Pharmacovigilance (PV) is undergoing a fundamental transformation driven by digital health technologies. Traditional pharmacovigilance, reliant on voluntary spontaneous reporting systems, has long been hindered by underreporting, delayed signal detection, and data fragmentation. The digital revolution—encompassing artificial intelligence (AI), machine learning, big data analytics, electronic health records (EHRs), mobile applications, and social media monitoring—is reshaping how adverse drug reactions (ADRs) are detected, assessed, and prevented.
This article provides a comprehensive medical overview of digital pharmacovigilance, examining its key components, benefits, and challenges. Drawing on evidence from the World Health Organization (WHO), the U.S. Food and Drug Administration (FDA), the European Medicines Agency (EMA), and recent peer-reviewed literature, we explore how digital tools are transforming drug safety surveillance from a reactive, manual process into a proactive, predictive, and patient-centred discipline. Real-world examples, including the WHO’s VigiBase with machine learning integration, the FDA’s Sentinel Initiative, EMA’s EudraVigilance with natural language processing, and mobile reporting applications such as VigiMobile, illustrate the practical impact of these technologies. Challenges including data quality, algorithmic bias, ethical concerns, and regulatory gaps are critically examined, alongside strategies for responsible implementation.
1. Introduction: The Evolution from Reactive to Proactive Drug Safety
Pharmacovigilance, defined by the World Health Organization as the science and activities relating to the detection, assessment, understanding, and prevention of adverse effects or any other drug-related problem, has traditionally been a reactive discipline. For decades, drug safety monitoring relied predominantly on spontaneous reporting systems (SRS), where healthcare professionals and patients voluntarily submitted individual case safety reports (ICSRs) of suspected ADRs to regulatory authorities.
While this system has successfully identified numerous safety signals—including the withdrawal of rofecoxib (Vioxx) and the identification of cardiovascular risks with certain diabetes medications—it has long been recognised as suffering from significant limitations: severe underreporting (estimated at over 90% of ADRs), delayed detection (often taking months or years), incomplete data, and fragmentation across different reporting systems.
A decade ago, pharmacovigilance was largely reactive. Clinicians reported ADRs manually, regulators reviewed cases in batches, and safety updates often lagged behind real-world risks. By 2025, the model had shifted dramatically. Artificial intelligence (AI), big data analytics, and globally networked health information systems are now central to pharmacovigilance, making it faster, broader, and in many cases, predictive.
Where regulators once relied on slow manual reviews, AI now sifts through millions of patient records, clinical notes, and spontaneous reports in near real-time, highlighting emerging safety concerns within days. This acceleration is saving lives, but it also creates new challenges. Poor-quality or biased data, ethical dilemmas surrounding accountability, and gaps in regulatory frameworks underscore the critical role of humans in pharmacovigilance.
This transformation has been catalysed by several converging factors: the exponential growth of real-world data (RWD) from electronic health records, claims databases, and patient registries; advances in AI and machine learning capable of processing unstructured data; the proliferation of mobile health applications enabling direct patient reporting; and the recognition that traditional PV systems are insufficient to ensure patient safety in an era of increasingly complex therapeutics. Global regulatory bodies including the WHO, EMA, and FDA are implementing AI-assisted PV systems, signalling a new paradigm in drug safety monitoring.
2. Key Components of Digital Pharmacovigilance
2.1 Spontaneous Reporting Systems Enhanced with AI and Big Data
2.1.1 VigiBase – The WHO Global Database
VigiBase, the WHO global database of individual case safety reports, is the world’s largest repository of ADR reports. As of 2025, VigiBase contains over 40 million ICSRs from more than 180 member countries. The database receives reports from national pharmacovigilance databases of WHO Programme for International Drug Monitoring (PIDM) members, including the US FDA Adverse Event Reporting System (FAERS) and the US Vaccine Adverse Event Reporting System (VAERS).
In recent years, VigiBase has incorporated machine learning tools that detect unusual reporting patterns more efficiently than traditional statistical methods. By integrating AI-driven signal detection with expert review, WHO has cut down the average time from initial report to global safety signal by weeks. The Uppsala Monitoring Centre (UMC), which manages VigiBase, has developed several AI-powered algorithms:
- vigiRank: A data-driven predictive model that uses a combination of disproportionality measures, report quality, and content to prioritise emerging safety signals. It has been the basis for UMC’s own signal detection in VigiBase since 2015.
- vigiMatch: A machine learning model that predicts duplicate case reports, representing the first statistical duplicate detection approach in pharmacovigilance deployed at scale. Duplicate reports are a significant problem for signal detection, but with the help of machine learning and new ways of comparing reports, they can be more effectively detected.
- vigiVec: A method for obtaining data-driven vector representations of medicines and adverse events based on reporting patterns in VigiBase, enabling semantic similarity analysis.
In March 2025, WHO and UMC introduced a new Data Access Conditions framework to clarify who can access VigiBase and under what circumstances, enhancing data sharing clarity and stakeholder confidence in global pharmacovigilance.
2.1.2 EudraVigilance – The European System
EudraVigilance is the central European system for collecting and analysing reports of suspected adverse reactions. As of 2025, the database contains over 34 million ICSRs for suspected AEs across Europe. The EMA has expanded EudraVigilance with natural language processing (NLP) to analyse narrative sections of ADR reports, improving the detection of complex safety issues, such as interactions involving multiple drugs.
The EMA and the Heads of Medicines Agencies (HMA) have published a joint workplan “Data and AI in medicines regulation to 2028,” setting out how the European medicines regulatory network plans to leverage large volumes of regulatory and health data as well as new tools. Key initiatives include developing guidance on AI in pharmacovigilance, fostering EU-wide and international collaboration, and providing the network with training on AI.
2.1.3 The FDA Sentinel Initiative
The U.S. Food and Drug Administration’s Sentinel Initiative, established in response to the FDA Amendments Act of 2007, is a national programme for active surveillance of medical products. The Sentinel System uses AI to analyse insurance claims, EHRs, and registry data. It contains linked electronic health records with claims data for over 25 million patients.
Sentinel has been able to detect cardiovascular risks linked to certain diabetes drugs much earlier than spontaneous reporting alone could. The Sentinel Real-World Evidence Data Enterprise (RWE-DE) is now integrating electronic health records with generative AI and machine learning to identify medication safety signals. Specific examples illustrate the application of generative AI/ML in identifying medication safety signals. TreeScan, a method developed by the Sentinel Program, is used for signal detection using claims and EHR databases.
2.2 Real-World Data (RWD) and Real-World Evidence (RWE)
Real-world data—data relating to patient health status and/or the delivery of healthcare routinely collected from a variety of sources—has become a cornerstone of digital pharmacovigilance. RWD sources include:
| Data Source | Description | PV Application |
|---|---|---|
| Electronic Health Records (EHRs) | Longitudinal patient records from healthcare institutions | Passive surveillance; identification of ADRs from clinical notes; pharmacoepidemiological studies |
| Administrative Claims Databases | Insurance and billing data | Large-scale population studies; drug utilisation studies; signal detection |
| Patient Registries | Disease-specific or product-specific registries | Long-term safety monitoring; post-authorisation safety studies |
| Laboratory and Diagnostic Data | Clinical laboratory results | Identification of biochemical ADRs; monitoring of drug-induced organ toxicity |
| Patient-Reported Outcomes (PROs) | Direct patient input via apps or portals | Patient-centred safety monitoring; real-time symptom tracking |
RWD ecosystems—including EHRs, administrative claims, and global spontaneous reporting systems—can be integrated with supervised machine learning to strengthen signal detection, risk prediction, and translational safety assessment. The HMA–EMA RWD catalogues represent a centralised digital approach for broad access to reliable data to ensure patient safety.
The EMA’s Data Analysis and Real World Interrogation Network (DARWIN EU®) is a federated network that provides a source of high-quality, validated real-world data on the uses, safety, and efficacy of medicines. DARWIN EU supports regulatory decision-making by delivering evidence that helps fill knowledge gaps and understand the use, safety, and benefits of medicines. A successful example of DARWIN EU outputs is the safety study requested during the review of doxycycline use and its association with suicidality ideation, which concluded that there was insufficient evidence to establish a causal relationship.
2.3 Mobile Applications and Digital Reporting Tools
Mobile health applications have emerged as powerful tools for patient engagement in pharmacovigilance, addressing the critical challenge of underreporting.
2.3.1 VigiMobile
VigiMobile is a customisable web application developed by the WHO and UMC for structured adverse event reporting by healthcare workers and the public. It is a free add-on to VigiFlow that works on any device and supports offline reporting, with forms saved and submitted once an internet connection is available. Reports can be filled out on a smartphone, tablet, or desktop – even without an internet connection. Once reconnected, reports are submitted to VigiBase, the WHO’s global database.
VigiMobile has been rolled out in multiple countries. In Mauritius, the Ministry of Health and Wellness, with WHO support, rolled out the National VigiMobile Programme across all hospitals and pharmacies, enabling healthcare workers and pharmacists to report ADRs and Adverse Events Following Immunization (AEFIs) in real time. The VigiMobile app provides real-time alerts on potential side effects of drugs and vaccines, allowing timely information and decision-making by healthcare professionals. Fiji became the first country in the Pacific to launch the VigiMobile app, enabling healthcare workers and the public to report unexpected side effects or reactions. The app uses WHO’s standard AEFI reporting form with the 25 core variables recommended by WHO for collecting AEFI data.
2.3.2 Other Digital Reporting Initiatives
The UMC has also launched an offline app in partnership with the WHO pharmacovigilance team for reporting possible side effects from vaccines. The Med Safety App is an international mobile tool for drug safety reporting. These tools enable direct patient reporting, which is particularly valuable for capturing patient-reported outcomes and experiences that may not be captured through traditional healthcare professional reporting.
2.4 Artificial Intelligence and Machine Learning in Signal Detection
AI is increasingly applied in pharmacovigilance to identify, prioritise, and interpret ADRs across real-world data sources. A comprehensive narrative review synthesising studies from 2015 to 2024 mapped AI use across four domains: extraction of ADRs from social and clinical text, supervised and ensemble signal detection in spontaneous reporting systems and EHRs, knowledge-graph-based discovery of drug-event associations, and prediction of outcome seriousness to support triage.
2.4.1 Natural Language Processing (NLP) for Unstructured Data
Large language models (LLMs) have shown particular promise in extracting ADRs from unstructured clinical notes. One study introduced an LLM-based model capable of extracting ADEs from free-text EHR notes with 94% accuracy and 84.1% recall. Another study evaluated the use of LLMs for pharmacovigilance of anti-Aβ therapies, providing preliminary indications that LLMs may serve as a pharmacovigilance signal detection tool.
2.4.2 Advanced Signal Detection Algorithms
AI-enhanced signal detection methods include:
- Disproportionality analysis: Using Reporting Odds Ratio (ROR) and Bayesian Confidence Propagation Neural Network (BCPNN) methods in VigiBase
- TreeScan: A method developed by the FDA Sentinel Program for signal detection using claims and EHR databases
- Omega interaction measure: For detecting drug-drug interactions
- Predictive modelling for duplicate detection: vigiMatch represents a scalable predictive modelling approach to identifying duplicate adverse event reports
2.5 Social Media and Patient Forums
Social media platforms and online patient forums offer a novel avenue for pharmacovigilance by providing a wealth of user-generated content on medication usage, adverse drug events, and public sentiment. Advanced techniques like natural language processing and machine learning are increasingly employed to extract meaningful information from social media data, aiding in early adverse drug event detection and real-time medication safety monitoring. Among the signals of disproportionate reporting identified both in social media and in VigiBase, up to 55.3% were identified earlier in co-occurrence NLP-processed social media.
However, the unstructured nature of social media content presents challenges in data analysis, including variability and potential biases. Despite these challenges, social media can fill critical gaps in real-time ADE detection and provide a broader understanding of public sentiment and patient experiences.
3. Benefits of Digital Pharmacovigilance
3.1 Enhanced Speed and Timeliness
AI models integrated with EHRs, insurance claims databases, laboratory systems, and spontaneous reporting platforms allow safety teams to identify unusual patterns rapidly. What once took months can now take weeks—sometimes less. A cluster of hepatic reactions flagged by an algorithm can trigger immediate investigation across multiple countries, reducing patient exposure. Digital tools offer potential benefits by providing real-time reporting, comprehensive safety data, and increased patient engagement.
3.2 Improved Signal Detection Sensitivity
AI systems can detect subtle patterns that may be missed by traditional statistical methods. Across domains, implementations most consistently enhance intake, coding, prioritisation, and the timeliness of safety assessment. Graph-based methods surface plausible associations for follow-up, and seriousness models aid risk stratification. This is particularly valuable for detecting rare ADRs, which may not be identified in clinical trials due to limited sample sizes.
3.3 Comprehensive Data Integration
Digital PV enables the integration of diverse data sources—spontaneous reports, EHRs, claims data, laboratory results, and patient-reported outcomes—creating a more complete picture of a drug’s safety profile. The FDA Sentinel RWE-DE contains linked electronic health records with claims data for over 25 million patients. The HMA-EMA RWD catalogues represent a centralised digital approach for broad access to reliable data.
3.4 Patient Engagement and Empowerment
Digital tools address significant challenges inherent in traditional systems, such as underreporting and data fragmentation. Mobile applications enable direct patient reporting, capturing patient experiences that may not be reported through traditional channels. The WHO’s VigiMobile app enables systematic collection of high-quality adverse event data, supporting patient safety and pharmacovigilance. Integrating digital solutions into pharmacovigilance can significantly enhance the timeliness and reliability of drug safety data, supporting more informed regulatory decision-making and ultimately improving patient outcomes.
3.5 Scalability and Global Reach
Digital PV tools can be deployed at scale across countries and healthcare systems. Cloud-based AI systems allow resource-constrained countries to participate in global safety monitoring without heavy infrastructure investments. The WHO’s global smart pharmacovigilance strategy offers a practical framework to help countries strengthen medicine and vaccine safety systems, promoting risk-based prioritisation, reliance, and integration into regulatory systems.
3.6 Cost-Effectiveness
While initial implementation costs may be significant, digital PV can reduce long-term costs through earlier detection of safety issues (reducing the cost of managing ADRs), improved efficiency in case processing (automating data intake and coding), and reduced duplication of effort.
4. Challenges and Limitations of Digital Pharmacovigilance
4.1 Data Quality and Completeness
Poor-quality or biased data is a significant challenge. Data quality issues include:
| Challenge | Description | Impact on PV |
|---|---|---|
| Underreporting | Estimates suggest >90% of ADRs are not reported | Significant gaps in safety data |
| Incomplete data | Missing fields in ICSRs; unstructured clinical notes | Limits ability to assess causality |
| Data fragmentation | Reporting systems are fragmented, especially at national level | Makes reporting process cumbersome and time-consuming |
| Data heterogeneity | Different coding systems, terminologies, and data formats | Hinders data integration and comparability |
AI systems are only as good as the data they are trained on. The unstructured nature of social media content presents challenges in data analysis, including variability and potential biases. Data quality, including deduplication and completeness, is essential for AI-based signal detection.
4.2 Algorithmic Bias and Fairness
Algorithmic bias is a well-recognised challenge in AI applications. Bias can arise from:
- Training data bias: If the data used to train AI models underrepresents certain populations (e.g., ethnic minorities, elderly, pregnant women), the models may perform poorly for those groups
- Selection bias: Spontaneous reporting systems inherently have selection bias—some events are more likely to be reported than others
- Confirmation bias: AI models may reinforce existing patterns without detecting truly novel signals
- Class imbalance: Rare events (especially for rare ADRs) are difficult for AI to detect
Evidence remains predominantly retrospective, with uneven external validation, underscoring the need for prospective studies, standardised reporting and calibration, fairness audits, and closer alignment with regulatory signal-management workflows.
4.3 Regulatory and Governance Challenges
| Challenge | Description |
|---|---|
| Lack of harmonised guidelines | Regulatory frameworks for AI in PV are still evolving |
| Accountability | PV accountability remains with sponsors despite AI use |
| Transparency | AI decision-making processes must be explainable |
| Data protection and privacy | Use of patient data must comply with GDPR, HIPAA, and other regulations |
| Validation requirements | AI models require rigorous validation before regulatory acceptance |
In January 2025, the FDA issued a draft guidance titled “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products,” which establishes a risk-based credibility assessment framework.
The EMA is developing guidance on AI in pharmacovigilance, with ongoing work focused on producing AI guidance for pharmacovigilance. The Council for International Organisations of Medical Sciences (CIOMS) working group focused on AI in pharmacovigilance published a specific report on the subject in 2025.
4.4 Technical Limitations
| Limitation | Description |
|---|---|
| Algorithm performance | LLM-based ADE extraction, while promising, requires further validation and evaluation of clinical integration |
| Interoperability | Lack of standardisation across different data sources and systems |
| Infrastructure requirements | AI and big data analytics require significant computational resources |
| Real-time processing | Near-real-time processing requires robust data pipelines |
4.5 Ethical Concerns
Ethical challenges—including accountability, transparency, and privacy—persist. Specific concerns include:
- Data privacy: Protecting patient confidentiality while using data for PV
- Informed consent: Whether patients consent to their data being used for PV
- Algorithmic accountability: Who is responsible when an AI system misses a safety signal?
- Digital divide: Unequal access to digital health tools may exacerbate health inequities
- Surveillance creep: Potential for PV data to be used for purposes beyond drug safety monitoring
4.6 Human Oversight Requirements
The future of pharmacovigilance is hybrid: AI for detection, humans for judgment. AI cannot replace clinical judgment. Key roles for human oversight include:
- Validating AI-generated signals: Distinguishing true safety signals from false positives
- Causality assessment: Determining whether a statistical association represents a causal relationship
- Contextual interpretation: Understanding clinical context that AI may miss
- Ethical decision-making: Making decisions about risk-benefit balance and patient communication
- Error correction: Identifying and correcting AI errors
Cross-cutting challenges include heterogeneous and shifting data, annotation burden, and concerns around transparency, privacy, and fairness.
5. Real-World Medical Examples
5.1 Example 1: Active Surveillance for Cardiovascular Risk (FDA Sentinel)
The FDA Sentinel Initiative identified cardiovascular risks associated with certain diabetes medications earlier than spontaneous reporting alone could. This active surveillance approach, using AI to analyse insurance claims, EHRs, and registry data, enabled regulatory action and updated product labelling.
5.2 Example 2: Drug-Induced Liver Injury Detection (VigiBase)
VigiBase, with its machine learning tools, has improved the detection of drug-induced liver injury (DILI) signals. By integrating AI-driven signal detection with expert review, WHO has reduced the time from initial report to global safety signal by weeks. The vigiRank algorithm prioritises emerging safety signals based on reporting patterns, report quality, and content.
5.3 Example 3: Mobile Reporting in Mauritius (VigiMobile)
In Mauritius, the national rollout of VigiMobile enabled healthcare workers and pharmacists to report ADRs and AEFIs in real time. The VigiMobile app provides real-time alerts on potential side effects of drugs and vaccines, allowing timely information and decision-making by healthcare professionals.
5.4 Example 4: Social Media Detection of Underreported ADRs
Research has demonstrated that social media can detect ADRs earlier than traditional reporting systems. Among signals of disproportionate reporting identified both in social media and in VigiBase, up to 55.3% were identified earlier in NLP-processed social media. This has particular value for capturing patient-reported symptoms that may not be reported to healthcare professionals.
5.5 Example 5: DARWIN EU in Regulatory Decision-Making
DARWIN EU has supported regulatory decision-making by providing high-quality, validated real-world data on medicine use, safety, and efficacy. A safety study requested during the review of doxycycline use and its association with suicidality ideation exemplifies how real-world evidence from DARWIN EU can inform regulatory conclusions.
6. Regulatory Landscape and Future Directions
6.1 Current Regulatory Frameworks
6.2 The AI Observatory and International Collaboration
The EU pharma regulators publish an annual AI observatory report summarising experiences and horizon scanning. EMA and FDA have articulated joint principles for the use of AI across the medicine lifecycle. The NDSG 2026-2028 workplan foresees guidance development on AI in clinical development and AI in pharmacovigilance.
6.3 Future Directions
AI offers clear opportunities across the medicines lifecycle. Key initiatives include supporting EMA’s scientific committees and the pharmaceutical industry in evaluating AI through the medicines lifecycle, developing guidance on AI in clinical development and in pharmacovigilance, fostering EU-wide and international collaboration, and providing the network with training on AI. The aim is to facilitate safe and responsible use of AI that benefits public and animal health.
7. Conclusion: The Path Forward
Digital pharmacovigilance represents a paradigm shift in drug safety monitoring. The integration of AI, big data, mobile health applications, and real-world evidence is transforming pharmacovigilance from a reactive, manual process into a proactive, predictive, and patient-centred discipline.
The benefits are substantial: enhanced speed and timeliness of signal detection, improved sensitivity for identifying rare and complex ADRs, comprehensive data integration across multiple sources, increased patient engagement, and global scalability. Digital tools offer real-time reporting, comprehensive safety data, and increased patient engagement, addressing the challenges of underreporting and incomplete data in pharmacovigilance.
However, significant challenges remain. Data quality and completeness, algorithmic bias, regulatory gaps, technical limitations, and ethical concerns must be addressed through rigorous validation, standardised reporting, fairness audits, and robust governance frameworks. The future of pharmacovigilance is hybrid: AI for detection, humans for judgment. By clarifying where AI is already dependable and where methodological and ethical gaps persist, the field can offer practical directions for integrating AI into routine PV with auditable thresholds, monitoring, and human oversight.
As the WHO’s global smart pharmacovigilance strategy emphasises, countries must strengthen medicine and vaccine safety systems through risk-based prioritisation, reliance, and integration into regulatory systems. The EMA and HMA’s joint workplan “Data and AI in medicines regulation to 2028” sets out a roadmap for leveraging large volumes of regulatory and health data as well as new tools to support regulatory decision-making.
In conclusion, digital pharmacovigilance is not about replacing human expertise but augmenting it. By combining the power of AI and digital tools with clinical judgment, regulatory oversight, and patient engagement, we can build a drug safety surveillance system that is faster, more comprehensive, and ultimately more protective of public health.
References
- World Health Organization. The global smart pharmacovigilance strategy. Geneva: WHO; 2025.
- Uppsala Monitoring Centre. VigiBase: WHO global database of individual case safety reports. Uppsala: UMC; 2025.
- Othman ZK, Nabashaho PM, Ojugbeli E, et al. Advancing pharmacovigilance through artificial intelligence: A review of applications and ethical considerations. Ann Trop Res. 2025;47(2):290.
- Getov I, Panayotova D, Dobrev S, Dimitrova M, Petkova V. Opportunities for implementing digital applications to monitor the impact of risk-minimization measures within the pharmacovigilance process: Scoping review of published literature. Pharmacia. 2025;72:1-6.
- European Medicines Agency. Data and AI in medicines regulation to 2028: Joint workplan. Amsterdam: EMA; 2025.
- European Medicines Agency. DARWIN EU®: Data Analysis and Real World Interrogation Network. Amsterdam: EMA; 2025.
- U.S. Food and Drug Administration. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (Draft Guidance). Silver Spring, MD: FDA; January 2025.
- Pharmacores. Pharmacovigilance and AI: How Drug Safety Is Changing in 2025. 2025.
- Uppsala Monitoring Centre. VigiMobile: Simplify reporting with VigiMobile. Uppsala: UMC; 2026.
- World Health Organization. Health: VigiMobile app launched to report potential side effects of medicines and vaccines. Geneva: WHO; 2025.
- Uppsala Monitoring Centre. vigiMatch: A machine learning model for duplicate detection. Uppsala: UMC; 2025.
- Council for International Organisations of Medical Sciences (CIOMS). Working Group on AI in Pharmacovigilance: Report. Geneva: CIOMS; 2025.
