularly when regulators request clarification on safety context or supporting evidence.
Such structured, in-depth analysis of literature data can improve response readiness, particularly when regulators request clarification on safety context or supporting evidence.
Conclusion:
Improving literature searches and reducing regulatory deficiency risk
AI-assisted literature mining can play an important role in building depth and defensibility into the preparation of regulatory documents that require comprehensive, up-to-date scientific evidence from published literature. By enabling extensive primary literature screens, from which relevant articles are selected based on semantic understanding, this tool greatly enhances efficacy while reducing the risk of missing critical publications. Human oversight is available at all stages of the selection process.
Well-designed AI workflows also improve traceability and auditability by recording search strategies, inclusion and exclusion logic, and the rationale for article prioritization. When paired with human-in-the-loop validation, this approach supports consistent, repeatable decisions and reduces the variability inherent in manual screening. Finally, structuring extracted data elements across relevant publications creates a foundation for deeper analysis (e.g., subpopulation insights, evidence grading, trend detection) and strengthens regulatory narratives and response readiness.
Broadly speaking, AI-enabled literature mining can substantially improve the completeness and quality of the evidence base to support regulatory submissions, thus greatly reducing the risk of deficiencies in the review-cycle.
Well-designed AI workflows also improve traceability and auditability by recording search strategies, inclusion and exclusion logic, and the rationale for article prioritization. When paired with human-in-the-loop validation, this approach supports consistent, repeatable decisions and reduces the variability inherent in manual screening. Finally, structuring extracted data elements across relevant publications creates a foundation for deeper analysis (e.g., subpopulation insights, evidence grading, trend detection) and strengthens regulatory narratives and response readiness.
Broadly speaking, AI-enabled literature mining can substantially improve the completeness and quality of the evidence base to support regulatory submissions, thus greatly reducing the risk of deficiencies in the review-cycle.
*Le fonti continuano di seguito
About the authors:
Shilpa Patil, PhD, is a Senior Scientist in the digital innovation group at Cencora. She has more than eight years of experience in the field of translational research and healthcare, specializing in drug discovery, biomarker research, and AI-assisted personalized medicine in oncology. Shilpa and her team are working towards defining and building AI-based solutions that aim at adding value and efficiency to the regulatory applications and consultancy services offered.
Angela Vogt-Eisele, PhD, is Director CDS, Team Lead Global Medical Writing, at Cencora. Angela has worked for the pharmaceutical industry for more than 15 years supporting nonclinical and clinical development programs for human medicinal products. Angela and the global writing team have a strong track record in writing and updating a wide range of regulatory documents across modalities, from small molecules to biologics and AMTPs, and are currently implementing AI-enabled capabilities to take these activities to the next level in collaboration with the Cencora digital innovation group.
Angela Vogt-Eisele, PhD, is Director CDS, Team Lead Global Medical Writing, at Cencora. Angela has worked for the pharmaceutical industry for more than 15 years supporting nonclinical and clinical development programs for human medicinal products. Angela and the global writing team have a strong track record in writing and updating a wide range of regulatory documents across modalities, from small molecules to biologics and AMTPs, and are currently implementing AI-enabled capabilities to take these activities to the next level in collaboration with the Cencora digital innovation group.
Dichiarazione di non responsabilità:
Le informazioni fornite in questo articolo non costituiscono una consulenza legale. Cencora, Inc. incoraggia vivamente i lettori a rivedere le informazioni disponibili relative agli argomenti discussi e a fare affidamento sulla propria esperienza e competenza nel prendere decisioni correlate.
Mettiti in contatto con il nostro team
Il nostro team di esperti di valore leader si dedica a trasformare evidenze, approfondimenti sulle politiche e informazioni di mercato in efficaci strategie di accesso al mercato globale. Lascia che ti aiutiamo a navigare con sicurezza nel complesso panorama sanitario odierno. Contattaci per scoprire come possiamo supportare i tuoi obiettivi.
Sources
1. Guideline on good pharmacovigilance practices (GVP), HMA, EMA, December 2013. https://www.ema.europa.eu/en/documents/scientific-guideline/guideline-good-pharmacovigilance-practices-gvp-module-vii-periodic-safety-update-report_en.pdf
2. Guidance for Industry, Applications Covered by Section 505(b)(2), FDA, October 1999. https://www.fda.gov/media/72419/download
