The role of digital technologies in the pharmacovigilance system

Introduction

Drug safety is one of the key priorities of state policy in the field of public health. In the Russian Federation, the legal foundation for this area is established by Federal Law No. 61‑FZ of 12 April 2010 “On the Circulation of Medicines”, which mandates continuous monitoring of the efficacy and safety of all medicinal products in civil circulation [1].

The instrument for implementing this requirement is pharmacovigilance, defined in the law of the Eurasian Economic Union as activities aimed at identifying, assessing, understanding, and preventing adverse reactions (ARs) to medicinal products, as well as any other issues related to their use. This pharmacovigilance system covers the entire lifecycle of a medicinal product, including ongoing benefit–risk assessment, processing of individual case safety reports (ICSRs), and detection of safety signals [2].

Pharmacovigilance practice, against the backdrop of an ever‑expanding pharmaceutical market, faces a steady increase in the volume of ICSRs [3]. Reports originate from qualitatively diverse sources: electronic health records [4, 5], scientific literature [6, 7], patient support programmes [8], digital communication platforms [9], telephone calls, emails, and others, which results in considerable heterogeneity in format, language, and structure [10, 11]. It is also important to consider the timeliness of AR detection. According to Federal Law No. 429‑FZ of 22 December 2014, an adverse reaction is an unintended harmful bodily response that may be associated with the use of a medicinal product [12]. For example, in a prospective study by Davies et al. (2009) involving a cohort of 3,695 hospitalised patients, ARs were recorded in 14.7% of cases (95% CI: 13.6–15.9%); half of the identified reactions were classified by the authors as “definitely” or “possibly preventable”, highlighting the continuing relevance of improving tools for their timely detection as a priority in clinical practice [13].

The development of the state system for drug evaluation in the Russian Federation began in November 1990, when, at the initiative of Professor V. K. Lepakhin, the All‑Union Scientific Centre for Drug Evaluation (VNTsELS) was established—the first centralised state body combining functions of documentation evaluation, clinical trial organisation, and monitoring of adverse drug reactions. This structure served as the foundation for the modern pharmacovigilance system administered by the Federal Service for Surveillance in Healthcare (Roszdravnadzor) [14]. Further evolution of the pharmacovigilance system took place amid the progressive digitalisation of healthcare and the pharmaceutical industry, which opened up practical opportunities for applying machine learning (ML) and natural language processing (NLP) methods to automated pharmacovigilance data analysis [15, 16].

Today, in the Russian Federation, the legal basis for integrating artificial intelligence technologies into healthcare is provided by Presidential Decree No. 490 of 10 October 2019 “On the Development of Artificial Intelligence in the Russian Federation” (as amended on 15 February 2024), Federal Law No. 258‑FZ of 31 July 2020 “On Experimental Legal Regimes in the Sphere of Digital Innovations in the Russian Federation”, Preliminary National Standard PNST 961–2024, and the Code of Ethics for the Use of Artificial Intelligence in Healthcare, version 2.1, approved by the Interdepartmental Working Group under the Ministry of Health of the Russian Federation on 14 February 2025 [17–20].

At the same time, full‑scale integration of digital technologies (including the comprehensive use of ML and NLP methods) into routine pharmacovigilance practice remains fragmentary. The gap between technological capabilities and their full application is attributable to a combination of factors: the absence of validated standards for AI models in signal detection tasks, limited algorithm interpretability in a regulatory context, heterogeneity of data requirements across the national health systems of EAEU member states, and the methodological gap between the current regulatory architecture of the Union and the current level of digital solutions [21–23].

This review systematises data on the application of digital solutions in pharmacovigilance tasks, identifies regulatory gaps in the context of EAEU law, and defines directions for implementing AI methods into the practice of the competent authorities of member states.

Objective

To consolidate and systematise current evidence on the use of digital technologies in pharmacovigilance tasks, identify existing gaps, and outline pathways to address them.

Structure and Data Sources in Pharmacovigilance

According to Order No. 3518 of the Federal Service for Surveillance in Healthcare of 17 June 2024 “On Approval of the Procedure for Pharmacovigilance of Medicinal Products for Medical Use”, which came into force on 1 March 2025, pharmacovigilance in the Russian Federation is based on six categories of data: spontaneous reports from subjects of medicines circulation submitted through the automated information system (AIS) of Roszdravnadzor; periodic updated safety reports (PUSRs); periodic safety reports for a developmental medicinal product (PSR‑D); materials from federal state control; risk management plans (RMPs); and special notifications of urgent safety issues [24].

Marketing authorisation holders are required to notify Roszdravnadzor of serious adverse reactions with a fatal outcome or life‑threatening within three working days; for other serious ARs, within 15 calendar days [24, 25]. Uniformity of medical terminology across the aforementioned sources is achieved through the mandatory use of the MedDRA dictionary (Medical Dictionary for Regulatory Activities); electronic exchange of individual case safety reports is carried out in the ICH E2B (R3) format (the international standard for electronic exchange of drug safety data), ensuring compliance with international databases such as the WHO VigiBase, the FDA Adverse Event Reporting System (FAERS), the European Medicines Agency’s EudraVigilance, and others (Table 1) [26, 27].

Table 1

International and national drug safety databases with open online access

Notes: AIS – automated information system; WHO – World Health Organization; EAEU – Eurasian Economic Union; FDA – Food and Drug Administration (USA); MHRA – Medicines and Healthcare Products Regulatory Agency (United Kingdom); ANSM – Agence Nationale de Sécurité du Médicament (France); ANVISA – Agência Nacional de Vigilância Sanitária (Brazil).

Spontaneous reports constitute an important channel for obtaining information on ARs to medicinal products in the Russian Federation. According to the EAEU Good Pharmacovigilance Practice Rules, a spontaneous report is a voluntary submission by a healthcare professional or consumer to the competent pharmacovigilance authority of information containing a description of one or more ARs in a patient who has taken one or more medicinal products, provided that such information has not been obtained during a clinical trial or any other method of organised data collection [25]. The obligation of medicines circulation subjects to report identified ARs is enshrined in Article 64 of Federal Law No. 61‑FZ of 12 April 2010 “On the Circulation of Medicines” [1]. However, the existence of a regulatory requirement does not fully deliver the desired outcome; in a systematic review by Hazell L. and Shakir S. A. (2006), based on 37 pharmacoepidemiological studies from 12 countries, the median under‑reporting of ARs in target databases was 94% (interquartile range: 82–98%), which is an extremely alarming figure [28]. This discrepancy limits the ability to objectively assess the true scale of the AR problem and necessitates the expansion of data sources in the pharmacovigilance system.

The current international trend in pharmacovigilance data increasingly points to the use of real‑world data (RWD) in the broadest sense [27]. In the context of our country, these include information of a heterogeneous profile: for example, data from electronic health records (EHRs) aggregated within the Unified Medical Information and Analytical System (UMIAS) and regional medical information systems; data from territorial compulsory health insurance funds containing ICD‑10 diagnosis codes, volumes of care provided, and hospitalisation data; results of post‑registration studies; materials from patient support programmes; and other sources. At the same time, the legal mechanism for working with anonymised patient data on a pilot basis is already set out in Federal Law No. 258‑FZ of 31 July 2020 “On Experimental Legal Regimes in the Sphere of Digital Innovations in the Russian Federation” [18].

Artificial Intelligence Methods in Adverse Drug Reaction Report Processing

The transformation of the pharmacovigilance system towards digital restructuring is driven by the growth in data volumes and the need for rapid response to signals on efficacy and safety of medicinal products [29]. The use of artificial intelligence methods, in particular machine learning and natural language processing systems, in pharmacovigilance covers several functional levels:

• automated extraction of data from unstructured texts;

• classification of AR seriousness;

• safety signal detection and ensuring interpretability of algorithmic solutions for regulatory authorities [30, 31].

According to Loginovskaya O. A. et al., modern pharmacovigilance information systems are in the phase of a planned transition from basic data entry automation to AI models that perform processing and automatic coding of ARs using the MedDRA dictionary [32].

In the medical literature, the results of applying AI models to specific pharmacological groups of drugs have already been verified. In a recent study by Stanekova et al. (2024), machine learning algorithms applied to text records of electronic health records of patients receiving penicillins made it possible to distinguish true allergic reactions from drug intolerance, as well as to stratify patients by the degree of allergic risk (we see potential for the practical application of similar digital systems within the framework of automated review of erroneously assigned allergic diagnoses for certain drugs) [33]. Another example: a retrospective analysis of the Russian database “Pharmacovigilance 2.0” conducted by Samsonova K. I. and including 1,890 spontaneous reports of ARs to fluoroquinolones for the period 2019–2023 did not reveal a single case of peripheral neuropathy, dysglycaemia, aneurysm, or aortic dissection, despite the fact that these ARs are included in the current instructions for medical use [34]. In their work, the authors attribute the obtained results to the low awareness of physicians about such ARs, rather than their actual absence in patients. In this case, one can clearly see the limitations of the traditional reporting system; the observed “miss” of clinically significant information defines the practical need for using AI methods for pharmacovigilance data analysis.

A different approach to using NLP for detecting ARs from unstructured clinical patient data was presented in a retrospective cohort study by Kawazoe et al. (2025), involving analysis of information from 2,935 patients receiving anthracyclines [35]. To extract the necessary information from patient medical records, the researchers used artificial intelligence tools, namely a model based on BERT (Bidirectional Encoder Representations from Transformers). This model represents a deep learning architecture that accounts for the contextual window simultaneously to the left and right of the required information extraction point, which is particularly important for medical documents with ambiguous terminology [36, 37]. Based on this model, the authors developed a specialised NLP tool called MedNERN, fine‑tuned on Japanese medical texts and configured to extract expressions indicating cardiotoxicity from medical histories, patient charts, and other information. The key advantage of this approach is that the extracted information did not depend on the presence of ICD‑10 codes (i.e., ARs recorded in free text but not entered into structured fields were taken into account).

In parallel, digital biomarkers—objective, continuously recorded physiological measures obtained from wearable devices—are being considered as a promising tool for active safety monitoring [38]. According to the definition of the European Medicines Agency (EMA), a “digital biomarker” is a quantifiable physiological and/or behavioural indicator whose clinical significance has been established through a verified algorithm and which can be used in clinical trials to assess the efficacy and safety of a medicinal product [39]. It is important to emphasise that digital biomarkers can provide substantially more objective and continuous information about a patient’s condition compared to traditional clinical methods. For example, in cardiology, ECG recording using smartwatches or other wearable devices allows the detection of previously unnoticed rhythm changes, including drug‑induced QT interval prolongation—an AR that, according to the Russian spontaneous reporting database, is clinically significant when using fluoroquinolones and is predominantly recorded during the inpatient stage of medical care.

Current Challenges and Prospects for Digital Technology Applications in Pharmacovigilance

Post‑marketing safety monitoring of medicinal products is conducted under conditions fundamentally different from those of registration clinical trials. The real‑world patient population includes groups systematically excluded from randomised controlled trials, such as pregnant women, elderly patients, persons with severe comorbidities, and those with polypharmacy. The spontaneous reporting system for adverse reactions, which forms the basis of post‑marketing pharmacovigilance, demonstrates structural limitations with respect to rare ARs with a long latency period: the low reporting rate and the absence of a denominator (the actual number of exposed patients) do not allow the calculation of the true incidence of such reactions [40]. In this situation, big data analytics methods, used in conjunction with digital technology tools, can create a methodological basis for overcoming these limitations [41].

To date, an analysis of published data on the topic of this work reveals an interesting trend: the majority of studies are aimed at solving one narrowly defined task using digital technologies (classification of spontaneous reports, identification of “incorrect” and conversely new data, monitoring social media for informational signals in the pharmacovigilance context, etc.) and do not assess the overall effect of the applied methods across the entire drug safety management cycle [42–44]. Bate A. and Tregunno P. M. (2026), in their article on the current state of artificial intelligence in pharmacovigilance, directly note this pattern and raise the question of transitioning from a task‑oriented to a systemic approach [45].

In this context, a “systemic approach” implies embedding digital tools into all operational levels of pharmacovigilance: primary intake and processing of ICSRs; signal detection and verification; causality assessment (which is particularly critical for evaluating new molecules and drugs); and development of risk minimisation measures, among others [46, 47].

To achieve such coverage, a promising solution is the concept of agentic AI: autonomous AI modules capable of interacting with each other and with deterministic rule‑based systems within a single analytical framework [47]. The key advantage of such a hybrid architecture is that machine learning methods provide “flexibility” when working with unstructured data, while the deterministic component ensures reproducibility and transparency of results, which are critically important for regulatory purposes [48].

Another challenge is automation. The practical implementation of this direction encounters obstacles at two levels. From a technological standpoint, key issues remain the heterogeneity of input data (structured ICSR databases, free‑text clinical notes—where, unfortunately, new challenges arise: handwriting recognition, small font, image quality, etc.); the need to identify and classify systemic algorithm errors in large volumes; and ensuring the analytical validity of the models used in the context of a continuously updated evidence base [49].

At the legal level, the use of automated systems is limited by the requirements of Good Pharmacovigilance Practices (GVP), standards for handling patient personal data, and the absence in most jurisdictions of agreed procedures for regulatory qualification of AI tools [50]. Nagar A. et al. (2025), after analysing the current state of regulatory AI application, put forward the thesis that routine use of digital technologies is achievable only when three key conditions are met: compliance with the regulatory requirements of the region’s legal framework, transparency of the algorithms themselves, and continuous validation [51]. Methodologically challenging remains the issue of detecting rare adverse reactions, particularly when analysing data for new drugs and potential candidate molecules.

The global transformation of the industry inevitably entails personnel adaptation. Healthcare professionals today need more than just clinical knowledge; full integration of training systems and the development of competencies in working with digital technologies are required. Moreover, the involvement of specialists outside clinical medicine (e.g., data engineers, AI system architects, digital ethics experts) in pharmacovigilance processes reflects an objective shift in the industry’s logic towards a “cross‑functional” level, as it is commonly termed in the pharmaceutical sector. Drug safety tasks are consistently acquiring a nature that requires competencies from several professional domains simultaneously.

The totality of the presented data allows us to formulate the following conclusion: digital technologies have proven potential to improve the operational efficiency of pharmacovigilance, but their full incorporation into regulatory practice requires a revision of not only technical but also organisational evaluation criteria. The question of under what conditions an automated system can assume functions traditionally assigned to a specialist remains open. There are grounds to believe that the answer to this question will determine the future of the pharmaceutical field.

Conclusion

This work analysed the key directions and opportunities for integrating artificial intelligence into the pharmacovigilance system. The study revealed that ensuring drug safety in the modern era is impossible without the large‑scale implementation of digital solutions.

It is clear that a global competitive struggle for leadership in medical technology is currently unfolding. However, the successful widespread integration of AI solutions requires overcoming “grey zones” and systemic barriers, including issues of data interoperability and ethical responsibility.

Further development of the industry within the EAEU space is impossible without coordinated efforts of the interdisciplinary community aimed at creating a sustainable digital ecosystem. Only through the synergy of innovative algorithms and professional responsibility of experts can new standards of safety, transparency, and effectiveness of drug therapy be achieved.