AI is revolutionizing pharmacovigilance by addressing critical challenges in managing, analyzing, and interpreting vast and complex datasets [16, 24]. In the following sections, the main contributions of AI to pharmacovigilance are presented according to thematic areas of application, highlighting its potential to automate processes, enhance risk prediction, and strengthen drug safety monitoring.

AI in signal detection & pharmacovigilance automation

A foundational step in pharmacovigilance automation is the detection of duplicate reports. Ensuring the integrity and uniqueness of individual case safety reports is essential before advancing to broader tasks such as safety surveillance and signal detection. Duplicate entries can distort safety analyses and lead to misleading conclusions about a medicine’s risk profile. To address this, the Uppsala Monitoring Centre (UMC) developed the vigiMatch algorithm, which applies ML techniques to identify potential duplicates by analysing similarities in patient demographics, drug information, and adverse event descriptions [25]. UMC has applied NLP techniques to analyze unstructured ADR reports, offering a deeper understanding of their content and identifying duplicates that traditional methods might overlook [26]. By integrating AI-driven duplicate detection, pharmacovigilance systems improve data quality, enhance confidence in detected safety signals, and support more informed decision-making, ultimately benefiting public health [27, 28].

Following the elimination of duplicate reports, AI plays a central role in expanding pharmacovigilance automation. It enables the proactive identification of safety signals through ML algorithms and predictive analytics, improving risk management and patient safety [8, 29]. AI supports automated surveillance systems capable of processing large-scale datasets continuously and efficiently [30, 31]. Real-time data analysis further contributes to the early detection of safety trends, enhancing the responsiveness of pharmacovigilance activities.

A key component of these advances is data mining, which allows the analysis of massive and heterogeneous datasets to identify potential safety signals. These may include both structured and unstructured sources, such as ADR reports, EHRs, and social media platforms [10]. Computational methods have improved the capacity to detect ADRs earlier in the product lifecycle by revealing associations that are not easily detected through traditional approaches.

Data mining techniques such as neural networks, decision trees, and clustering algorithms are applied to classify and detect relationships between drugs and ADRs. Harpaz et al. demonstrated that such methods improve signal detection rates in large pharmacovigilance databases [8, 16]. Additionally, text mining plays a key role in analysing the narrative sections of ADR reports. Studies by Botsis et al. and others show that these techniques uncover patterns and linguistic cues often missed in manual review [32, 33]. Together, these AI-driven tools enhance the sensitivity and precision of pharmacovigilance systems, supporting a shift from reactive to anticipatory models of drug safety surveillance. Moreover, DL approaches have also emerged as powerful tools in this space. As highlighted by Chen et al., DL is particularly effective in handling large datasets with complex, non-linear relationships [34, 35]. Techniques like association rule mining and Bayesian frameworks have further proven helpful in identifying unexpected links between drugs and ADRs, especially in spontaneous reporting systems [36]. Convolutional neural networks (CNNs) improve detection accuracy when analyzing textual and visual data, while Recurrent Neural Networks provide insights into the temporal progression of drug safety signals using sequential datasets such as timestamped ADR reports [37, 38]. ML models also support predictive pharmacovigilance by identifying novel patterns in historical data, allowing for proactive ADR prevention [11]. Hybrid models that combine traditional disproportionality analysis with ML algorithms are increasingly being adopted to enhance signal detection by improving sensitivity and specificity without compromising interpretability [39].

The integration of AI in pharmacovigilance is also evident in automated signal detection systems. Unlike conventional retrospective approaches, these models continuously monitor incoming data from sources like EHRs, spontaneous reports, and social media [24, 40]. Key methods include NLP, clustering, and supervised learning, which allow for the real-time identification of high-priority safety signals [41, 42]. This proactive strategy contrasts with traditional methods that rely on manual data exploration or predefined queries [43]. Automated systems not only accelerate ADR detection but also enhance reliability. By analyzing data in real time, they facilitate timely risk identification and help healthcare professionals focus on interpretation and decision-making rather than laborious data review [44]. Coste et al. emphasized that such systems outperform manual techniques in both speed and scope [24] emphasized that such systems outperform manual techniques in both speed and scope [24]. Furthermore, automated tools offer a more comprehensive view of drug safety by integrating diverse data sources like reporting systems, EHRs, and patient-reported outcomes [45].

The role of AI in continuous drug monitoring is increasingly vital for patient safety. NLP, for instance, accelerates data extraction from free-text ADRs, as shown by Hu et al. [9]. Meanwhile, reinforcement learning (RL) models are gaining traction for their ability to refine detection algorithms based on feedback, making them more relevant to real-world scenarios [8]. As these technologies evolve, the incorporation of explainable AI (XAI) will be essential to ensure transparency, accountability, and regulatory compliance in pharmacovigilance systems [46].

Data integration and evidence generation

Understanding the safety and effectiveness of medications after they are marketed requires integrating data and studying RWE. The challenge is in efficiently combining and evaluating the growing amount of data provided from real-world sources, such as ADR reports, insurance data, and EHR. AI makes the integration of complex and heterogeneous data easier, which helps derive valuable insights for pharmacovigilance [47].

Multidimensional data integration

AI integrates diverse data sources, including structured and unstructured formats like medical images, clinical notes, and test results. ML and NLP techniques enable a more contextualized understanding of adverse events [48]. Additionally, AI-driven harmonization standardizes heterogeneous datasets, improving interoperability. Algorithms align terminologies from various sources, such as SNOMED CT and MedDRA, enhancing consistency and cross-referencing [49].

Real-world evidence analysis

When it comes to understanding how medications are used outside of the controlled setting of clinical trials, RWE analysis offers important information. AI makes it easier to analyze massive amounts of RWE by pointing out links and patterns that conventional analyses could miss [50]. For example, AI expedites the identification of rare or long-term harmful effects by automating safety signal detection in real-time [51]. RWE analysis is essential for this purpose. Additionally, applying AI to RWE makes it possible to detect differential safety signals and effectiveness patterns across subpopulations, such as age groups or comorbidity profiles, supporting a more individualised pharmacovigilance strategy [51].

A key advancement in RWE analysis with AI is federated learning, which enables collaborative data analysis across institutions without sharing sensitive patient information [52]. This method preserves privacy while incorporating diverse datasets, enhancing the generalizability of safety insights [53].

However, issues like inconsistent data quality and the need for transparency in AI models remain. AI is valuable for data integration and RWE analysis, processing large datasets quickly and identifying complex patterns [54].

Predictive models for adverse drug reactions and drug-drug interactions

Predictive models are revolutionizing pharmacovigilance by enabling early detection of ADRs and DDIs. These models, powered by ML algorithms, analyze large-scale datasets, including spontaneous reports, EHRs, and clinical trials [16]. Identifying patterns in patient demographics, genetic data, and pharmacological properties, provides a more proactive and precise risk assessment, which is particularly relevant given the increasing prevalence of polypharmacy, especially in older populations [55].

Different ML techniques have unique strengths for ADR and DDI prediction. Random forests are effective in identifying ADR risk factors like drug-induced liver injury [56]. DL and neural networks detect complex DDIs by analyzing molecular structures and biological interactions [57]. RNNs excel in processing sequential data, making them valuable for detecting temporal ADR patterns in longitudinal patient records [58, 59].

Algorithm selection depends on data type. Random forests excel with structured datasets, while DL models are better for unstructured data, like genetic sequences or medical images [60]. Decision trees and support vector machines are often used to classify ADR risks based on drug properties and patient variables [61], as shown in Table 1.

Table 1 Common machine learning algorithms for predicting adverse drug reactions and drug-drug interactions

Predictive models assess risks in real-time, often before clinical symptoms appear, helping healthcare professionals make informed prescribing decisions, especially in complex cases with multiple medications [62]. These models can complement post-marketing surveillance, improving patient outcomes and reducing serious ADR occurrences [63]. An illustrative example is the use of AI-based models to predict the likelihood that a reported case will be classified as serious. Although regulatory definitions of seriousness exist, in practice, many reporters rely on personal perception or patient-reported impact rather than formal criteria. Predictive models that identify key clinical or sociodemographic variables associated with case seriousness can support pharmacovigilance centres in validating reports and prioritising follow-up [64].

Predictive models face challenges, with accuracy relying on the quality and diversity of input data [65]. Biases in datasets can lead to flawed predictions, threatening patient safety if not validated. The "black-box" nature of ML algorithms, especially DL models, raises concerns about transparency and interpretability [66], key for clinical decision-making (Table 2).

Table 2 Benefits and challenges of artificial intelligence-driven predictive models for adverse drug reactions and drug-drug interactions

Predictive models will develop to incorporate new data types, such as proteomics, metabolomics, and genomes, as healthcare data volume and complexity continue to rise [67]. This will allow for the provision of individualized ADR and DDI forecasts. Big data analytics and AI integration can further improve pharmacovigilance by enabling precision medicine strategies that account for patient variability when evaluating drug risks [68]. Furthermore, XAI approaches are anticipated to enhance model transparency, boosting patient and clinician confidence [69].

Causality assessment: from traditional to AI-driven approaches

In pharmacovigilance, causality evaluation links drug administration to adverse events. Traditionally, expert judgment and the clinical differential diagnosis method, using Bradford Hill criteria, guide this process. Standard tools like the Naranjo algorithm and Bayesian inference-based methods provide semi-quantitative or qualitative support. However, these methods are time-consuming, variable between evaluators, and depend on trained professionals expertise [70, 71].

The advent of AI has transformed causality evaluation by introducing systems capable of analyzing vast and heterogeneous datasets with speed and precision. AI-based approaches, such as ML models, are designed to incorporate probabilistic reasoning, operational algorithms, and even expert-driven methodologies like introspection. For instance, NLP techniques can extract critical information from unstructured data, such as patient narratives and clinical notes, while neural networks uncover patterns in complex datasets that may not be evident through traditional methods [72]. These advancements reduce the cognitive burden on human experts, enhance objectivity, and facilitate the integration of RWE for more robust and scalable causality assessments. Importantly, XAI ensures transparency in these automated processes, enabling regulators and clinicians to trust the outputs and incorporate them into decision-making frameworks (Table 3) [73].

Table 3 Summary of tools, benefits, challenges, and the role of experts in causality assessment with artificial intelligence [3, 7, 25]Expert-defined Bayesian network in a pharmacovigilance centre

Causality attribution in ADR reporting is a critical component of the pharmacovigilance process, as it determines the likelihood that a specific drug is responsible for a reported adverse event. While various methods exist for this evaluation, the Portuguese Pharmacovigilance System employs the Global Introspection method [71, 74]. In this approach, clinical experts assess the probability of a causal relationship between a drug and an ADR based on criteria recommended by the World Health Organization (WHO). These assessments assign degrees of probability, typically ranging from “certain” to “unlikely,” reflecting the confidence level in the drug-reaction association [75].

At the Porto Pharmacovigilance Centre (UFPorto, in the Portuguese acronym), one of the regional pharmacovigilance centers within the Portuguese Pharmacovigilance System, this task is carried out by a multidisciplinary team of pharmacists and physicians. Their expertise is essential for interpreting intricate clinical data, accounting for confounding factors, and evaluating alternative explanations for adverse reactions, ensuring robust and reliable causality assessments. To enhance efficiency and ensure continuity in ADR reporting, especially during periods of limited expert availability, UFPorto implemented an AI system in 2018 based on a Bayesian network [76]. This system acts as a “proxy” for global introspection, emulating the reasoning processes of clinical evaluators. A Bayesian network is a probabilistic graphical model that represents variables and their conditional dependencies through a directed acyclic graph (Fig. 1). In the context of pharmacovigilance, it models the relationships between various factors in ADR reports, such as the suspected drug, dechallenge/rechallenge information, and the characteristics of the ADR [76].

Fig. 1

Graphical representation of the expert-defined Bayesian network implemented in a Pharmacovigilance Centre

Leveraging data from ADR reports and expert assessments spanning over 12 years at UFPorto, the AI system generates probability scores that simulate how a human expert might assess each case. By automating this process, the system ensures consistent causality assessments and supports timely reporting during periods of high workload or when clinical experts are unavailable. This innovative approach demonstrates how AI can complement traditional pharmacovigilance methods, maintaining the reliability of causality assessments while improving operational efficiency.

The implementation of the Bayesian network system at the UFPorto has introduced several advantages, significantly enhancing the efficiency and effectiveness of causality assessments. One of the primary benefits is its ability to ensure that these evaluations are completed within regulatory timelines, even during periods of limited human resources. By automating key aspects of the process, the system maintains the centre’s capacity to meet its obligations while alleviating the workload on clinical experts.

Although the Bayesian network system does not replace the expertise of human evaluators, it serves as a valuable tool for providing preliminary assessments, supporting decision-making, and ensuring the continuity of operations. A particularly noteworthy feature of this system is its ability to deliver highly comparable results to those of human experts. A study conducted by the team demonstrated a high degree of concordance between the AI’s causality attributions and the judgments made by clinical evaluators [76]. The positive predictive value for the highest probability levels was 71.4% (Definitive level) and 87.3% (Probable level).

The Bayesian network's errors were always "conservative," as whenever the network made a mistake, it assigned the case the probability level immediately below the level assigned by the specialist. These promising outcomes have bolstered the centre’s confidence in the system as a reliable adjunct to human expertise, underscoring its role in maintaining high pharmacovigilance standards.

However, this Bayesian network has an inherent limitation, i.e., it may perpetuate potential biases from the clinical expert (gold standard) since its construction was based on cases assessed by that expert. Additionally, the lack of continuous updates with new cases hinders its ability to evolve and adapt to emerging notifications. Notably, updating the network (which dates back to 2018) with ADR reports received since then would be valuable, particularly those from the 2021 COVID-19 vaccination campaign. This campaign significantly increased the volume of ADR reports across all pharmacovigilance centers, including the UFPorto, and introduced a broader range of reports, including those from consumers (non-health professionals) and reports on vaccines under additional monitoring, which were subject to intense media attention.

Another operational limitation is the network's graphical interface, which is not very user-friendly. Users require prior training to input the necessary data for the network to function and interpret its results. To address this, plans are underway to develop a more intuitive user interface, facilitating the network's dissemination to other national and potentially international pharmacovigilance centers.

Case study

The spontaneous report refers to a case submitted by a pharmacist about a male patient, aged 50, who developed urticaria associated with the use of ciprofloxacin 500 mg, with unknown therapeutic indication, with a dosage regimen of one tablet (500 mg) every 12 h. The ADR began on the same day as the administration of the suspected medication (the exact time-to-onset is unknown), lasted for one day, and required the withdrawal of ciprofloxacin and specific treatment with deflazacort. There is a reference to the concomitant use of paracetamol. No history of ADRs to any drug is known. No relevant clinical history is known ADR Outcome: Recovered.

The image below illustrates how to assess this case using the Bayesian network, and the interpretation of the result is as follows: “The network has 83.15 per cent confidence that the expert assessor would give this case the grade of Probable” (Fig. 2).

Fig. 2

Example of Bayesian network application to a real pharmacovigilance case with probability estimation

The integration of a Bayesian network system at UFPorto has enhanced the centre’s ability to meet regulatory deadlines while maintaining consistency with expert clinical reasoning. As advancements in AI progress, such systems hold significant potential to further streamline pharmacovigilance processes, providing reliable and efficient tools for causality assessment in drug safety monitoring.