Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Writing · pharmacovigilance literature surveillance

Literature Surveillance in Pharmacovigilance: Trends, Challenges and Prospects

Literature surveillance is a regulatory obligation with a fixed reporting timeline rather than a research activity. ICH E2D(R1) expects marketing authorisation holders to monitor the worldwide scientific literature using large reference databases, and an expedited individual case safety report (ICSR) is due no later than 15 calendar days after day zero, which, for a literature case, begins when sufficient information has been identified to meet the ICSR criteria. The principal difficulties of the task follow from this requirement.

Pubrica AcademyWritten by the medical writing and biostatistics teams Reviewed byPubrica biostatistics team Published9 October 2026 Last checked9 October 2026 Reading time16 minutes Sources56 peer-reviewed

Published case reports are not a marginal source of safety signals. Shah, in Drug Safety — Case Reports, found that adverse drug reaction case reports were used as evidence in 330 of 462 safety-related drug withdrawals worldwide between 1950 and 2014, or 71.4%. The literature is nonetheless the most difficult source to screen reliably: adverse effects are poorly indexed, duplicates propagate across systems, and the automation now being deployed is considerably better at identifying entities than at linking a drug to an event. This article examines what the obligation requires, where the problems of retrieval and de-duplication lie, what automation currently delivers, and how the process should be documented so that an inspector or reviewer can follow it.

Learning objectives

  • The requirements of the literature-monitoring obligation, and the point at which the 15-day clock starts
  • The coverage of the EMA Medical Literature Monitoring service, and the responsibilities it leaves with marketing authorisation holders
  • Why adverse effects are systematically hard to retrieve from bibliographic databases
  • How duplicate case reports distort signal detection, and the value of detecting them
  • The current performance of automated screening on recall, precision and relation extraction, and its limitations
  • How to document a surveillance process so that it is reproducible and inspectable

Common misconceptions

  • “Spontaneous reporting is where signals come from; the literature is supplementary.” Ochyra, Szewczyk and Przybylkowski, in Wien Klin Wochenschr, examined 4,160 individual case safety reports covering 73 drug-event associations and found about 33% came from the scientific literature. Literature ICSRs were 100% of the reports for two associations and more than 60% for three others, and the average delay between the first non-literature ICSR and the first literature ICSR was 38 months.
  • “If a trial reported adverse effects, a database search will find them.” Derry, Loke and Aronson showed in BMC Medical Research Methodology that of 81 clinical trials indexed in both MEDLINE and EMBASE, 25 — about 31% — were not indexed for adverse effects in either database. Only 66 of 107 trials reporting adverse effects, about 62%, mentioned them in the title or abstract, and combining textword and indexing terms retrieved only 82 of 107 papers, 77%.
  • “Duplicate reports are a tidiness problem.” They are a signal-detection problem. Norén, in Drug Safety, gives the arithmetic: a single report copied four times, listing seven drugs and four adverse events, could generate 84 false drug-event leads. The original vigiMatch implementation estimated its match weights from no more than 38 pairs of confirmed duplicates.
  • “Text mining has solved literature screening.” It has improved recall at the cost of precision, and relation extraction lags substantially behind entity extraction. Li, Liu and Yu, in JMIR Medical Informatics, reported named-entity recognition at an F1 of 84.1% but relation extraction at 65.9%, and medication-to-adverse-event relations at only 45.5%. Identifying the drug and the event in a sentence is not equivalent to establishing that one is reported as causing the other.
  • “If the EMA monitors a substance, the marketing authorisation holder’s literature obligation is discharged.” Only in part, and only for what the service covers. For the active substances in scope, marketing authorisation holders are not required to monitor that literature or to report the resulting suspected adverse reactions into EudraVigilance, and EMA makes the cases available so that holders can meet reporting obligations outside the European Economic Area (EEA). Everything outside the covered substance list, outside the screened databases and outside the EEA reporting framework remains the responsibility of the marketing authorisation holder.

Requirements of the literature-monitoring obligation

The text of the guideline is the appropriate starting point, because most process failures in literature surveillance arise from imprecise reading of it. ICH E2D(R1), adopted at Step 4 on 15 September 2025, states that each marketing authorisation holder is encouraged — and in some regions required — to regularly monitor the worldwide scientific literature using large reference databases with broad coverage 3,12. Screening frequency follows regional and local requirements rather than a single global interval.

Three operational details in the same guideline decide how a compliant process is built. The primary source for a literature case is the first listed author, or the corresponding author where one is specified. Expedited individual case safety reports are due as soon as possible and no later than 15 calendar days after day zero. And day zero for a literature case starts when sufficient information is identified to meet the ICSR criteria — an identifiable patient, an identifiable reporter, a suspect medicinal product and a suspected adverse reaction 3.

This last point is where compliance with the timeline is most often lost. Day zero is not the publication date, the date the search was run, or the date on which a reviewer read the abstract. It is the moment at which the four criteria can be satisfied from the information identified, which in practice is often the reading of the abstract rather than the full text. A process that logs retrieval dates but not the point at which the minimum criteria were met cannot demonstrate when the clock started, and this is typically the first question an inspector asks. The EU framework in Good Pharmacovigilance Practices Module VI sets out the collection, management and submission expectations within which this clock operates 6. See PSUR and DSUR writing and post-market surveillance for where the output goes.

Scope and limits of the EMA Medical Literature Monitoring service

EMA screens the medical literature centrally for a defined set of active substances and enters relevant information into EudraVigilance, which removes duplicated effort across holders of the same substance. When the service entered full operation, EMA described it as covering 400 active substance groups — 300 chemical and 100 herbal — on behalf of more than 4,000 companies 8.

The published coverage has changed since, and the figures are reported here as they stand rather than reconciled. The current service page lists active-substance groups numbered 1 to 309 and herbal-substance groups 1 to 100, valid from 1 June 2020, with 9 additional substances added in that June 2020 revision during the COVID-19 period, and states that EMA can monitor only a set number of substances based on its annual budget 5,27. The figure of 400 groups in the launch announcement and the enumeration of 309 plus 100 groups on the current list are therefore both accurate statements about different points in time, and any process built against the scope should be based on the live list rather than a remembered total.

The database coverage has also grown. The original detailed guide described three databases, one searched daily on working days and two monthly. The current description of the journal and reference databases used names four: EMBASE, International Pharmaceutical Abstracts, the Allied and Complementary Medicine Database and MEDLINE, with MEDLINE added on 1 April 2024 2. The declared journal coverage is substantial — Embase contributing some 3,200 additional journals, IPA 800 and AMED 500 2.

The critical point for a holder is what the service does not do. For the active substances it covers, holders are not required to monitor that literature or to report the resulting suspected adverse reactions into EudraVigilance, and EMA makes the cases available so holders can discharge reporting duties outside the EEA. Outside that scope — other substances, other databases, local and non-indexed literature, and non-EEA obligations — the duty is unchanged 5,6. Treating the service as a general exemption is the most common scope error observed in a surveillance SOP.

EMA Medical Literature Monitoring: what the published figures say, and when

AspectAt full operationOn the current published service description
Active substance groups covered400 groups in total — 300 chemical and 100 herbalActive-substance groups listed 1–309 and herbal-substance groups 1–100, valid from 1 June 2020, with 9 substances added in that revision
Companies relieved of the duty for covered substancesMore than 4,000Not restated as a single figure; scope is defined by the substance list
Databases screenedThree — one searched daily on working days, two monthlyFour — EMBASE, International Pharmaceutical Abstracts, AMED and MEDLINE, the last added on 1 April 2024
Declared journal coverage—Embase contributing around 3,200 additional journals, IPA 800 and AMED 500
Stated constraint on scope—EMA can monitor only a set number of substances, based on its annual budget
What it removes from the holderMonitoring of the covered literature and reporting of the resulting suspected adverse reactions into EudraVigilanceUnchanged; EMA makes the cases available so holders can meet reporting duties outside the EEA

Figures as published by the European Medicines Agency: the full-operation announcement for the launch column, and the medical literature monitoring service pages and the description of the journal and reference databases used for the current column. The two columns describe different dates and are not reconciled here on purpose — a surveillance scope should be built against the live substance list, not a remembered total.

The distinctive evidential weight of the published literature

The case for literature surveillance is not that it finds more cases than spontaneous reporting. It is that it finds different ones, sometimes earlier, and that it provides the narrative detail a signal needs to be assessable.

Shah, in Drug Safety — Case Reports, found that adverse drug reaction case reports were used as evidence in 330 of 462 safety-related drug withdrawals worldwide between 1950 and 2014 — 71.4% 46. The same work traces the cisapride sequence: after early case reports in 1995, the drug accumulated 341 reports of arrhythmias, including 80 fatalities, before its withdrawal in July 2000. Garcia-Doval and colleagues make the parallel argument for the value of case reports in British Journal of Dermatology, and Pontes, Clement and Rollason, in Drug Safety, present four practical examples in which a literature-detected signal affected a drug’s lifecycle, alongside an analysis of the screening requirements that pharmacovigilance guidelines impose 38,51.

Ochyra, Szewczyk and Przybylkowski, in Wien Klin Wochenschr, give the quantitative version. Across 4,160 individual case safety reports covering 73 drug-event associations, about 33% came from the scientific literature; literature ICSRs made up 100% of reports for two drug-event associations and more than 60% for three others; and the average delay between the first non-literature ICSR and the first literature ICSR was 38 months 5,8,33,47,50. This delay has implications in both directions: for most associations the spontaneous reporting system detected the association first and the literature followed years later, which supports literature surveillance as a means of corroboration and characterisation rather than as an early-warning system. For a small number of associations, however, the literature was the only available source. The pages on uses of case reports in medical literature and case report writing in evidence-based practice cover the reporting side of this.

The literature as a source of safety evidence

FindingFigureSource
Safety-related drug withdrawals worldwide, 1950–2014, in which ADR case reports were used as evidence330 of 462 — 71.4%Shah RR, Drug Safety — Case Reports
Cisapride arrhythmia reports accumulated after early case reports in 1995, before withdrawal in July 2000341 reports, including 80 fatalitiesShah RR, Drug Safety — Case Reports
Individual case safety reports originating from the scientific literatureAbout 33% of 4,160 ICSRs across 73 drug-event associationsOchyra B, Szewczyk M, Przybylkowski A, Wien Klin Wochenschr
Drug-event associations for which literature ICSRs were the only source2 associations at 100%; 3 further associations above 60%Ochyra B, Szewczyk M, Przybylkowski A, Wien Klin Wochenschr
Average delay from the first non-literature ICSR to the first literature ICSR38 monthsOchyra B, Szewczyk M, Przybylkowski A, Wien Klin Wochenschr
Worked examples of a literature-detected signal affecting a drug’s lifecycle4 cases analysedPontes H, Clement M, Rollason V, Drug Safety

Figures as reported in each publication. The 38-month delay is a reason to treat literature surveillance primarily as corroboration and characterisation; the two wholly literature-derived associations are the reason it cannot be treated as optional.

The retrieval problem: poor indexing of adverse effects

Literature surveillance has an irreducible retrieval problem that is not solved by searching harder. Adverse effects are frequently not the subject of the paper that reports them, so neither the indexing nor the abstract reliably flags them.

Derry, Loke and Aronson quantified this in BMC Medical Research Methodology, and the figures remain the clearest statement of the difficulty. Of 81 clinical trials with records in both MEDLINE and EMBASE, 25 — about 31% — were not indexed for adverse effects in either database 54. Only 66 of 107 trials that reported adverse effects, about 62%, mentioned them in the title or abstract 31. And simultaneous use of textword and indexing terms retrieved only 82 of the 107 papers, 77%.

Three consequences follow for the design and defence of a search strategy. First, a strategy that relies on adverse-effect indexing terms alone will systematically miss a third of relevant trial reports, so drug-plus-condition and drug-alone strategies must contribute to retrieval even though they return far more records to screen 54. Second, abstract-only screening is insufficient because of how records are structured, not because of a lack of diligence: if the effect is absent from the abstract in nearly four cases out of ten, full-text review must be triggered by a criterion other than a mention of harm in the abstract. Third, the requirement for at least two databases is substantive — the records not indexed for adverse effects in MEDLINE are not the same records as those missed in EMBASE, which is why the combined figure exceeds either alone yet remains incomplete. Search strategy peer review is the stage at which a surveillance strategy should be tested before it is incorporated into an SOP.

31%of trials indexed in both MEDLINE and EMBASE were not indexed for adverse effects in either (25 of 81)
62%of trials reporting adverse effects mentioned them in the title or abstract (66 of 107)
77%of papers were retrieved by combining textword and indexing terms (82 of 107)

Why adverse effects cannot be retrieved by indexing alone. Figures from Derry S, Loke YK, Aronson JK. Incomplete evidence: the inadequacy of databases in tracing published adverse drug reactions in clinical trials. BMC Medical Research Methodology.

Duplicate reports as a signal-detection problem

The same case can enter a safety database several times — from the reporter, from the holder, from a regulator, from a published report of the same patient, and from a review that re-describes it. Rule-based screening detects some of these duplicates, but those that remain are not harmless.

Norén, in Drug Safety, sets out the mechanism with an illustrative example: a single report copied four times, listing seven drugs and four adverse events, could generate 84 false drug-event leads 47. Disproportionality methods count co-occurrences, so a duplicated case inflates every drug-event pair it contains, and a case with many drugs and many events inflates a great many pairs at once. Norén also notes a limitation of the tooling that should temper confidence in it: the original vigiMatch implementation estimated its match weights from no more than 38 pairs of confirmed duplicates.

Tregunno and colleagues, in Drug Safety, evaluated a probabilistic duplicate-detection method against national centre data and the results show both the value and the heterogeneity. Across reports with enough information for the method to evaluate, 2.5% were flagged as suspected duplicates 28,50. The predictive value for confirmed or likely duplicates among flagged reports was 86% for the United Kingdom, 64% for Denmark and 33% for Spain 33,50. Of the confirmed duplicates, 89% were new to the Spanish centre, 60% to the UK centre and 38% to the Danish centre, despite existing rule-based screening. The variation across centres is the principal finding: performance depends on the data a centre holds and the screening it already performs, so a detection method should not be adopted on the basis of predictive values obtained elsewhere. It should be evaluated against the organisation’s own database and the result reported.

4copies of one report
7 × 4drugs and adverse events listed on it
84false drug-event leads this could generate

How one duplicated case inflates disproportionality analysis. Figures from Norén GN. The Power of the Case Narrative — Can it be Brought to Bear on Duplicate Detection? Drug Safety. The same paper notes that the original vigiMatch implementation estimated its match weights from no more than 38 pairs of confirmed duplicates.

Current performance of automation

Automation in literature surveillance is now routine rather than experimental, and the published performance is better than sceptics often assume but considerably lower than vendor descriptions imply. Salas and colleagues, in Pharmaceutical Medicine, searched Embase and MEDLINE for work published from 1 January 2015 to 9 July 2021, identified 66 articles for evaluation, and found that about 57.6% applied AI to identifying adverse drug events and adverse drug reactions — indicating that the application is concentrated in the screening task examined in this article 27,31,54.

The best-case results are of practical value. The LiSA assisted literature search pipeline, reported by Martenot and colleagues in BMC Medical Informatics and Decision Making, reached 81.1% precision, 88.6% recall and an F1-score of 84.7% at sentence level on PubMed articles, with 78.5% precision at document level, benchmarked on 1,231 sentences from 988 unique documents; a medical reviewer using it found about 2.5 times as many relevant documents as with a simple PubMed keyword search 28,50. Tafti and colleagues, in JMIR Medical Informatics, demonstrated feasibility at a scale no manual process reaches: 97,246,719 sentences from 1,451,413 articles screened, with 92.7% accuracy, 93.6% precision, 93.0% recall and an AUC of 0.905 45.

Three findings qualify these results. Relation extraction lags behind entity extraction: Li, Liu and Yu reported named-entity recognition at F1 84.1% but relation extraction at 65.9%, and medication-to-adverse-event relations at only 45.5% 42. Precision in realistic screening settings is low: Eskildsen and colleagues, in BMC Medical Informatics and Decision Making, benchmarked two text-mining methods against a standard pharmacovigilance screening method and found recall of 90.4%, 88.1% and 78.5% respectively, with precision of 3.4%, 1.9% and 1.6% — a consequence of base rates, since of 2,533 manually annotated individual case safety reports only 42 were true cases, with inter-annotator agreement at a Cohen’s kappa of 0.96 37. And performance degrades outside the training distribution: Yu and Vydiswaran, in JMIR Medical Informatics, reported a validation F1 of 0.625 falling to 0.51 and 0.394 on two external datasets 26.

The areas of improvement are sufficiently clear to inform planning. Kopacheva and colleagues, in JMIR Formative Research, show fine-tuned clinical language models reaching an NER F1 of 0.845 and a relation macro-F1 of 0.813 against a classical baseline of 0.281, with an adverse-drug-event relation F1 of 0.704 16. Balogh and colleagues, in Briefings in Bioinformatics, found that purpose-built adverse-event embeddings trained on around 15 million FAERS reports reached an AUROC of 0.88 for drug-event association prediction, against 0.76 to 0.83 for general large language model embeddings — a finding that challenges the assumption that a general model is the strongest option 3,12. Combination also outperforms any single component: Xu and Wang, in BMC Bioinformatics, extracted 2,787,797 drug-side effect pairs from FAERS at an initial precision of 0.025, and combining the signals with MEDLINE raised top-ranked precision to 0.371, about a 13.8-fold increase 49. Only 9.6% of the FAERS pairs — 269,040 of 2,787,797 — appeared in MEDLINE abstracts at all, and manual curation of 617 drug-cardiovascular pairs found in both sources gave a precision of 0.519 with 320 true positives, of which 258, or 80.6%, were absent from the SIDER resource. Jeong and colleagues, in PLoS ONE, make the same point for signal-detection algorithms: combining the outputs of several existing algorithms outperformed any single one, with sensitivity from 0.593 to 0.793 and AUROC from 0.737 to 0.816 43. See the challenges of delivering clinical impact with machine learning for the validation framing these results sit inside.

Reported performance of automated approaches to adverse-event detection in text

Approach and sourceRecall / sensitivityPrecisionOther reported figures
LiSA assisted literature search pipeline — Martenot and colleagues, BMC Medical Informatics and Decision Making88.6% at sentence level81.1% at sentence level; 78.5% at document levelF1 84.7%; benchmark of 1,231 sentences from 988 documents; about 2.5× the relevant documents found versus simple keyword search
Large-scale literature mining — Tafti and colleagues, JMIR Medical Informatics93.0%93.6%97,246,719 sentences from 1,451,413 articles; 92.7% accuracy; AUC 0.905
End-to-end extraction from EHR notes — Li, Liu and Yu, JMIR Medical Informatics——NER F1 84.1%; relation extraction F1 65.9%; medication-to-adverse-event relation F1 45.5%
Fine-tuned clinical language models — Kopacheva and colleagues, JMIR Formative Research——NER F1 0.845; relation macro-F1 0.813 against a baseline of 0.281; ADE relation F1 0.704
Text mining versus standard screening — Eskildsen and colleagues, BMC Medical Informatics and Decision Making90.4% (CPR), 88.1% (standard), 78.5% (I2E)3.4% (standard), 1.9% (CPR), 1.6% (I2E)2,533 annotated ICSRs containing only 42 true cases; inter-annotator Cohen’s kappa 0.96
Social media ADE detection, internal then external — Yu and Vydiswaran, JMIR Medical Informatics——Validation F1 0.625, falling to 0.51 and 0.394 on two external datasets; 1,544 ADE tweets from 34,536 screened
Purpose-built adverse-event embeddings — Balogh and colleagues, Briefings in Bioinformatics——AUROC 0.88 from around 15 million FAERS reports, against 0.76–0.83 for general large language model embeddings

Figures exactly as reported in each publication. The recall and precision columns are not comparable across rows: the tasks, units of analysis and base rates differ, and the Eskildsen precision figures in particular reflect a setting in which only 42 of 2,533 annotated reports were true cases.

Inherited limitations: quality of case documentation

A surveillance process cannot be better than the cases it collects, and the documentation quality of individual case safety reports is poor and has been deteriorating.

The vigiGrade work on identifying well-documented individual case reports assessed around 7 million reports and found a mean completeness score of 0.45, with only 13% classified as well documented 56. The difference by reporter type is marked: 24% of physician reports against 4% of consumer reports met the well-documented threshold. Rolfes and colleagues, in Drug Safety, compared the clinical information quality of patient and healthcare professional reports directly, and this is an appropriate companion reading — the point is not that patient reports are less valuable but that they carry different information and require different handling.

Automated coding reflects the same asymmetry. A validation of artificial intelligence for automatic coding of patient adverse drug reaction reports using nationwide pharmacovigilance data reported an F-measure of 0.82 and an AUC of 0.97 for identifying the reaction, but only an F-measure of 0.65 against an AUC of 0.87 for seriousness 55. Reaction identification approaches usable performance; seriousness assessment — the determination that drives expedited reporting — does not. This asymmetry should determine which steps are automated and where a human decision remains mandatory, and it is arguably the most important design constraint in a modern surveillance workflow.

Documenting the process for inspection

A literature-surveillance process is judged on whether another person could reproduce the work performed on a given day, which is the same standard applied to a systematic review search. PRISMA-S, with its sixteen items, is a useful template even though it was written for reviews: it requires that databases, search dates, full search strategies and supplementary search methods all be specified, and these are exactly the components a documented surveillance process must record 35.

Translated into surveillance terms, the record for each cycle needs: the databases and interfaces searched; the full search strings as executed, with their version history, because strings are revised and an audit covers the string in force on the day; the date and time of each run and the coverage window; the number of records retrieved per source; who screened at abstract and at full text, with the criteria applied; the date on which minimum ICSR criteria were met for each case, which is day zero; the submission date against the 15-day deadline; the de-duplication method and its outcome; and the local, non-indexed and non-database sources covered separately 3,12.

Two details should be incorporated from the outset. The reason for every exclusion at full-text stage should be recorded, not only the decision — this is the part of the file that demonstrates that judgement was exercised rather than that records were discarded. The scope rationale should also be maintained as a live document: which substances are in scope, which are covered by the EMA service and which are not, which territories impose their own requirements, and the date each determination was last reviewed. Scope drift is an inconspicuous failure mode, because nothing in the daily process signals that a new product or indication has fallen outside the standing strategy. Regulatory writing and health authority response support are the contexts in which this documentation is tested most rigorously.

The literature-surveillance section of a periodic safety report

Asserts a process

A comprehensive review of the scientific literature was performed throughout the reporting interval to identify any published reports of adverse reactions associated with the product. Relevant publications were reviewed by the pharmacovigilance team and valid cases were processed in accordance with company procedures and applicable regulatory timelines. No new safety concerns were identified.

“Comprehensive” is unverifiable. No databases, no interfaces, no strings, no frequency — nothing an inspector can re-run or an auditor can check against a date.

No day zero. The one date that governs compliance is absent, so there is no way to demonstrate the 15-day clock was met rather than asserted.

“In accordance with company procedures” points at a document instead of stating what happened. It is the regulatory equivalent of citing a guideline in place of a method.

No counts, no screening record, no exclusion reasons. The volume screened and the judgements made are both invisible, so the absence of findings cannot be distinguished from the absence of screening.

No de-duplication and no scope statement. Nothing says which substances are covered by the EMA service and which the holder monitors itself, which is where scope drift becomes a finding.

“No new safety concerns were identified” is unsupported. On this record it is a claim about a process the reader cannot see.

Documents a process

Literature screening was performed against EMBASE (Ovid) and MEDLINE (PubMed) on each working day throughout the interval, with International Pharmaceutical Abstracts and AMED screened monthly; the active-substance search strings as executed, with their version history and the date each revision took effect, are provided in Annex 4. The weekly retrieval counts per source and the screening decisions at abstract and full-text level, with the reason recorded for each exclusion, are in Annex 5. For each case meeting the minimum criteria, day zero was recorded as the date on which an identifiable patient, an identifiable reporter, a suspect product and a suspected adverse reaction were all identifiable from the material reviewed; the primary source was taken as the first listed author, or the corresponding author where specified. All 11 valid individual case safety reports were submitted within 15 calendar days of day zero, with a median of 6 days; the submission date against day zero for each is tabulated in Annex 6. Duplicate screening was applied across the interval using the method described in Annex 7, which was evaluated against this database and identified 3 suspected duplicates, 2 of which were confirmed. Local and non-indexed sources screened outside the four databases, and the substances within the scope of the EMA Medical Literature Monitoring service for which the holder does not monitor or report into EudraVigilance, are listed in Annex 8 with the date the scope determination was last reviewed.

Sources are named with interfaces and screening frequency, and the executed strings with version history are annexed — so the search in force on any given day can be reconstructed.

Day zero is defined operationally, against the four minimum criteria, and recorded per case; the 15-day performance is reported with a median and tabulated per report.

The primary-source rule is stated explicitly, which matters because it determines whose publication the case is attributed to.

Retrieval counts, screening decisions at both levels and the reason for each exclusion are available, so a nil finding is evidence rather than an assertion.

The duplicate method is named and — critically — reported as evaluated against this database, which is what the wide between-centre variation in published predictive values requires.

Scope is stated in both directions, with the review date, so what the holder monitors and what the EMA service covers are separable at audit.

Everything in the second version already exists if the process was run properly; recording it requires only an annex and addresses the inspector’s central question, which is whether the process described is the process performed.

Future directions and realistic expectations

Three directions are well enough evidenced to plan for, and each comes with a stated limit.

Assisted screening will keep improving and will not remove the reviewer. The combination that is effective at present is high-recall automated retrieval feeding a human decision, which is what LiSA’s 2.5-fold increase in relevant documents found by a medical reviewer describes — a reviewer working more effectively, not a reviewer replaced 28,50. The precision figures from realistic benchmarks explain why: at 3.4% precision for the best-performing method in the Eskildsen comparison, a fully automated pipeline would route a very large volume of false positives into a regulated decision 37.

Multi-source combination will outperform any single stream. The FAERS-plus-MEDLINE result — a 13.8-fold precision increase for top-ranked pairs — and the algorithm-ensemble result from Jeong and colleagues point the same way 43,49. The limit is that only 9.6% of FAERS pairs appeared in MEDLINE abstracts at all, so combination sharpens what both sources see and does nothing for what neither sees.

Electronic health records are a promised source that has not yet arrived. Golder and colleagues’ scoping review in Drug Safety searched six databases in July 2023 for NLP and machine learning applied to adverse drug event detection in electronic health and medical records, and found only seven studies, reported across 10 publications from 2013 to 2023, that met the inclusion criteria; area under the ROC curve for a drug-adverse event association model was 75.3% unadjusted and 80.4% adjusted for confounders 17. Seven studies in a decade is a thin evidence base for a data source frequently described as transformative. Kim and colleagues, in Medicine, mapped analytic method to data source across 72 studies published between 2015 and 2020, which is the better place to start than any single demonstration 27,31. In summary, the methods are maturing faster than the external validation behind them, and the gap between internal and external performance — an F1 of 0.625 falling to 0.394 in the Yu and Vydiswaran results — is the figure that should inform any proposal to deploy one 26.

Key points

Key points

Literature surveillance is an obligation with an operational timeline. ICH E2D(R1), adopted at Step 4 on 15 September 2025, expects regular monitoring of the worldwide scientific literature in large reference databases, with expedited reports due no later than 15 calendar days after day zero; day zero begins when sufficient information has been identified to meet the ICSR criteria, not when the paper was published. The task is difficult for structural reasons: Derry and colleagues found that 31% of trials indexed in both MEDLINE and EMBASE were not indexed for adverse effects in either, and only 62% of trials reporting adverse effects mentioned them in the title or abstract. It matters because the literature carries evidential weight that no other source does: adverse drug reaction (ADR) case reports were evidence in 330 of 462 safety-related withdrawals from 1950 to 2014, and about 33% of 4,160 ICSRs across 73 drug-event associations came from the literature, with two associations wholly literature-derived. The EMA Medical Literature Monitoring service removes part of the obligation for covered substances and none of it outside them. Automation assists at the retrieval stage but should not be relied upon at the decision stage: relation extraction performs well below entity extraction, precision in realistic screening settings is in the low single figures, and external performance falls sharply relative to internal validation.

Application to your study

Application to your study

  1. Day zero should be defined operationally in the standard operating procedure (SOP). For each case, the record should state the date on which an identifiable patient, reporter, suspect product and suspected reaction were all identifiable from the material reviewed, rather than the publication or retrieval date, and the submission date should be reported against it.
  2. The search should be designed to withstand the indexing gap. Adverse-effect index terms should not be relied upon alone; drug-alone and drug-plus-condition strands should be included, at least two bibliographic databases used, and full-text review triggered by criteria other than the appearance of harm in the abstract.
  3. Search strings should be version-controlled and every run logged. The log should record the string as executed, the date each revision took effect, records retrieved per source, screening decisions at abstract and full-text stages, and the reason for every exclusion.
  4. Duplicate detection should be evaluated against the organisation’s own database before it is relied upon. Published predictive values among flagged reports ranged from 33% to 86% across national centres, so figures from other settings do not transfer; the performance achieved locally should be reported.
  5. Retrieval, rather than seriousness assessment, should be automated. Automated coding performs well in identifying the reaction and poorly in assessing seriousness, and seriousness determines the expedited timeline. That decision should remain with human reviewers, and the SOP should state this.
  6. Scope should be reviewed on a stated schedule, with the review date recorded. The review should establish which substances fall within the EMA Medical Literature Monitoring service, which territories impose separate obligations, and which local or non-indexed sources are screened directly.

Frequently asked questions

It is the systematic, scheduled screening of published scientific literature for reports of suspected adverse reactions to a company’s products, and the processing of qualifying reports as individual case safety reports. ICH E2D(R1), adopted at Step 4 on 15 September 2025, states that each marketing authorisation holder is encouraged, and in some regions required, to regularly monitor the worldwide scientific literature using large reference databases with broad coverage, with screening frequency following regional and local requirements. In the EU, the collection, management and submission expectations sit in Good Pharmacovigilance Practices Module VI. It is distinct from a literature review: the output is regulatory case reports against a deadline, not a synthesis, and the process has to be documented well enough to be inspected.

Day zero for a literature case starts when sufficient information is identified to meet the individual case safety report criteria — an identifiable patient, an identifiable reporter, a suspect medicinal product and a suspected adverse reaction. It is not the publication date, not the date the search was run, and not the date a full text was obtained. In practice this often means the moment a reviewer reads an abstract containing all four elements, which is why a process that logs only retrieval dates cannot demonstrate when the clock began. Expedited reports are due as soon as possible and no later than 15 calendar days after day zero under ICH E2D(R1). The primary source is taken as the first listed author, or the corresponding author where one is specified.

Only for the active substances it covers, and only within the EU reporting framework. For substances in scope, marketing authorisation holders are not required to monitor that literature or to report the resulting suspected adverse reactions into EudraVigilance, and EMA makes the cases available so holders can meet reporting duties outside the EEA. Everything else remains the responsibility of the holder: substances not on the list, databases the service does not screen, local and non-indexed literature, and obligations in other territories. The scope also changes — the service was announced at full operation as covering 400 active substance groups on behalf of more than 4,000 companies, while the current published list enumerates active-substance groups 1 to 309 and herbal groups 1 to 100, and EMA states it can monitor only a set number of substances based on its annual budget. Scope should be built against the live list.

At least two bibliographic databases, and in the EU context the databases used by the regulator itself are informative. The EMA service’s current description names four — EMBASE, International Pharmaceutical Abstracts, the Allied and Complementary Medicine Database and MEDLINE, the last added on 1 April 2024 — with declared journal coverage including some 3,200 additional journals from Embase, 800 from IPA and 500 from AMED. The original detailed guide described three databases, one searched daily on working days and two monthly, which indicates a defensible frequency pattern. Two databases constitute a minimum rather than a preference because Derry, Loke and Aronson found 31% of trials indexed in both MEDLINE and EMBASE were not indexed for adverse effects in either, and the records each database misses are not the same records.

Because they are frequently not what the paper is about, so neither the indexing nor the abstract flags them. Derry, Loke and Aronson quantified this in BMC Medical Research Methodology: of 81 clinical trials with records in both MEDLINE and EMBASE, 25 — about 31% — were not indexed for adverse effects in either database; only 66 of 107 trials reporting adverse effects, about 62%, mentioned them in the title or abstract; and simultaneous use of textword and indexing terms retrieved only 82 of 107 papers, 77%. The practical consequence is that a strategy built on adverse-effect index terms will miss roughly a third of relevant trial reports, and that abstract-level screening alone is insufficient as a matter of how records are structured rather than how carefully they are read.

Enough to distort signal detection measurably. Norén, in Drug Safety, gives the mechanism: a single report copied four times, listing seven drugs and four adverse events, could generate 84 false drug-event leads, because disproportionality methods count co-occurrences and a duplicated multi-drug case inflates many pairs simultaneously. Detection methods work but transfer poorly between settings. Tregunno and colleagues found that 2.5% of evaluable reports were flagged as suspected duplicates, with predictive value for confirmed or likely duplicates among flagged reports of 86% in the United Kingdom, 64% in Denmark and 33% in Spain — and of the confirmed duplicates, 89% were new to the Spanish centre, 60% to the UK centre and 38% to the Danish centre despite existing rule-based screening. Any method should be evaluated against the organisation’s own database.

Not at the decision stage, on current evidence. Performance at the retrieval stage is strong: the LiSA pipeline reached 88.6% recall and 81.1% precision at sentence level with an F1 of 84.7%, and a medical reviewer using it found about 2.5 times as many relevant documents as with a simple keyword search; Tafti and colleagues screened 97,246,719 sentences from 1,451,413 articles at 92.7% accuracy. However, relation extraction lags substantially behind entity extraction — Li, Liu and Yu reported NER at F1 84.1% against medication-to-adverse-event relations at 45.5% — precision in realistic settings is low, with Eskildsen and colleagues reporting 3.4%, 1.9% and 1.6% for three methods, and external performance drops sharply, with one validation F1 of 0.625 falling to 0.394 on an external dataset. High-recall retrieval feeding a human decision is the configuration the evidence supports.

Sufficient detail for another person to reproduce the process as performed on a given day. Specifically: databases and interfaces; full search strings as executed, with version history and the date each revision took effect; the date, time and coverage window of every run; records retrieved per source; who screened at abstract and full text and against what criteria; the recorded reason for each full-text exclusion; day zero per case, defined against the four minimum ICSR criteria; submission date against the 15-day deadline; the de-duplication method and its evaluated performance on the organisation’s own data; the local and non-indexed sources screened outside the databases; and the scope determination — which substances the EMA service covers and which the holder monitors itself — with the date it was last reviewed. PRISMA-S, with its sixteen items on sources, dates, strategies and supplementary methods, is a serviceable template.

References

  1. European Medicines Agency. EudraVigilance: electronic reporting. European Medicines Agency. 2026. read 9 October 2026. ema.europa.eu
  2. European Medicines Agency. Monitoring of medical literature and the entry of relevant information into the EudraVigilance database by the European Medicines Agency: Description of the Journal/Reference databases used. European Medicines Agency. 2026;EMA/141813/2015 Rev 5. read 9 October 2026. ema.europa.eu
  3. International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH). ICH Harmonised Guideline E2D(R1): Post-Approval Safety Data: Definitions and Standards for Management and Reporting of Individual Case Safety Reports (Step 4, adopted 15 September 2025). International Council for Harmonisation. 2025. read 9 October 2026. database.ich.org
  4. European Medicines Agency. EudraVigilance system overview. European Medicines Agency. 2024. read 9 October 2026. ema.europa.eu
  5. European Medicines Agency. Medical literature monitoring. European Medicines Agency. 2024. read 9 October 2026. ema.europa.eu
  6. European Medicines Agency and Heads of Medicines Agencies. Guideline on good pharmacovigilance practices (GVP) Module VI: Collection, management and submission of reports of suspected adverse reactions to medicinal products (Rev 2). European Medicines Agency. 2017;EMA/873138/2011 Rev 2. read 9 October 2026. ema.europa.eu
  7. European Medicines Agency, Inspections and Human Medicines Pharmacovigilance Division. Detailed guide regarding the monitoring of medical literature and the entry of relevant information into the EudraVigilance database by the European Medicines Agency. European Medicines Agency. 2015;EMA/161530/2014. read 9 October 2026. ema.europa.eu
  8. European Medicines Agency. EMA's medical literature monitoring enters into full operation. European Medicines Agency. 2015. read 9 October 2026. ema.europa.eu
  9. Brosch S. EMA Medical Literature Monitoring (MLM) Service. 8th Stakeholder Forum, Session 3 - Process improvement for public health and product lifecycle support. European Medicines Agency. 2014. read 9 October 2026. ema.europa.eu
  10. International Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use. ICH Harmonised Tripartite Guideline E2E: Pharmacovigilance Planning. International Council for Harmonisation. 2004;E2E(Step 4). read 9 October 2026. database.ich.org
  11. International Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use. E2E: Pharmacovigilance Planning (Final Concept Paper). International Council for Harmonisation. 2002;E2E. read 9 October 2026. database.ich.org
  12. Balogh OM, Petervari M, Csernak AM, Puhl E, Horvath A, Ferdinandy P, Agg B. Contrastive learning of adverse events to provide effective and interpretable vector representations for machine-assisted pharmacovigilance. Briefings in Bioinformatics. 2026;27(5):bbag463. doi:10.1093/bib/bbag463 read 9 October 2026. doi.org/10.1093/bib/bbag463
  13. Elhawary MA, Edwards B, Rostom H, Pal SN, Macolic Sarinic V, Bahri P, McCarthy D, Nambasa VP, Harmark L, Bate A, et al. Global Perspective on the Evolution and Future of Pharmacovigilance: Deliverables from the 24th Annual Meeting of the International Society of Pharmacovigilance Celebrating 25 Years of Excellence. Drug Safety. 2026;49(7):733-745. doi:10.1007/s40264-026-01680-8 read 9 October 2026. doi.org/10.1007/s40264-026-01680-8
  14. Xu S, Zhao Z, Liu X, Meng X-L. A comparative study of screening performance between abstrackr and GPT models: Systematic review and contextual analysis. BMC Medical Informatics and Decision Making. 2025;25(1):293. doi:10.1186/s12911-025-03138-w read 9 October 2026. doi.org/10.1186/s12911-025-03138-w
  15. Algarvio RC, Conceicao J, Rodrigues PP, Ribeiro I, Ferreira-da-Silva R. Artificial intelligence in pharmacovigilance: a narrative review and practical experience with an expert-defined Bayesian network tool. International Journal of Clinical Pharmacy. 2025;47(4):932-944. doi:10.1007/s11096-025-01975-3 read 9 October 2026. doi.org/10.1007/s11096-025-01975-3
  16. Kopacheva E, Henriksson A, Dalianis H, Hammar T, Lincke A. Identifying Adverse Drug Events in Clinical Text Using Fine-Tuned Clinical Language Models: Machine Learning Study. JMIR Formative Research. 2025;9(1):e71949. doi:10.2196/71949 read 9 October 2026. doi.org/10.2196/71949
  17. Golder S, Xu D, O'Connor K, Wang Y, Batra M, Gonzalez Hernandez G, et al. Leveraging Natural Language Processing and Machine Learning Methods for Adverse Drug Event Detection in Electronic Health/Medical Records: A Scoping Review. Drug Safety. 2025;48(4):321-337. doi:10.1007/s40264-024-01505-6 read 9 October 2026. doi.org/10.1007/s40264-024-01505-6
  18. Booth A, Sutton A, Clowes M, Martyn-St James M. Systematic Approaches to a Successful Literature Review (4th edition). 2025;4th edition. read 9 October 2026. collegepublishing.sagepub.com
  19. Webb P, Bain C, Page A. Essential Epidemiology (5th edition). 2024;5th edition. read 9 October 2026. swslhd.libguides.com
  20. Greenhalgh T, Dijkstra P. How to Read a Paper: The Basics of Evidence-Based Healthcare (7th edition). 2024;7th edition. read 9 October 2026. swslhd.libguides.com
  21. Harmsen W, de Groot J, Harkema A, van Dusseldorp I, de Bruin J, van den Brand S, van de Schoot R. Machine learning to optimize literature screening in medical guideline development. Systematic Reviews. 2024;13(1):177. doi:10.1186/s13643-024-02590-5 read 9 October 2026. doi.org/10.1186/s13643-024-02590-5
  22. Aldahdooh J, Tanoli Z, Tang J. Mining drug-target interactions from biomedical literature using chemical and gene descriptions-based ensemble transformer model. Bioinformatics Advances. 2024;4(1):vbae106. doi:10.1093/bioadv/vbae106 read 9 October 2026. doi.org/10.1093/bioadv/vbae106
  23. Marusic A (editor). A Guide to Responsible Research. 2023. read 9 October 2026. swslhd.libguides.com
  24. Application of artificial intelligence and machine learning in early detection of adverse drug reactions (ADRs) and drug-induced toxicity. Artificial Intelligence Chemistry. 2023. read 9 October 2026. doaj.org
  25. Jagadeesh G, Pitchai B, Senatore F (editors). The Quintessence of Basic and Clinical Research and Scientific Publishing. 2023. read 9 October 2026. swslhd.libguides.com
  26. Yu D, Vydiswaran VGV. An Assessment of Mentions of Adverse Drug Events on Social Media With Natural Language Processing: Model Development and Analysis. JMIR Medical Informatics. 2022;10(9):e38140. doi:10.2196/38140 read 9 October 2026. doi.org/10.2196/38140
  27. Kim HR, Sung M, Park JA, Jeong K, Kim HH, Lee S, Park YR. Analyzing adverse drug reaction using statistical and machine learning methods. Medicine. 2022;101(25):e29387. doi:10.1097/MD.0000000000029387 read 9 October 2026. doi.org/10.1097/MD.0000000000029387
  28. Martenot V, Masdeu V, Cupe J, Gehin F, Blanchon M, Dauriat J, Horst A, Renaudin M, Girard P, et al. LiSA: an assisted literature search pipeline for detecting serious adverse drug events with deep learning. BMC Medical Informatics and Decision Making. 2022;22(1):338. doi:10.1186/s12911-022-02085-0 read 9 October 2026. doi.org/10.1186/s12911-022-02085-0
  29. Price C. Syntheses Synthesized: A Look Back at Grant and Booth's Review Typology. Evidence Based Library and Information Practice. 2022;17(2). doi:10.18438/eblip30093 read 9 October 2026. doi.org/10.18438/eblip30093
  30. Khan KS, Zamora J. Systematic Reviews to Support Evidence-Based Medicine (3rd edition). 2022;3rd edition. read 9 October 2026. swslhd.libguides.com
  31. Salas M, Petracek J, Yalamanchili P, Aimer O, Kasthuril D, Dhingra S, Junaid T, Bostic T. The Use of Artificial Intelligence in Pharmacovigilance: A Systematic Review of the Literature. Pharmaceutical Medicine. 2022;36(5):295-306. doi:10.1007/s40290-022-00441-z read 9 October 2026. doi.org/10.1007/s40290-022-00441-z
  32. Dasgupta S, Jayagopal A, Jun Hong AL, Mariappan R, Rajan V. Adverse Drug Event Prediction Using Noisy Literature-Derived Knowledge Graphs: Algorithm Development and Validation. JMIR Medical Informatics. 2021;9(10):e32730. doi:10.2196/32730 read 9 October 2026. doi.org/10.2196/32730
  33. Ochyra B, Szewczyk M, Przybylkowski A. Impact of literature reports on drug safety signals. Wien Klin Wochenschr. 2021;133:188-193. doi:10.1007/s00508-020-01677-y read 9 October 2026. doi.org/10.1007/s00508-020-01677-y
  34. Sardella M, Belcher G, Lungu C, Ignoni T, Camisa M, Stenver DI, Porcelli P, D'Antuono M, Castiglione NG, Adams A, et al. Monitoring the manufacturing and quality of medicines: a fundamental task of pharmacovigilance. Therapeutic Advances in Drug Safety. 2021;12. doi:10.1177/20420986211038436 read 9 October 2026. doi.org/10.1177/20420986211038436
  35. Rethlefsen ML, Kirtley S, Waffenschmidt S, Ayala AP, Moher D, Page MJ, Koffel JB; PRISMA-S Group. PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews. Systematic Reviews. 2021;10(1):39. doi:10.1186/s13643-020-01542-z read 9 October 2026. doi.org/10.1186/s13643-020-01542-z
  36. Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, Shamseer L, Tetzlaff JM, Akl EA, Brennan SE, Chou R, Glanville J, Grimshaw JM, Hrobjartsson A, Lalu MM, Li T, Loder EW, Mayo-Wilson E, McDonald S, McGuinness LA, Stewart LA, Thomas J, Tricco AC, Welch VA, Whiting P, Moher D. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. doi:10.1136/bmj.n71 read 9 October 2026. doi.org/10.1136/bmj.n71
  37. Eskildsen NK, Eriksson R, Christensen SB, Aghassipour TS, Bygso MJ, Brunak S, Hansen SL. Implementation and comparison of two text mining methods with a standard pharmacovigilance method for signal detection of medication errors. BMC Medical Informatics and Decision Making. 2020;20(1):94. doi:10.1186/s12911-020-1097-0 read 9 October 2026. doi.org/10.1186/s12911-020-1097-0
  38. Garcia-Doval I, Segovia E, Hunter H, Frew J, Naldi L. The value of case reports in pharmacovigilance. British Journal of Dermatology. 2020;183(5):795-796. doi:10.1111/bjd.19504 read 9 October 2026. doi.org/10.1111/bjd.19504
  39. Norman CR, Leeflang MMG, Porcher R, et al., Neveol A. Measuring the impact of screening automation on meta-analyses of diagnostic test accuracy. Systematic Reviews. 2019;8(1):243. doi:10.1186/s13643-019-1162-x read 9 October 2026. doi.org/10.1186/s13643-019-1162-x
  40. Sutton A, Clowes M, Preston L, Booth A. Meeting the review family: exploring review types and associated information retrieval requirements. Health Information and Libraries Journal. 2019;36(3):202-222. doi:10.1111/hir.12276 read 9 October 2026. doi.org/10.1111/hir.12276
  41. Schulz K, Grimes DA. Essential Concepts in Clinical Research: Randomised Controlled Trials and Observational Epidemiology (2nd edition). 2018;2nd edition. read 9 October 2026. swslhd.libguides.com
  42. Li F, Liu W, Yu H. Extraction of Information Related to Adverse Drug Events from Electronic Health Record Notes: Design of an End-to-End Model Based on Deep Learning. JMIR Medical Informatics. 2018;6(4):e12159. doi:10.2196/12159 read 9 October 2026. doi.org/10.2196/12159
  43. Jeong E, Park N, Choi Y, Park RW, Yoon D. Machine learning model combining features from algorithms with different analytical methodologies to detect laboratory-event-related adverse drug reaction signals. PLoS ONE. 2018;13(11):e0207749. doi:10.1371/journal.pone.0207749 read 9 October 2026. doi.org/10.1371/journal.pone.0207749
  44. Tricco AC, Lillie E, Zarin W, O'Brien KK, Colquhoun H, Levac D, Moher D, Peters MDJ, Horsley T, Weeks L, Hempel S, Akl EA, Chang C, McGowan J, Stewart L, Hartling L, Aldcroft A, Wilson MG, Garritty C, Lewin S, Godfrey CM, Macdonald MT, Langlois EV, Soares-Weiser K, Moriarty J, Clifford T, Tuncalp O, Straus SE. PRISMA extension for scoping reviews (PRISMA-ScR): checklist and explanation. Annals of Internal Medicine. 2018;169(7):467-473. doi:10.7326/M18-0850 read 9 October 2026. doi.org/10.7326/M18-0850
  45. Tafti AP, Badger J, LaRose E, Shirzadi E, Mahnke A, Mayer J, Ye Z, Page D, Peissig P. Adverse Drug Event Discovery Using Biomedical Literature: A Big Data Neural Network Adventure. JMIR Medical Informatics. 2017;5(4):e51. doi:10.2196/medinform.9170 read 9 October 2026. doi.org/10.2196/medinform.9170
  46. Shah RR. Importance of Publishing Adverse Drug Reaction Case Reports: Promoting Public Health and Advancing Pharmacology and Therapeutics. Drug Safety - Case Reports. 2017;4(1):11. doi:10.1007/s40800-017-0053-0 read 9 October 2026. doi.org/10.1007/s40800-017-0053-0
  47. Noren GN. The Power of the Case Narrative - Can it be Brought to Bear on Duplicate Detection? Drug Safety. 2017;40(7):543-546. doi:10.1007/s40264-017-0548-8 read 9 October 2026. doi.org/10.1007/s40264-017-0548-8
  48. Rolfes L, van Hunsel F, van der Linden L, Taxis K, van Puijenbroek E. The Quality of Clinical Information in Adverse Drug Reaction Reports by Patients and Healthcare Professionals: A Retrospective Comparative Analysis. Drug Safety. 2017;40:607-614. doi:10.1007/s40264-017-0530-5 read 9 October 2026. doi.org/10.1007/s40264-017-0530-5
  49. Xu R, Wang Q. Large-scale combining signals from both biomedical literature and the FDA Adverse Event Reporting System (FAERS) to improve post-marketing drug safety signal detection. BMC Bioinformatics. 2014;15(1):17. doi:10.1186/1471-2105-15-17 read 9 October 2026. doi.org/10.1186/1471-2105-15-17
  50. Tregunno PM, Fink DB, Fernandez-Fernandez C, et al. Performance of Probabilistic Method to Detect Duplicate Individual Case Safety Reports. Drug Safety. 2014;37:249-258. doi:10.1007/s40264-014-0146-y read 9 October 2026. doi.org/10.1007/s40264-014-0146-y
  51. Pontes H, Clement M, Rollason V. Safety Signal Detection: The Relevance of Literature Review. Drug Safety. 2014;37:471-479. doi:10.1007/s40264-014-0180-9 read 9 October 2026. doi.org/10.1007/s40264-014-0180-9
  52. Grant MJ, Booth A. A typology of reviews: an analysis of 14 review types and associated methodologies. Health Information and Libraries Journal. 2009;26(2):91-108. doi:10.1111/j.1471-1842.2009.00848.x read 9 October 2026. doi.org/10.1111/j.1471-1842.2009.00848.x
  53. Kelly WN, Arellano FM, Barnes J, Bergman U, Edwards RI, Fernandez AM, Freedman SB, Goldsmith DI, Huang KA, Jones JK, McLeay R, Moore N, Stather RH, Trenque T, Troutman WG, van Puijenbroek E, Williams F, Wise RP. Guidelines for Submitting Adverse Event Reports for Publication. Drug Safety. 2007;30(5):367-373. doi:10.2165/00002018-200730050-00001 read 9 October 2026. doi.org/10.2165/00002018-200730050-00001
  54. Derry S, Loke YK, Aronson JK. Incomplete evidence: the inadequacy of databases in tracing published adverse drug reactions in clinical trials. BMC Medical Research Methodology. 2001;1(1):7. doi:10.1186/1471-2288-1-7 read 9 October 2026. doi.org/10.1186/1471-2288-1-7
  55. Validation of Artificial Intelligence to Support the Automatic Coding of Patient Adverse Drug Reaction Reports, Using Nationwide Pharmacovigilance Data. doi:10.1007/s40264-022-01153-8 read 9 October 2026. doi.org/10.1007/s40264-022-01153-8
  56. vigiGrade: A Tool to Identify Well-Documented Individual Case Reports and Highlight Systematic Data Quality Issues. doi:10.1007/s40264-013-0131-x read 9 October 2026. doi.org/10.1007/s40264-013-0131-x

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 9). Literature Surveillance in Pharmacovigilance: Trends, Challenges and Prospects. Pubrica. https://pubrica.com/academy/latest-topics/recent-trends-challenges-and-future-prospects-in-literature-surveillance-of-pharmacovigilance/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

Literature surveillance cycle record

One such record is completed per screening cycle. It is designed so that the compliance questions — what was searched, when, by whom, when the clock started, and whether the deadline was met — can each be answered from the record rather than from the SOP.

LITERATURE SURVEILLANCE CYCLE RECORD

CYCLE
  Product(s) / active substance(s): ______________________
  Coverage window: ______ to ______
  Cycle run date: ______    Performed by: ______________

SCOPE  (review and date this, do not inherit it)
  Substances in scope of the EMA MLM service (holder does NOT monitor
    or report into EudraVigilance for these): ______________________
  Substances the holder monitors directly: ______________________
  Non-EEA territories with separate requirements: ______________
  Local / non-indexed / congress sources screened separately: ______
  Scope last reviewed on: ______  By: ______________

SOURCES SEARCHED
  Database ................ Interface ......... Frequency .... Records
  1. ____________________  ________________    __________     ______
  2. ____________________  ________________    __________     ______
  3. ____________________  ________________    __________     ______
  4. ____________________  ________________    __________     ______
  Search string version in force this cycle: ______ (Annex ____)
  Date this string version took effect: ______
  Strings as executed attached?  [ ]

SCREENING
  Abstract level:  screened ______  excluded ______  by ______________
  Full-text level: assessed ______  excluded ______  by ______________
  [ ] A reason is recorded against EVERY full-text exclusion
  Full-text review triggered by criteria other than harm appearing in
    the abstract?  [ ] yes - criteria: ______________________

CASES AND THE CLOCK  (one row per valid case)
  Ref | Primary source (1st listed / corresponding author) | Day zero |
      | Date all four minimum criteria identifiable        | Submitted |
      | Days elapsed | Within 15 calendar days? |
  ____|___________________________________________|__________|_________
  ____|___________________________________________|__________|_________
  ____|___________________________________________|__________|_________
  Minimum criteria used for day zero:
    [ ] identifiable patient   [ ] identifiable reporter
    [ ] suspect product        [ ] suspected adverse reaction
  Median days to submission this cycle: ______

DE-DUPLICATION
  Method used: ______________________
  Evaluated against OUR database?  [ ] yes, on ______  [ ] no
  Performance observed locally: ______________________
  Suspected duplicates flagged: ______  Confirmed: ______

AUTOMATION  (if used)
  Tool / model: ______________________  Version: ______
  Used for: [ ] retrieval  [ ] ranking  [ ] entity extraction
            [ ] relation extraction  [ ] NOT used for seriousness
  Human decision point retained at: ______________________
  Local recall/precision last measured: ______  On date: ______

SIGN-OFF
  Nil-finding statement supported by the counts above?  [ ]
  Reviewed by: ______________  Date: ______

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s safety and regulatory writing teams develop and document literature-surveillance processes to inspection standard: search strings that withstand the adverse-effect indexing gap, version-controlled run records, day zero defined against the minimum ICSR criteria rather than the publication date, and a scope statement distinguishing what the EMA Medical Literature Monitoring service covers from what remains to be monitored. Pubrica also writes the surveillance sections of periodic safety reports so that a nil finding is supported by counts rather than asserted, and states explicitly where an automated tool’s published performance will not transfer to the data in question.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Pubrica reviews surveillance standard operating procedures and periodic report sections to identify questions an inspector would raise that the documents cannot currently answer.

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.