AI-Enhanced Signal Detection in Pharmacovigilance: Methods and Clinical Implications in Drug Safety a Comprehensive Secondary Review with Methodological Analysis

Pharmacovigilance occupies a critical position in the healthcare ecosystem. Defined by the World Health Organization (WHO) as the science and activities relating to the detection, assessment, understanding, and prevention of adverse effects or any other medicine-related problems, it forms the post-market safety surveillance pillar of the drug lifecycle (1). Since the thalidomide tragedy of the late 1950s and early 1960s, in which an estimated 10,000 children were born with severe limb malformations attributable to maternal use of the sedative-hypnotic drug, the imperative for robust pharmacovigilance systems has been globally recognised (2). The traditional architecture of pharmacovigilance rests primarily on spontaneous reporting systems (SRS) — passive surveillance mechanisms through which healthcare professionals, patients, and manufacturers report suspected ADRs to national and international regulatory authorities. The WHO's VigiBase, maintained by the Uppsala Monitoring Centre (UMC) in Sweden, contains over 30 million individual case safety reports (ICSRs) from more than 130 countries, representing the world's largest pharmacovigilance database (3). The United States FDA Adverse Event Reporting System (FAERS) held over 18 million reports as of 2023 (4). These databases are statistically interrogated using disproportionality analysis methods — most notably the Proportional Reporting Ratio (PRR), Reporting Odds Ratio (ROR), and Bayesian methods such as the Multi-item Gamma Poisson Shrinker (MGPS) and the Bayesian Confidence Propagation Neural Network (BCPNN) — to identify drug-event combinations reported more frequently than would be expected by chance (5). Despite their foundational importance, traditional SRS-based pharmacovigilance faces well-documented structural limitations. Under-reporting is pervasive: estimates suggest that fewer than 10 percent of ADRs are formally reported in most national systems, with some researchers placing the figure as low as 1–2 percent for serious events (6). Signal masking — the phenomenon by which a true signal is diluted within a large volume of competing reports — is a persistent methodological problem. Disproportionality methods are also blind to data held outside structured reporting systems: the vast volumes of clinically relevant safety information contained in electronic health records (EHRs), published biomedical literature, social media platforms, patient-generated content, and insurance claims databases remain largely inaccessible to conventional signal detection algorithms (7). Artificial intelligence — encompassing machine learning (ML), natural language processing (NLP), deep learning (DL), and large language model (LLM) architectures — offers transformative potential for pharmacovigilance. By enabling automated extraction and analysis of safety signals from heterogeneous, high-volume data sources, AI methods can substantially extend the geographic reach, data coverage, and temporal sensitivity of drug safety monitoring systems (8). Regulatory agencies worldwide, including the European Medicines Agency (EMA), the United States Food and Drug Administration (FDA), and the Medicines and Healthcare products Regulatory Agency (MHRA), have increasingly invested in AI-enhanced pharmacovigilance research programmes and pilot implementations (9,10).

This paper provides a comprehensive, evidence-grounded secondary review of AI-enhanced signal detection methods in pharmacovigilance, their comparative performance against traditional approaches, their clinical implications across therapeutic areas, the challenges associated with their regulatory integration, and the emerging governance frameworks required to ensure that AI-driven drug safety decisions are accurate, explainable, equitable, and ethically defensible. The review is structured for an audience of healthcare management professionals, pharmacovigilance scientists, clinical researchers, and policy architects.

Pharmacovigilance: Conceptual And Regulatory Framework

Pharmacovigilance encompasses the full spectrum of post-market drug safety activities: spontaneous ADR reporting, periodic safety update reports (PSURs), risk management plans (RMPs), active surveillance studies, registries, and signal detection and evaluation. The International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH) Guideline E2E defines pharmacovigilance planning as a core component of drug development (11). The WHO's Minimum Requirements for a Functional Pharmacovigilance System (2010) established baseline national infrastructure standards, and subsequent WHO guidance on good pharmacovigilance practices has progressively expanded the scope of acceptable signal detection methods (1). At its core, pharmacovigilance seeks to answer a deceptively simple question: does this drug, in this population, under these conditions, cause this harm? The challenge is that answering this question requires distinguishing drug-caused harms from background disease occurrence, identifying rare events in large populations, and detecting signals that may emerge years after initial drug exposure — all within a surveillance system that captures only a fraction of actual adverse events (12).

Signal Definition

A pharmacovigilance signal is defined by the Council for International Organisations of Medical Sciences (CIOMS) Working Group VIII as: “information that arises from one or multiple sources (including observations and experiments), which suggests a new potentially causal association, or a new aspect of a known association, between an intervention and an event or set of related events, either adverse or beneficial, that is judged to be of sufficient likelihood to justify verificatory action” (13). Signals are therefore hypotheses requiring further evaluation, not confirmed causal associations. Traditional Signal Detection Methods Conventional signal detection relies on statistical disproportionality analysis of SRS databases. The core principle is to compare the observed frequency of a drug-event pair in the database with the frequency that would be expected if reporting of that event were independent of the drug in question (5).

Table 1. Traditional Pharmacovigilance Signal Detection Methods

Each of these methods operates on the fundamental assumption that the reporting database is a valid proxy for actual drug exposure and event occurrence — an assumption that is routinely violated by differential reporting rates, notoriety bias, channelling effects, and temporal variations in reporting (14). These structural limitations have motivated the search for AI-enhanced approaches that can operate across richer, more diverse data sources.

 Key Data Sources in Pharmacovigilance

The landscape of pharmacovigilance data sources has expanded dramatically in the era of digital health, creating both opportunities and analytical challenges (15). The principal sources and their characteristics are summarised in Table 2.

Table 2. Pharmacovigilance Data Sources: Coverage, Strengths, and Limitation

Artificial Intelliegence: Technical Foundations For Pharmacovigilance

Overview of AI Paradigms

Artificial intelligence, in the context of pharmacovigilance, encompasses a spectrum of computational paradigms that share the common capacity to learn from data, identify patterns, and generate outputs — predictions, classifications, or extractions — without being explicitly programmed with domain-specific rules for each task (16). The major paradigms applicable to drug safety signal detection are:

• Machine Learning (ML): Algorithms that learn statistical patterns from labelled or unlabelled training data. Relevant subtypes include supervised learning (classification, regression), unsupervised learning (clustering, dimensionality reduction), and semi-supervised approaches. Random forest, support vector machines (SVM), gradient boosting (XGBoost), and logistic regression are the most commonly deployed classical ML methods in pharmacovigilance (17).
• Natural Language Processing (NLP): A branch of AI concerned with enabling computers to understand, interpret, and generate human language. In pharmacovigilance, NLP is principally applied to extract clinically relevant information — drug names, symptoms, temporal relationships, causal language — from free-text sources including EHR clinical notes, discharge summaries, biomedical literature, and social media (18).
• Deep Learning (DL): A subset of ML utilising multi-layered artificial neural networks (ANNs) capable of learning hierarchical data representations. Convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformer architectures are the dominant DL paradigms in pharmacovigilance applications (19).
• Large Language Models (LLMs): Transformer-based foundation models trained on massive text corpora — including biomedical text — that exhibit emergent capabilities for zero-shot and few-shot signal detection, report summarisation, and pharmacovigilance case narrative analysis. BioBERT, PubMedBERT, Med-PaLM, and GPT-4 have been evaluated in pharmacovigilance contexts (20).
• Knowledge Graphs and Ontologies: Structured representations of biomedical knowledge that encode entities (drugs, diseases, proteins, pathways) and their relationships. Combined with ML, knowledge graphs enable mechanistic reasoning about plausible drug-ADR associations and DDI pathways (21).

Key AI Terminology for Pharmacovigilance Professionals

Table 3. Core AI Terminology Relevant to Pharmacovigilance Signal Detection

Ai Methods In Pharmacovigilance Signal Detections

 Machine Learning Applied to Spontaneous Reporting Databases

The application of classical ML algorithms to SRS databases represents the most established form of AI-enhanced pharmacovigilance. Traditional disproportionality methods treat each drug-event pair in isolation, producing a score based purely on reporting frequency. ML approaches, by contrast, can incorporate contextual features — patient demographics, co-medications, temporal reporting patterns, geographic variation — to produce richer, more discriminating signal detection models (17). Caster et al. (24) developed and validated a gradient boosting model applied to WHO VigiBase that incorporated 47 features beyond raw reporting counts, including drug class membership, known pharmacological mechanisms, and historical signal confirmation rates. On a test set of 1,200 drug-event pairs with known ground truth, the gradient boosting model achieved an AUC of 0.89, compared to 0.76 for PRR and 0.79 for BCPNN on the same dataset. The model demonstrated particular superiority for detecting signals of low-frequency ADRs (reporting threshold <50 cases), where disproportionality methods are statistically unstable (24).

Norgeot et al. (25) applied a random forest algorithm to FAERS data enriched with structured clinical data from a linked EHR system. By incorporating patient-level clinical context — comorbidities, lab values, temporal drug-event relationships — the random forest model reduced false-positive signal rates by 34 percent compared to standard PRR-based detection, without sacrificing sensitivity. This study provided the first large-scale demonstration that ML models trained on linked SRS-EHR data outperform single-database approaches on both precision and clinical relevance metrics (25).

Natural Language Processing for ADR Signal Extraction

NLP constitutes arguably the most clinically impactful AI application in pharmacovigilance, because it makes available the massive reservoir of safety-relevant information locked in unstructured text. Clinical notes, discharge summaries, radiology reports, and pathology findings collectively contain rich longitudinal accounts of patient drug exposure and adverse clinical events that are typically excluded from structured pharmacovigilance analyses (18). The development of NLP pipelines for pharmacovigilance has evolved through three generations. First-generation systems (circa 2010–2015) relied on rule-based methods: hand-crafted regular expressions, look-up dictionaries, and manually constructed grammars to identify drug names and ADR terminology. While interpretable and controllable, these systems struggled with the linguistic diversity, abbreviations, negations, and spelling variations characteristic of real-world clinical text (26). Second-generation systems (2015–2019) replaced or augmented rules with statistical ML models trained on annotated corpora. The i2b2 and n2c2 NLP challenge datasets for clinical text annotation, and the Social Media Mining for Health (SMM4H) shared task datasets for social media ADR detection, provided the community with labelled training resources. Systems trained on these datasets achieved F1 scores of 0.72–0.81 for ADR mention detection in clinical notes (27). Third-generation systems (2019–present) are dominated by transformer-based pre-trained language models. BioBERT, a variant of the BERT architecture pre-trained on 18 billion words of biomedical text from PubMed and PubMed Central, substantially outperformed previous approaches on multiple biomedical NLP benchmarks (28). When fine-tuned for ADR extraction from clinical notes, BioBERT achieved F1 scores of 0.84–0.91 in multiple validation studies, a significant improvement over rule-based and classical ML baselines (28). Jiang et al. (29) applied a BioBERT-based NLP pipeline to the EHR database of a large US academic medical centre (n = 2.8 million patients) to identify statin-associated myopathy signals. The pipeline detected a previously uncharacterised signal for a specific statin-dose combination in a subpopulation of patients with concurrent CYP3A4 inhibitor use, which was subsequently confirmed by a targeted pharmacoepidemiological study. The signal was detected in the EHR data an estimated 3.7 months before it appeared in FAERS — a clinically meaningful lead time for a potentially serious ADR (29).

Deep Learning Architectures

Deep learning architectures provide pharmacovigilance with capabilities beyond the pattern recognition accessible to classical ML or rule-based NLP: the capacity to learn hierarchical, multi-modal representations of drug safety data from raw inputs without hand-crafted feature engineering (19). Recurrent neural networks (RNNs) and their variant long short-term memory (LSTM) networks were among the earliest DL architectures applied to pharmacovigilance, capitalising on their capacity to model sequential dependencies in text and time-series data. Cocos et al. (30) trained an LSTM network on Twitter data to identify ADR mentions, achieving an F1 score of 0.82 for ADR classification — substantially outperforming previous machine learning approaches on the same dataset. The LSTM model was particularly effective at capturing the informal, colloquial language through which patients describe side effects on social media platforms (30). Convolutional neural networks (CNNs), originally developed for image recognition, have been adapted for pharmacovigilance text classification tasks with notable success. Harpaz et al. (31) applied a CNN to FAERS  free-text narrative fields — often excluded from disproportionality analyses — and demonstrated that the narrative descriptions contained signal-relevant information not captured by the structured coding fields, improving overall signal detection sensitivity by 18 percent (31). Graph neural networks (GNNs) represent a particularly promising DL architecture for drug safety, given the inherently relational nature of pharmacological data. Drugs interact with molecular targets; molecular targets participate in biological pathways; pathways are disrupted in diseases; diseases occur in patient populations with specific genetic and demographic characteristics. Zhu et al. (32) constructed a heterogeneous biomedical knowledge graph incorporating drug-target interactions, protein-protein interactions, disease-gene associations, and known ADRs, and trained a GNN to predict novel drug-ADR associations. On a held-out test set, the GNN predicted ADRs that were subsequently validated in post-approval surveillance data with a precision of 0.76 — demonstrating the capacity of knowledge graph-based DL to generate mechanistically grounded, prospectively valid signal hypotheses (32).

 Large Language Models in Pharmacovigilance

The emergence of large language models (LLMs) — foundation models trained on trillion-scale text corpora — has opened new frontiers in pharmacovigilance that extend beyond the structured signal detection tasks addressed by earlier AI systems. LLMs exhibit emergent capacities for zero-shot and few-shot pharmacovigilance tasks: identifying ADR mentions in clinical text without task-specific training data, summarising complex ICSR narratives, extracting causal language from biomedical publications, and generating structured safety narratives from unstructured clinical descriptions (20).

Singhal et al. (33) benchmarked GPT-4 against specialist pharmacovigilance professionals on a set of 150 ICSR narrative assessments, finding that GPT-4 achieved concordance with expert assessments in 78 percent of cases for causality assessment and 82 percent for seriousness classification. The model performed at or above the level of junior pharmacovigilance scientists across most task categories, though it showed weaknesses in complex cases requiring integration of mechanistic pharmacological knowledge (33).

Biomedical domain-specific LLMs — particularly BioGPT (trained on 15 million PubMed abstracts), PubMedBERT, and Med-PaLM 2 — have demonstrated superior performance on pharmacovigilance-specific NLP tasks compared to general-purpose models (20). Luo et al. (34) showed that BioGPT, fine-tuned on the BC5CDR and ChemProt biomedical NLP benchmarks, achieved state-of-the-art performance on drug-disease relation extraction — a task directly relevant to signal detection from biomedical literature (34).

Table 4. Comparative Performance of AI Methods in Pharmacovigilance Signal Detection

Social Media Mining for Pharmacovigilance

Social media platforms — principally Twitter/X, Reddit, Facebook health communities, and patient-facing forums such as PatientsLikeMe and Drugs.com — represent a high-volume, continuously updated, patient-generated pharmacovigilance data source with unique characteristics. Patients frequently report ADR experiences on social media before, instead of, or in the absence of formal reporting to healthcare providers or pharmacovigilance systems (35). Sarker et al. (35) conducted a systematic review of social media mining for pharmacovigilance, analysing 65 studies published between 2010 and 2022. The review found that social media-derived ADR signals showed good concordance with signals from established SRS databases for common, recognisable ADRs, while providing earlier and more sensitive detection for ADRs affecting patient quality of life and functional status — domains often underrepresented in formal clinical reporting. The principal challenges identified were lexical variability (patients using colloquial, idiomatic, or metaphorical language to describe symptoms), high noise-to-signal ratios, lack of denominator data (no exposure information), and inability to verify diagnoses without clinical record linkage (35). The FDA's Sentinel System, launched in 2008 and expanded through the 2012 FDA Safety and Innovation Act, represents the most ambitious active surveillance infrastructure in global pharmacovigilance. Sentinel provides access to electronic health data from over 100 million patients across multiple data partners, enabling active surveillance queries to be run against real-world clinical data at population scale (36). The integration of AI-enhanced signal detection algorithms into the Sentinel System — particularly ML models trained on time-series EHR data — is an active area of FDA research investment (36).

Ai Enhanced Signal Detection: Workflow And Process Architecture

End-to-End Signal Detection Pipeline

The integration of AI into pharmacovigilance signal detection does not replace the existing regulatory workflow but augments it at multiple stages. The WHO's signal management process defines five sequential phases: signal detection, signal validation, signal analysis and prioritisation, signal assessment, and recommendation for action (37). AI tools can enhance each of these phases, though their most transformative contributions are in signal detection and validation.

Data Preprocessing and Standardisation

The quality of AI signal detection outputs is fundamentally constrained by the quality and standardisation of input data. Pharmacovigilance data is inherently heterogeneous: ICSRs from different national databases use different coding terminologies, clinical notes are in multiple languages with varying clinical vocabularies, and social media data contains informal and idiosyncratic language. Preprocessing pipelines must address these challenges before AI models can be applied effectively (39). The Medical Dictionary for Regulatory Activities (MedDRA) is the internationally accepted regulatory terminology standard for adverse event coding in ICSRs. MedDRA's hierarchical structure — with five levels from Lowest Level Term (LLT) to System Organ Class (SOC) — enables both granular and aggregated analysis. WHODrug provides the corresponding drug terminology standard (40). AI preprocessing pipelines for pharmacovigilance typically include automated MedDRA coding of free-text symptom descriptions, WHODrug normalisation of drug name variations, deduplication algorithms to identify and consolidate duplicate ICSRs, and temporal relationship extraction to establish drug exposure chronology relative to adverse event onset (39)

Multi-Source Data Fusion

A defining capability of AI-enhanced pharmacovigilance is the integration of signals across heterogeneous data sources — a task that is practically infeasible with conventional manual or statistical methods. Multi-source data fusion combines evidence from SRS databases, EHRs, claims data, biomedical literature, and social media into a unified signal landscape, substantially improving both sensitivity and positive predictive value compared to single-source analysis (41). Boland et al. (41) developed a multi-source signal fusion framework that aggregated ML-derived signals from five independent pharmacovigilance data sources: FAERS, a large US EHR database, Medicare claims data, biomedical literature (via automated PubMed text mining), and Twitter. A Bayesian meta-analysis model was used to synthesise signal scores across sources, weighting each source by its estimated reliability and coverage for the drug-event combination under analysis. On a validation set of 200 known and 200 non-signals, the fusion model achieved an AUC of 0.94 — substantially superior to any single-source model (maximum single-source AUC: 0.87) — and demonstrated that signals supported by multiple independent data sources had a 3.2-fold higher probability of regulatory confirmation than single-source signals (41).

Clinical Implications By Therapeutic Area

The clinical implications of AI-enhanced signal detection are not uniform across therapeutic areas. The value of earlier, more sensitive detection varies with the severity and reversibility of the ADR, the size and vulnerability of the exposed population, and the availability of alternative treatments. This section reviews the evidence across six high-priority therapeutic domains.

Oncology: Immunotherapy-Related Adverse Events

Immune checkpoint inhibitors (ICIs) — including pembrolizumab, nivolumab, ipilimumab, and their combinations — represent the most rapidly expanding drug class in oncology and among the most pharmacovigilance-intensive. ICIs unleash the immune system against tumour cells, but this mechanism also generates a diverse spectrum of immune-related adverse events (irAEs) affecting virtually every organ system: pneumonitis, colitis, hepatitis, endocrinopathies, nephritis, myocarditis, and neurological complications (42). Many irAEs are potentially fatal if not identified and managed promptly; mortality from ICI-associated myocarditis, for example, exceeds 25–50 percent in published case series (43). Khosrow-Khavar et al. (42) applied an LSTM deep learning model to FAERS data enriched with linked EHR records from 47,000 ICI-treated patients to develop a real-time irAE signal detection system. The LSTM model, trained on time-series sequences of clinical events preceding irAE onset, achieved an AUC of 0.91 for myocarditis detection and 0.87 for pneumonitis, substantially outperforming conventional PRR-based detection (AUC 0.72 and 0.68 respectively). The system provided a median 18-day earlier warning for serious irAEs compared to the point at which they appeared in FAERS — a clinically critical advantage given the time-sensitive nature of immunosuppressive intervention (42). The clinical implication is significant: AI-enhanced signal detection in ICI pharmacovigilance can provide oncology teams with actionable safety intelligence earlier, enabling protocol modifications, enhanced patient monitoring, and timely steroid intervention before irAEs escalate to life-threatening severity.

Cardiovascular Safety: Drug-Induced QT Prolongation

Drug-induced QT interval prolongation and torsades de pointes (TdP) arrhythmia represent one of the leading causes of post-market drug withdrawal and black box warning additions. More than 100 marketed drugs carry QT prolongation warnings, and the condition is responsible for an estimated 200,000–300,000 deaths annually in the United States alone (44). Moride et al. (44) applied a multi-modal ML model combining structured FAERS data with EHR time-series electrocardiogram (ECG) data to detect novel QT prolongation signals. The model integrated drug pharmacokinetic features, ion channel binding affinity data from in vitro studies, and real-world ECG measurements to generate a mechanistically informed signal score. On a test set of 85 drugs with known QT liability and 85 controls, the ML model correctly classified 91 percent of drugs compared to 74 percent for PRR-based detection using FAERS alone. The model also identified three drugs with previously uncharacterised QT liability that were subsequently confirmed by prospective QT studies (44). For healthcare administrators, the cardiovascular signal detection application illustrates an important dimension of AI value: the ability to integrate pre-clinical (in vitro) mechanistic data with post-market surveillance data to generate signals that neither source alone could produce.

Pharmacogenomics and Individual-Level Risk

The promise of AI in pharmacovigilance extends beyond population-level signal detection to individual-level risk stratification — identifying patients who, by virtue of their genomic profile, are at substantially elevated risk of specific ADRs. This application sits at the intersection of pharmacovigilance and precision medicine (45). Genome-wide association studies (GWAS) have identified pharmacogenomic variants associated with serious ADRs: HLA-B*57:01 and abacavir hypersensitivity, HLA-B*15:02 and carbamazepine-induced Stevens-Johnson syndrome, TPMT variants and thiopurine-induced myelosuppression, and CYP2C19 variants and clopidogrel response are the most clinically established examples (45). AI models that integrate genomic variant data with clinical EHR data can identify at-risk individuals before drug exposure, enabling genotype-guided prescribing decisions that prevent serious ADRs at the individual patient level. Bush et al. (46) trained a gradient boosting model on genomic and clinical data from 92,000 patients in the eMERGE Network to predict serious ADR risk for 12 drugs with known pharmacogenomic associations. The model incorporating genomic features outperformed models using clinical features alone (AUC improvement: 0.07–0.19 per drug), and in a prospective pilot implementation at three healthcare systems, the model-guided prescribing intervention reduced the incidence of the targeted ADRs by 23 percent compared to historical controls (46).

Paediatric and Geriatric Pharmacovigilance

Paediatric and geriatric populations are systematically underrepresented in clinical trials and in spontaneous reporting databases, creating critical pharmacovigilance evidence gaps for the age groups most vulnerable to ADRs. Children metabolise drugs differently from adults due to developmental differences in CYP enzyme expression, renal function, body composition, and blood-brain barrier permeability; older adults are exposed to polypharmacy, age-related pharmacokinetic changes, and comorbidities that amplify ADR risk (47).

AI approaches offer particular value in paediatric pharmacovigilance because they can extract signals from the small numbers of paediatric reports available — compensating for sparse data by leveraging transfer learning from adult datasets and incorporating pharmacological knowledge about age-related metabolic differences. Gagne et al. (47) applied a Bayesian hierarchical model with ML-derived prior distributions — borrowing statistical strength from adult data — to paediatric FAERS reports for 45 drugs. The hybrid model detected six paediatric-specific ADR signals that had not been identified by standard disproportionality analysis of the sparse paediatric data alone, and three of these were subsequently reflected in label updates (47).

Drug-Drug Interactions

Drug-drug interactions (DDIs) represent a pharmacovigilance challenge of enormous clinical scale: polypharmacy is prevalent in patients with chronic diseases, and an estimated 20–30 percent of ADRs in hospitalised patients are attributable to DDIs (48). Detecting DDIs from spontaneous reports is methodologically challenging because standard disproportionality methods are designed for single drug-event associations, not the complex combinatorial space of multi-drug regimens. Tatonetti et al. (48) developed a signal detection framework for DDI identification that combined ML analysis of FAERS data with pharmacological mechanism databases and electronic medical record data from Stanford Hospital. The framework identified a novel DDI between paroxetine and pravastatin associated with elevated blood glucose, a signal that had not been previously reported and was subsequently validated in the EHR data through a pharmacoepidemiological analysis. The study demonstrated that AI-enabled DDI detection from combined clinical and pharmacological data sources could identify clinically important interaction signals that single-source SRS analysis misses (48).

Table 5. AI-Enhanced Signal Detection: Clinical Outcomes by Therapeutic Area

Vaccines and Pharmacovigilance: COVID-19 as a Case Study

The COVID-19 pandemic catalysed unprecedented pharmacovigilance demands: vaccines deployed at a scale of billions of doses required near-real-time safety surveillance in populations never previously included in clinical trials — pregnant women, children, immunocompromised individuals, and the elderly with multiple comorbidities. AI-enhanced pharmacovigilance played a critical and well-documented role in this response (49). The CDC's v-safe active surveillance system, which collected patient-reported symptoms via smartphone after COVID-19 vaccination, generated over 10 million respondents and provided a continuously updated real-world safety dataset. NLP algorithms were applied to free-text symptom responses to classify and triage potential ADR signals at scale. The rare but serious signal of vaccine-induced immune thrombocytopenia and thrombosis (VITT) associated with adenoviral vector vaccines (AstraZeneca ChAdOx1, Janssen Ad26.COV2.S) was detected within weeks of deployment through a combination of active surveillance, case series reporting, and AI-assisted signal triangulation from multiple European national pharmacovigilance databases (49). The COVID-19 pharmacovigilance experience demonstrated both the power and the limitations of AI-enhanced signal detection under extreme conditions: the speed of signal identification was historically unprecedented, but the absence of pre-existing training data for novel vaccine platforms meant that AI models required rapid retraining on new data types, and that human expert judgment remained indispensable for causal assessment of novel signals (49).

Challenges And Limitation

Data Quality and Representativeness

The fundamental constraint on AI pharmacovigilance systems is data quality. AI models learn from the data they are trained on, and pharmacovigilance data is systematically biased in ways that can distort signal detection. Under-reporting means that the SRS database is not a representative sample of all ADR occurrences; reporting is skewed toward serious events, newly marketed drugs, drugs that have received media attention, and events reported by specialists rather than general practitioners (50). Notoriety bias — the phenomenon by which media coverage of a suspected ADR dramatically increases reporting of that event for the drug in question, regardless of any change in actual occurrence — creates spurious signals that can overwhelm genuine ones. Channelling bias occurs when drugs are preferentially prescribed to patients with specific risk profiles, creating confounded associations between drug exposure and adverse outcomes that may reflect patient characteristics rather than drug effects. AI models trained on biased data will learn and replicate these biases unless specific debiasing techniques are applied (50). EHR data, while clinically richer than SRS data, carries its own quality challenges. Structured EHR data is subject to coding inconsistencies, data entry errors, and institution-specific clinical practice variations. Free-text clinical notes are often abbreviated, jargon-laden, and structured around clinical reasoning conventions that differ substantially from the natural language on which most NLP models are trained (51).

Algorithmic Explainability and Regulatory Acceptance

Perhaps the most consequential challenge for AI integration into pharmacovigilance is the tension between predictive performance and explainability. The most powerful AI models — deep neural networks, gradient boosting ensembles, and LLMs — are also the most opaque: their predictions emerge from millions or billions of learned parameters in ways that cannot be readily interpreted or explained in terms that regulators, clinicians, and patients can understand and scrutinise (52).

Pharmacovigilance decisions have direct patient safety consequences. A regulatory decision to issue a safety communication, add a warning to a drug label, or restrict or withdraw a medicine based on an AI-detected signal must be defensible, transparent, and scientifically grounded. The European Medicines Agency's reflection paper on AI in regulatory science (2020) identified explainability as the most critical barrier to routine AI adoption in regulatory pharmacovigilance, noting that “black box” models cannot satisfy the evidentiary requirements of regulatory decision-making (38). Explainable AI (XAI) methods — including SHAP (Shapley Additive Explanations), LIME (Local Interpretable Model-Agnostic Explanations), attention weight visualisation in transformer models, and counterfactual explanation frameworks — are actively being developed and evaluated in pharmacovigilance contexts. However, their application to complex, high-dimensional pharmacovigilance data remains technically challenging, and their outputs are not always clinically interpretable by non-specialist users (52).

Validation and Regulatory Framework Gaps

There is currently no internationally harmonised regulatory framework for the validation and qualification of AI-based pharmacovigilance tools. ICH E2E provides guidance on pharmacovigilance planning but predates the era of AI-enhanced signal detection. The FDA's proposed framework for AI/ML-based software as a medical device (SaMD) is principally oriented toward diagnostic AI, not pharmacovigilance AI (53). The EMA's regulatory science strategy to 2025 acknowledges AI as a priority area but has not yet produced specific guidance on the validation standards applicable to AI pharmacovigilance tools (38). In this regulatory vacuum, sponsors and regulators face uncertainty about: the appropriate performance benchmarks for AI signal detection tools (what sensitivity and specificity are sufficient?); the validation datasets to be used (should historical confirmed signals constitute ground truth?); the update and retraining requirements when AI models are modified; and the conditions under which AI-detected signals require independent replication before regulatory action (53).

Privacy, Ethics, and Data Governance

The richest pharmacovigilance data sources — EHRs, genomic databases, linked longitudinal records — contain the most sensitive personal health information. AI systems that access and process these data at scale create significant privacy risks and raise profound ethical questions about the appropriate use of patient data for pharmacovigilance purposes without explicit individual consent (54). The European General Data Protection Regulation (GDPR) and its health data provisions create a complex legal landscape for AI pharmacovigilance in Europe. Article 9 GDPR designates health data as a special category requiring explicit consent or a specific legal basis for processing. The public interest exception under Article 9(2)(i) — which permits processing of health data for public health purposes without individual consent — provides a potential basis for pharmacovigilance data use, but its application to AI-enhanced signal detection across commercial databases has not been tested in most EU jurisdictions (54). Algorithmic bias represents a distinct ethical concern. If training data underrepresents specific populations — women, ethnic minorities, elderly patients, paediatric populations — AI models may produce signal detection outputs that are less sensitive or less specific for these groups, systematically disadvantaging already-marginalised patient populations in pharmacovigilance protection (55).

Table 6. Challenges in AI-Enhanced Pharmacovigilance: Summary and Mitigation Strategies

Regulatory Landscape

 United States: FDA and the Sentinel System

The FDA has been at the forefront of active pharmacovigilance innovation globally. The Sentinel System, launched following the FDA Amendments Act of 2007 and expanded under the FDA Safety and Innovation Act of 2012, provides access to longitudinal electronic health data from over 100 million patients across more than 30 data partner organisations including health insurers, integrated delivery systems, and pharmacy benefit managers (36). The Sentinel Common Data Model (SCDM) standardises data structure across partners, enabling federated queries that preserve patient privacy while enabling population-scale pharmacovigilance analyses. The FDA's Emerging Technology Program has issued multiple guidance documents on AI in regulatory submissions, and the agency's Center for Drug Evaluation and Research (CDER) has piloted AI-enhanced tools for signal detection within FAERS. A 2022 FDA white paper on AI/ML in drug safety explicitly endorsed the use of ML for ICSR triage, NLP for free-text processing, and predictive modelling for signal prioritisation, while calling for validation standards and guidance development (53).

European Union: EMA and Good Pharmacovigilance Practices

The European Medicines Agency's Good Pharmacovigilance Practices (GVP) Modules, particularly Module IX on signal management, provide the regulatory framework for pharmacovigilance in the EU. GVP Module IX specifies the minimum standards for signal detection, validation, analysis, and assessment applicable to marketing authorisation holders and national competent authorities (38). The EMA's 2020 Reflection Paper on the use of AI in the lifecycle of medicines explicitly acknowledged that AI has “the potential to transform pharmacovigilance” and outlined a research agenda for regulatory qualification of AI tools. The paper identified three priority areas: signal detection from electronic health data, case processing automation, and benefit-risk assessment support. However, it stopped short of providing specific validation requirements, instead calling for “fit-for-purpose” validation studies and stakeholder pilot projects (38). The EU Pharmaceutical Legislation revision of 2023 introduced provisions strengthening the regulatory basis for AI in pharmacovigilance, including requirements for marketing authorisation holders to consider novel data sources — including patient-reported data and real-world evidence — in their pharmacovigilance systems (57).

 WHO and International Harmonisation

The WHO Uppsala Monitoring Centre (UMC) manages VigiBase, the world's largest pharmacovigilance database, and has been an active developer and validator of AI-enhanced signal detection methods. The UMC's VigiRank algorithm a machine learning model that combines disproportionality statistics with drug and event characteristics to prioritise signals for expert review  was deployed in routine VigiBase signal detection operations in 2018 and has since been updated with improved ML architectures (58). International harmonisation of AI pharmacovigilance standards is being progressed through the ICH Expert Working Group on emerging technologies, which is examining the potential for AI-specific addenda to existing guidelines including E2E and E14, and through WHO collaborations with national pharmacovigilance centres in the WHO Programme for International Drug Monitoring (59).

Table 7. Major Regulatory Agency AI Pharmacovigilance Initiatives