The current stage of societal development is characterized by the rapid introduction of digital technologies, among which artificial intelligence (AI) occupies a special place [1]. Neural networks play a particular role in the advancement of modern intelligent systems – mathematical models that operate on the principle of biological neural systems in humans [2]. Neural networks represent a collection of interconnected artificial neurons organized into layers, where each neuron receives input data, processes it, and passes the result to the next layer. During training, the neural network adjusts the weight coefficients of connections between neurons, increasing the weight of correct answers and decreasing the weight of erroneous ones [3]. Modern neural network technologies enable the processing of large data arrays, identification of hidden patterns, prediction of outcomes, and automation of decision-making processes. Owing to their capacity for self-learning and adaptation, neural networks have found wide application in medicine, economics, industry, education, data analytics, and information technology.

Currently, there are numerous neural network architectures differing in their organization and information processing principles. Examples include feed-forward neural networks (FFNN), recurrent neural networks (RNN), long short‑term memory networks (LSTM), and convolutional neural networks (CNN) [4].

A separate and most contemporary direction in neural network development is the architecture based on the attention mechanism – in particular, the Transformer. The Transformer is a neural network architecture that is highly effective for processing data sequences. Like recurrent neural networks, Transformers are designed to handle sequences such as natural language text [5, 6].

One type of AI is large language models (LLMs), which are based on the Transformer architecture and trained on vast text corpora to perform a wide range of natural language processing (NLP) tasks, including text generation, translation, summarization, question‑answering, coding, and others [7].

The primary goal of NLP is to enable effective interaction between humans and computer systems through natural language. Modern NLP models are used in intelligent search engines, voice assistants, chatbots, automatic translation systems, sentiment analysis, document processing, and text generation [8]. Text mining encompasses a set of methods used to characterize and transform text. Within text mining, NLP represents a collection of syntactic and/or semantic processing algorithms based on rules or statistics that can be used for segmentation, extraction, or analysis of textual data. The distinction between approaches is that text mining uses words themselves as units of analysis (e.g., frequency, presence or absence of specific keywords), whereas NLP methods utilize underlying metadata, including content and phrase patterns. Both NLP and text mining are employed in several online health‑related domains, such as mental health, oncology, and infectious diseases [9].

Currently, NLP models are actively implemented in both domestic and foreign intelligent systems. Among Russian developments, platforms such as GigaChat, Alice, and Gureev Pro AI can be highlighted. Among foreign NLP platforms, the most widespread are Claude AI, ChatGPT, Google Gemini, DeepSeek, Kimi, Qwen, and Perplexity AI. These systems are based on the Transformer architecture and are capable of performing deep information analysis, generating texts, conducting intelligent searches, processing scientific materials, creating program code, and automating analytical activities [6].

The development of AI technologies opens new opportunities for automating routine pharmacovigilance processes; however, their implementation requires assessment of effectiveness, reliability, and compliance with data protection requirements. The relevance of this problem is driven by the growing volume of medicinal product safety information and the need to improve the quality and speed of its processing. Research into the potential of NLP systems and other AI models is of great importance for improving drug safety monitoring processes and further advancing digital technologies in the pharmaceutical industry.

When assessing the potential for integrating AI models into PV processes, it is necessary to consider a number of significant aspects, detailed in the report of the CIOMS (Council for International Organizations of Medical Sciences) Working Group XIV, "Artificial Intelligence in Pharmacovigilance" [10].

Among these, one fundamental aspect is data confidentiality protection. In their professional activities, PV specialists work with patient personal data, and current legislation, regulations, and guidance documents provide a set of measures aimed at ensuring confidentiality, anonymity, autonomy, and control over sensitive and potentially identifiable medical data. In addition, a significant part of PV activities involves internal company documents, which, under confidentiality agreements, are not to be disclosed to third parties.

The combination of these restrictions precludes the use of publicly available solutions such as Claude, ChatGPT, Google Gemini, and others, despite the fact that these models represent ready‑to‑use tools with a wide range of functionalities that do not require lengthy training. At the same time, a number of PV processes involve working with open data, which permits the use of the aforementioned models. These processes are discussed below.

Literature search is one of the processes covering virtually all areas of PV: causality assessment, signal analysis, preparation of Periodic Benefit‑Risk Evaluation Reports (PBRERs) and Risk Management Plans (RMPs), development of educational materials as part of additional risk minimization measures, benefit‑risk assessment, preparation of responses to regulatory authority queries, and several other tasks.

It should be emphasized that this section does not address the use of AI for weekly or periodic monitoring of the scientific medical literature. Rather, it concerns the possibility of thematic literature searches based on predefined criteria.

This direction is also the only one in this article where we consider the application of semantic search models rather than NLP models.

The process of finding the required literature is often carried out manually. The specialist turns to bibliographic databases (PubMed, Google Scholar, CyberLeninka, eLIBRARY), formulates a query using keywords, filters, and Boolean operators (AND, OR, NOT), and then sequentially reviews titles and abstracts to assess relevance. This process is quite labor‑intensive. One may need to screen a significant number of publications, most of which do not meet the selection criteria, or encounter few or no publications at all. As a result, one has to broaden the query, remove filters, or generalize the criteria, which inevitably leads to additional work that does not always yield the desired result.

The use of AI models can significantly optimize this process, primarily by eliminating the stage of exhaustive screening of all suggested publications. Compared with manual searching in bibliographic databases, the application of AI models offers a number of significant advantages:

Thus, the use of AI models for scientific literature search and systematization can greatly enhance the speed and efficiency of one of the most labor‑intensive auxiliary processes in PV.

The next step after finding a relevant publication is reading it to extract the required data. Experienced specialists typically do not read the entire article but purposefully search for the needed information using keywords or directly consult the corresponding section. Nonetheless, even this approach takes time and does not eliminate the risk of missing important data, especially when the sought‑after information is embedded in large tables or figures, or when the specialist works under high workload.

The use of LLMs can considerably accelerate this process. It suffices to formulate a query specifying the required data, after which the model either presents a summary of the requested information or reports its absence in the article. An additional advantage of this approach is the ease of verification, as any fragment generated by the model can be checked by locating the corresponding text in the original publication. A practical example of this scenario is the extraction of data for the risk characterization in Section 16.4 of the PBRER or Module CVII of the RMP (e.g., risk factors, time from drug initiation to adverse reaction onset, specific clinical manifestations, outcomes, and reversibility measures).

Another scenario is the preparation of a publication summary, for example, for Sections 11, 17.1, 17.2, or 18.1 of the PBRER, or Module CI of the RMP. In this case, it is sufficient to specify which aspects of the article should be reflected in the summary. It should be noted that the writing style generated by the model in the first iterations may not fully meet the requirements of the specific document. To achieve acceptable results, preliminary model calibration is required. When working with open solutions, in most cases, it is enough to manually edit several texts, upload them as examples, and instruct the model to adhere to the given style. After such calibration, the model becomes a ready‑to‑use tool that can significantly reduce the time spent on document preparation.

The use of LLMs in this manner is also possible for working with internal company documents, for example, for extracting data on adverse events recorded during a clinical trial from the corresponding report. However, in this case, the use of open models is inadmissible; internal solutions trained on corporate data with an established system of periodic quality control are required. This aspect is discussed in more detail below.

The number of ICSRs received by pharmacovigilance departments of pharmaceutical companies is steadily increasing. At the same time, the quality and volume of the data they contain vary considerably: from brief reports with a minimum set of information (patient demographics and description of the adverse reaction) to extensive unstructured texts or copies of documents in PDF or JPG format, including complex laboratory values and clinically significant information that a specialist might overlook during manual processing. Examples of such information include discrepancies between the actual indication for use and the registered indication (e.g., off‑label use for renal cell carcinoma due to erroneous interpretation caused by visual similarity of terms with the registered indication – hepatocellular carcinoma), or dosing errors relative to patient body weight, etc. Additional strain arises during periods of sharp increases in the flow of ICSRs (as observed during the coronavirus pandemic), which, with limited staff, inevitably lead to reduced quality of case processing and the risk of missing clinically important information. At the same time, pharmaceutical companies retain the obligation to respond promptly to adverse reaction reports in order to protect public health and to rapidly identify and address safety issues.

Previously, to address these tasks, a Named Entity Recognition (NER) approach was used – a technology that identifies and classifies key information elements in unstructured texts (names of adverse reactions, reporter qualification, country, specific terms) using predefined markup. However, in practice, this approach did not consistently improve ICSR processing [10].

The application of LLMs opens new possibilities for accelerating ICSR processing while maintaining quality. A well‑trained LLM can extract from the source data (e.g., E2B files, CIOMS forms, scientific literature publications) a predefined list of information: the four minimum criteria for a valid ICSR, medical history, date of initiation of the suspect drug, time of onset and resolution of the adverse reaction, outcome, and concomitant therapy. Based on the extracted data, the model can perform case analysis, causality assessment, and identify missing information that needs to be requested for a comprehensive evaluation. As a next step, the LLM generates a structured case summary in a predetermined format and writing style (e.g., following the anamnesis morbi logic).

The CIOMS XIV Working Group has already described in its report [10] a practical scenario for using LLMs for data extraction in ICSR processing. It presented the results of an experimental study in which a semi‑automated ICSR processing workflow was established, consisting of three stages. In the first stage, input documents in various formats were harmonized. Then, the GPT‑4 model extracted data using a structured JSON template containing hints about the likely location of information in the document. In the third stage, the textual responses from the model were automatically mapped to predefined field values. To evaluate efficiency, a graphical interface was developed that allowed comparison of fully manual processing with the AI‑assisted process. Four experienced specialists were randomly assigned to the two workflow variants, and the system recorded the net processing time.

According to the results, extraction accuracy ranged from 85–100% for clinical trial documents, 60–100% for patient support program reports, and 67–100% for cases from the medical literature. The use of LLMs reduced processing time by 39%, saving on average about 20 minutes per case. Of the approximately 69 data points extracted by the model, only 2.4 required manual correction [10].

Thus, this approach can significantly shorten ICSR processing time, reduce the risk of missing clinically significant information, and maximally standardize the style of case descriptions.

An important area of PV is the preparation of information materials for healthcare professionals (e.g., physician guides) and patients (e.g., patient leaflets) as part of additional risk minimization measures.

According to the requirements of Good Pharmacovigilance Practice [11], educational tools must be adapted to the needs of the target audience. In practice, this means ensuring that the text is understandable both for healthcare professionals and for patients without medical training. Adapting professional medical text containing highly specialized terminology to a patient‑accessible format presents some difficulties even for experienced specialists and requires time. For products with innovative mechanisms of action (e.g., gene therapies), preparing accessible educational material may also pose challenges for the physician audience.

The use of LLMs can substantially accelerate this process. The model can promptly formulate complex descriptions, adapting language and style to the specific target audience – from healthcare professionals to patients. Similar to the ICSR processing described above, LLMs can generate structured text based on available source materials (e.g., the Summary of Product Characteristics, scientific publications, other information sources) according to a predefined structure or template. This approach is also applicable when preparing the safety sections of the patient package leaflet. The LLM can adapt the text from the Summary of Product Characteristics, simplifying the medical narrative while fully preserving its meaning.

Thus, the use of LLMs in preparing information materials can significantly reduce the time spent on creating and adapting texts for various target audiences. This approach is particularly valuable in complex therapeutic areas where adapting professional medical text for patients is a labor‑intensive task.

One of the key objectives of PV is to assess whether the benefit of a medicinal product outweighs its potential risks, based on accumulated safety data. The integration of NLP into the benefit‑risk assessment process represents a promising direction for PV development.

To date, many methods have been developed for evaluating the potential benefits and risks of drug therapy, classified as quantitative, semi‑quantitative, and qualitative. However, in practice, qualitative methods predominate, among which the most widespread are PrOACT‑URL, the Benefit‑Risk Framework, the BRAT (Benefit‑Risk Action Team) model, and several others [12]. Despite methodological differences, all these approaches can be integrated with NLP systems.

At the same time, among all the application areas of AI considered in this article, this direction is the most labor‑intensive in terms of implementation. Qualitative methods typically involve answering a defined set of questions, for example:

To answer each of these questions, the model requires specific training data: the Summary of Product Characteristics, clinical study reports, epidemiological data, clinical guidelines, and other sources. Thus, developing such a model requires considerable time for data collection, training, and subsequent quality validation.

Nevertheless, the implementation of such a model allows automation of the routine part of the work, namely the analysis and systematization of large data volumes with subsequent description, leaving the most critical stage – final weighing and decision‑making – to the specialist. This approach ensures transparency of the assessment for regulatory authorities and contributes to improving the quality, objectivity, evidence‑based nature, reproducibility, and reliability of the evaluation, reducing the influence of human bias, where a specialist might unintentionally focus on data supporting their view of the product. It is fundamentally important that AI does not replace the expert judgment of the specialist but augments it, concentrating human resources on the stages that require professional and ethical responsibility.

An example of the application of an AI model for a similar scenario is presented in the CIOMS report [10]. The Working Group investigated the applicability of AI models for automating, standardizing, and supporting expert decision‑making in causality assessment. For training and testing, a large heterogeneous dataset was used, combining spontaneous reports, clinical trial data, and medical literature. The dataset included patient demographics, medical history, comorbidities, a wide range of drugs and vaccines, as well as various types of adverse events coded according to MedDRA terminology. The model architecture involved feature extraction from structured fields and textual narratives using NLP and LLMs, computation of surrogate metrics, such as the Naranjo scale or WHO criteria, and final classification of causality from "doubtful" to "certain."

This model demonstrated high reproducibility of results, significantly reducing subjective factors. The AI model effectively accounted for hidden co‑factors such as age and concomitant therapy, which are often overlooked in manual analysis. The AI model successfully classified cases by uncertainty level: obvious and well‑documented reactions were processed automatically, while complex and borderline cases were referred for in‑depth expert consultation [10].

Although this example pertains to causality assessment rather than benefit‑risk analysis directly, it clearly demonstrates the fundamental applicability of a similar approach in this field. The process logic is largely analogous: training the model on a large array of heterogeneous data followed by filling in predefined fields based on that data. The difference is only that, in the case of benefit‑risk analysis, these fields are defined not by the Naranjo scale or WHO criteria, but, for example, by the PrOACT‑URL framework structure.

All documents in the PV field must comply with established requirements and follow a defined structure, including mandatory sections regulated by documents of the relevant regulatory authorities and industry standards (EAEU, CIOMS, FDA, EMA, ICH).

Although the main requirements for document structure are harmonized, certain aspects may vary, creating additional difficulties for pharmaceutical companies operating in multiple countries with different regulatory bodies. The need to thoroughly study each regulatory document and template to identify such differences is a labor‑intensive and time‑consuming task, carrying the risk of missing important requirements and consequently leading to rejections and resubmissions.

One tool to address this problem is the use of NLP models in two main scenarios.

The first scenario – verification of compliance with formal structural requirements. NLP models with template recognition or specialized checklists integrated with LLMs can verify the presence of all mandatory sections and identify inconsistencies with CIOMS, FDA, and other regulatory authority formats.

The second scenario – verification of document content compliance. An NLP model can analyze the document text and compare it with the requirements of regulatory guidance. For example, it can check whether the PBRER structure complies with ICH E2C(R2) recommendations, whether the requirements of Module IX of Good Pharmacovigilance Practice are reflected in the safety signal description, and whether the seriousness of the adverse reaction and causality are correctly interpreted. This approach is of particular practical value in situations where the same ICSR needs to be submitted simultaneously to Roszdravnadzor, FDA, and EMA, each with its own requirements regarding language, format, and level of detail.

The introduction of NLP models into document verification processes allows the PV specialist to focus on the substantive and medical components of the document, without being distracted by formal compliance verification. A reduction in regulatory queries due to formal inconsistencies is also expected. In the event of such queries, NLP models can additionally be used to analyze and systematize the regulator's comments, thereby forming an expanded requirement base that will be taken into account in subsequent document checks.

The previous sections discussed the advantages of integrating AI models into the practical work of PV specialists, namely: time savings in routine tasks, increased standardization and quality of documents, the ability to focus on expert analytical work, and reduction of human error when processing large data volumes. At the same time, the use of AI models entails a number of significant limitations and risks, which are examined in this section.

Risk of errors and unreliable generation ("hallucinations"). Despite the gradual improvement in the quality of modern LLMs, they remain susceptible to so‑called "hallucinations" – the generation of plausible but factually unreliable responses containing fabricated facts, false references, or logical errors. In the PV context, this can have serious consequences, including: incorrect assessment of case seriousness and consequent delayed submission of ICSRs to regulatory authorities, detection of false signals or missing existing ones, incorrect benefit‑risk assessment, inaccurate wording in patient leaflets or package inserts that may mislead patients. Since the accuracy of information in PV directly affects patient safety and the timeliness of safety issue identification, the risk of unreliable generation must be mandatorily taken into account when planning the use of LLMs [10].

A similar risk exists when using semantic search models for literature systematization. In the absence of publications meeting the specified criteria, the model may generate a non‑existent article with plausibly indicated authors, journal name, and even a DOI identifier.

Non‑determinism. Generative models are non‑deterministic – they produce different responses to the same query without any change in the model itself. This complicates verification of system stability [10].

Dependence on data quality and quantity. The performance of AI models depends on the quality of training data. However, PV data are often characterized by inconsistency, incompleteness, coding errors, and high variability. Moreover, AI models may show unsatisfactory results for certain population subgroups (by race, gender, age, or socio‑economic factors) or for rare adverse reactions if these were underrepresented in the training set. All of this reduces the accuracy of AI models [10].

Transfer and amplification of human bias. Since AI models are trained on human‑generated data, they will adopt and possibly amplify hidden or overt human biases and discriminatory patterns [10].

Risk of confidential data leakage. As noted earlier, the use of open AI models when working with patient personal data and internal company documentation is categorically unacceptable. At the same time, even when developing an in‑house AI model operating within the corporate infrastructure, it is not possible to completely eliminate the risk of data leakage [10].

The limitations and risks described above are critically important when planning the implementation of AI models into routine PV processes and necessitate validation, systematic fact‑checking, quality control of training and subsequent generated data, as well as the implementation of technical measures to prevent unauthorized access and information leaks. These measures are discussed in detail in the next section.

For the full‑scale integration of AI models into PV specialist activities, it is necessary to ensure the validity and reliability of the decisions they make. Validity means that the model achieves its intended purpose within acceptable performance parameters [10]. This requires preliminary definition of acceptable performance metrics, selection of representative data for training and testing, evaluation of model effectiveness under realistic conditions, and its integration into the existing quality assurance system. Reliability, in turn, means stable achievement of the stated objectives despite variability in input data. In its report, CIOMS identifies five key aspects that must be considered when implementing an AI model [10]:

An essential condition for the deployment of AI models in PV processes is ensuring their transparency [10]. Open communication is required regarding when and how AI models are used to address key tasks. The description of the applied solutions must be exhaustive: the model architecture, the nature of expected inputs and outputs, as well as the level and form of human‑AI interaction should be disclosed.

A particular form of transparency is explainability – the ability to understand how the system arrived at a specific result. The practical value of explainability lies in enabling users to understand the reasoning behind an AI model's response and, if necessary, to challenge the result.

At the same time, a significant limitation of generative LLMs must be considered: although such models can produce outwardly persuasive justifications for their outputs, these justifications are linguistic constructs, not reflections of the model's actual internal computations, and are essentially post‑hoc rationalizations. Therefore, for PV applications, the Retrieval‑Augmented Generation architecture is preferred, which prevents the model from "fantasizing" by forcing it to seek answers strictly within an isolated document base. Another technical solution to combat hallucinations, false argumentation, and non‑determinism is source grounding – requiring the model to provide justifications tied to specific evidence (with references to fragments of input data or original documents). Before incorporating LLM‑generated justifications into the audit trail, it is recommended to conduct tests for their veracity [10].

Regarding data quality, one requirement for the formation of training data is the structural completeness of each training item. Such an item must contain a predetermined set of mandatory elements, and if at least one is missing, the training item is deemed unfit for inclusion in the training set, regardless of the quality of the remaining content.

To minimize the transfer of human biases and errors, it is possible to involve multiple specialists in the assessment of training data quality, as well as specialists with different professional profiles (e.g., clinical pharmacologists, regulatory experts, and drug safety specialists). Discrepancies in assessments will serve as a basis for correction of the training material before its inclusion in the training set.

However, there is currently no ready‑made solution for the problem of insufficient data quantity for training AI models. The use of synthetic data in PV cannot be considered safe and reliable by default. Therefore, in situations where the volume of available training data is insufficient, a more justifiable approach is a temporary deferral of AI model application in that area, combined with the systematic accumulation of necessary data for subsequent model re‑training.

Ensuring data confidentiality when working with AI models requires the implementation of a set of measures. According to CIOMS recommendations [10], the main protective measures are anonymization and pseudonymization of data. A practical example is training algorithms on de‑identified ICSRs in which personal data (patient initials, addresses, etc.) are replaced with surrogate identifiers or only the patient's age in full years is provided instead of the exact date of birth. Also, when preparing training data, it is necessary to ensure that they contain only the necessary information and lack extraneous information that is not needed for model training.

Other protective measures include the use of closed local models operating within a secure perimeter, monitoring of queries and data, regular audits of input and output data, implementation of data retention and deletion policies, incorporation of data protection requirements into contractual agreements with contractors, conducting data protection impact assessments, designating a data protection officer, and ensuring regular staff training and audits [10].

For the successful implementation of all the above aspects, one or two people are insufficient. It is necessary to form a multidisciplinary team. Successful integration of AI into PV processes requires interdisciplinary expertise at all stages of the system lifecycle – from development to daily operation. This is achieved through close collaboration among specialists from various fields: PV, quality assurance, data science, information technology, platform analytics, ethics, law, personal data protection, and project management [10].

It is fundamentally important that the employees directly involved, for example, in ICSR processing, participate in the design, testing, and implementation stages of AI models. This is a necessary condition to ensure that the developed solutions meet the real tasks and expectations of the end users.

Even a well‑trained and validated AI model, developed on a large and diverse dataset, requires constant human oversight. The level of control over the AI model should correspond to the potential consequences of its errors or distorted outputs [10]. Depending on the nature of the observed activity and the degree of system autonomy, three forms of human oversight are distinguished:

In the initial stages of implementation, a high level of human involvement (HITL) is foreseen. As confidence in the system's stable performance grows, the degree of PV specialist involvement in quality control of the generated outputs may be gradually reduced.

The aspects outlined above are not exhaustive (especially regarding validation) but can serve as a basis for the initial stage of implementing AI models into routine PV processes.

AI models, and in particular NLP models, possess significant potential for integration into routine PV processes, as they can substantially reduce the time spent on repetitive tasks, enhance the standardization and quality of documents, and reduce human error. All this allows PV specialists to devote more time to expert analytical work.

At the same time, successful integration of AI models is impossible without a systematic approach to managing associated risks. It is fundamentally important that AI does not replace the professional judgment of the specialist but rather serves as a support tool. Responsibility for the decisions made remains with humans.

Implementing AI models into routine PV processes is not a one‑time solution. It is a continuous process requiring interdisciplinary collaboration, clear task specification, systematic validation, and gradual growth of trust in the system based on verified results. Subject to these conditions, AI can become a robust tool for increasing the efficiency of PV processes.