Signal Detection and Temporal Analysis of Aromatase Inhibitor‐Associated Hepatotoxicity: A Pharmacovigilance Study Integrating Bayesian Belief‐Propagation Network and Frequentist Metrics

ABSTRACT

Background

This study assessed the real‐world hepatotoxicity of third‐generation aromatase inhibitors (AIs) for breast cancer using pharmacovigilance approaches.

Methods

FAERS data (Q1 2004–Q1 2025) were analysed using a data‐driven disproportionality analysis framework incorporating traditional frequentist metrics and an information‐theoretic Bayesian network.

Results

A total of 24 liver‐related adverse events and 7 clinical outcomes were extracted in this study. Letrozole associated with the highest number of hepatotoxicity cases and showed the highest BCPNN‐supported reporting signal. Exemestane exhibited the earliest hepatotoxicity onset (median 49 days), significantly earlier (p = 0.047) than anastrozole (61.5 days) and letrozole (56 days).

Conclusion

AIs present distinct hepatotoxic profiles. Exemestane requires early vigilance due to its rapid onset, while letrozole exhibits the highest signal. Proactive liver monitoring and individualised management are crucial. Future research should integrate artificial intelligence with multi‐modal real‐world data for predictive risk assessment.

1. Introduction

Aromatase inhibitors (AIs) are important agents in the treatment of oestrogen‐dependent breast cancer, particularly in postmenopausal patients with early breast cancer, and are usually required for 5–10 years of treatment [1, 2].

Currently, highly active and selective third‐generation aromatase inhibitors (AIs), including steroid AIs (exemestane) and non‐steroid AIs (letrozole, anastrozole), are commonly used in clinical practice [3, 4, 5]. Steroid AIs contain androstenedione, which can irreversibly bind to aromatase and inhibit its activity [6]. Non‐steroid AIs compete with endogenous substrates for the active site of aromatase, reversibly bind to it, and inhibit its activity [3].

Although AIs have shown remarkable efficacy in the treatment of breast cancer [7, 8], there have been many reports of adverse events (AEs) related to AIs, including hot flashes, bone loss, arthralgia, increased risk of cardiovascular disease, and metabolic imbalance [9, 10]. Besides, their potential risk of hepatotoxicity should not be overlooked. AIs have a higher incidence of hepatotoxicity compared with tamoxifen [11]. Reports of AI‐induced hepatotoxicity have been gradually increasing in recent years [12, 13, 14, 15], suggesting the need for an in‐depth study of this issue. Hepatotoxicity can lead to interruptions in patients' treatment, affecting therapeutic outcomes and increasing healthcare costs [16].

The FDA Adverse Event Reporting System (FAERS) database (https://www.fda.gov/drugs/drug‐approvals‐and‐databases/fda‐adverse‐event‐reporting‐system‐faers) is the largest adverse events (AEs) database in the world, including all adverse event information and medication error information collected by the FDA. With the exponential growth of real‐world medical data, traditional pharmacovigilance methods relying solely on frequentist statistics (e.g., ROR) may lack robust sensitivity and precision when detecting complex or rare safety signals natively hindered by high noise levels. To address these limitations, early neural‐network‐inspired Bayesian methods were introduced into pharmacovigilance. Specifically, the Bayesian Confidence Propagation Neural Network (BCPNN) utilises Bayesian shrinkage to calculate Information Components (IC). While not functioning as a modern multi‐layer predictive deep neural network, BCPNN pioneers the transition from rigid tabular counts to probabilistic machine intelligence, stabilising safety signal detection in highly noisy unstructured reporting systems.

In this study, we conducted a pharmacovigilance analysis based on information on liver‐related AEs associated with AIs (exemestane, letrozole, and anastrozole), which was collected from the FAERS database. By applying a dual computational framework combining traditional statistical bounds and Bayesian probabilistic approach, BCPNN, our study aimed to clarify their differences in hepatotoxicity and provide data‐driven guidance for clinical medication in special populations.

2. Methods

2.1. Data Source and Data Preprocessing

All available data from the first quarter of 2004 (Q1 2004) through the first quarter of 2025 (Q1 2025) were accessed from the FAERS online public platform (available at https://fis.fda.gov/extensions/FPD‐QDE‐FAERS/FPD‐QDE‐FAERS.html, accessed on 22 April 2025). Ethical approval and informed consent were not applicable in this study, as FAERS is a publicly accessible database and records are de‐identified. During the data deduplication process, the latest FDA_DT (the date of FDA case receipt) was retained for reports sharing identical CASEIDs (the number for identifying a FAERS case). When both CASEID and FDA_DT were identical, the report with the highest PRIMARYID (the unique identifier for FAERS cases) was selected [17]. The search terms ‘Anastrozole’, ‘Letrozole’ and ‘Exemestane’ were filled in the PROD_AI field, or we imported ‘Arimidex’, ‘Anastrozole’, ‘Letrozole’, ‘Femara’, ‘Exemestane’ and ‘Aromasin’ in the DRUGNAME field, and limited them to Primary Suspect Drug (PS). We also extracted the time of occurrence data for further analysis by extracting the START_DT field, defined as the date of medication initiation, the EVENT_DT field, defined as the date of event occurrence, and the time to onset of hepatotoxicity, defined as the interval between the date of medication initiation and the date of event occurrence.

The preferred terms (PT) were employed as keywords under the heading of ‘Liver‐related investigations, signs and symptoms’ in the Standardised MedDRA Queries SMQ in the MedDRA 27.0 Dictionary, to broadly capture potential hepatic safety signals. Following the exclusion of PTs that were unrelated to the drugs in question, a search was conducted in the PT field, as illustrated in Figure 1.

FIGURE 1.

Flowchart for obtaining and filtering reports in the FAERS database. PT, preferred term.

2.2. Statistical Analysis

Patient characteristics were described in terms of frequency and percentage. The odds ratio is calculated using the dual disproportional analysis based on the Reporting Odds Ratio (ROR) and the probabilistic Bayesian disproportionality approach, Bayesian Confidence Propagation Neural Network (BCPNN). The BCPNN algorithm utilises the Information Component (IC), which assesses the ratio of the observed co‐occurrence of drug‐event pairs to the expected co‐occurrence under the assumption of random independence; a higher IC value indicates a stronger association. A signal is defined to satisfy both the lower 95% CI of ROR > 1, n ≥ 3 and the lower 95% CI of IC calculated by BCPNN > 0. Any PT that met all the pre‐established criteria was documented as an adverse reaction signal and included in the study. We also analysed time‐to‐onset subgroup by drug. Survival analysis and chi‐square test were performed using SPSS 25.0, p < 0.05 was defined as statistically significant, and the data were visualised using R version 4.3.2.

3. Results

3.1. Demographic and Regional Characteristics of This Study

22 775 812 reports were extracted from the FAERS online public platform from Q1 2004 to Q1 2025. A total of 1426 cases of the target preferred term (PT) were obtained, comprising 24 PTs and 7 clinical outcomes. There were 237 liver‐related AE reports for anastrozole (16.6% of total cases), 920 for letrozole (64.5% of total cases), and 269 for exemestane (18.9% of total cases). The majority of the cases were female patients, which is associated with breast cancer as the main indication for AIs, and most of these women were 45–64 years old (34.2% of total cases) (Table 1). Within 1426 cases, age data were missing in a significant proportion (37%). Due to the inherent constraints of spontaneous reporting databases such as FAERS, these deficiencies and limitations were inevitable.

TABLE 1.

Characteristics of reports of AIs associated hepatotoxicity.

Characteristics	Anastrozole (n = 237)	Letrozole (n = 920)	Exemestane (n = 269)
Gender
Female	219 (92.41%)	877 (95.33%)	244 (90.71%)
Male	5 (2.11%)	10 (1.09%)	1 (0.37%)
Not specified	13 (5.49%)	33 (3.59%)	24 (8.92%)
Weight (kg)
< 60	36 (15.19%)	113 (12.28%)	37 (13.75%)
60–80	53 (22.36%)	182 (19.78%)	55 (20.45%)
> 80	25 (10.55%)	74 (8.04%)	14 (5.20%)
Not specified	123 (51.90%)	551 (59.89%)	163 (60.59%)
Median (kg)	68 (44.5–101)	67 (38–167)	63.75 (40–143)
Age (year)
21–44	10 (4.22%)	89 (9.67%)	8 (2.97%)
45–64	81 (34.18%)	269 (29.24%)	109 (40.52%)
≥ 65	60 (25.32%)	192 (20.87%)	78 (29.00%)
Not specified	86 (36.29%)	370 (40.22%)	74 (27.51%)
Median (year)	62 (35–98)	59 (21–91)	62 (39–92)
Occupation of reporters
Consumer	37 (15.61%)	151 (16.41%)	60 (22.30%)
Physician	124 (52.32%)	490 (53.26%)	119 (44.24%)
Pharmacist	36 (15.19%)	191 (20.76%)	52 (19.33%)
Other health‐professional	26 (10.97%)	68 (7.39%)	24 (8.92%)
Not specified	14 (5.91%)	20 (2.17%)	14 (5.20%)
Reported countries
Germany	61 (25.74%)	338 (36.74%)	49 (18.22%)
United States of America	42 (17.72%)	68 (7.39%)	38 (14.13%)
Others	134 (56.54%)	514 (55.87%)	182 (67.66%)

Note: Median age and weight calculations were performed based on a complete‐case analysis, excluding ‘not specified’ or missing entries.

The clinical applications of the three AIs are comparable, yet the number of liver‐related AEs associated with the three drugs included in this study is markedly disparate. A notable increase in the number of reports pertaining to letrozole has been observed since 2017 (Figure 2), predominantly from Germany. The AE data for this study came mainly from doctors, pharmacists and patients. Among them, doctors reported the largest number of cases, including 52.3% of anastrozole cases, 53.3% of exemestane cases, and 44.2% of letrozole cases (Table 1). Most of the liver‐related AEs of the three drugs came from European and American countries, with the largest number of reports from Germany, followed by the United States. It is notable that a considerable number of reports do not specify the country where the incident occurred (Table 1).

FIGURE 2.

Annual report cases of AIs associated hepatotoxicity. Data for 2025 only includes the first quarter, and the dotted line does not represent the trend from 2024 to 2025.

3.2. Reporting Disproportionality Signal Detection Between Liver‐Related AEs and AIs

We conducted disproportional analysis based on ROR and the Bayesian neural network‐based signal detection algorithm (BCPNN), as shown in Table 2.

TABLE 2.

Signal strength of liver‐related AEs: Dual analysis using ROR and BCPNN.

Aromatase inhibitors	Reported number	ROR (95% CI)	IC (IC025)
Anastrozole	560	1.33 (1.22–1.44)	0.41 (0.28)
Letrozole	1896	3.09 (2.95–3.24)	1.60 (1.53)
Exemestane	485	2.90 (2.65–3.17)	1.51 (1.38)

Note: Reported number refers to total drug‐event combinations.

Abbreviations: IC: information component; ROR: reporting odds ratio.

3.3. Reported Time‐to‐Onset Patterns of Liver‐Related AEs Associated With AIs

A total of 849 cases reported the date of administration and AE occurrence at the same time and 73 cases of abnormal date (the date of occurrence was earlier than the date of medication) were excluded. A total of 776 cases were included, as shown in Figure 3. AIs associated with liver‐related investigations, signs, and symptoms occurred at a median time of 56 days (95% CI: 46.5–60 days). The median reported time to onset of liver‐related AEs associated with anastrozole was 61.5 days (95% CI: 34–119.5 days), letrozole was 56 days (95% CI: 44–60 days), and exemestane was 49 days (95% CI: 35–64 days), with a statistical difference (p = 0.047).

FIGURE 3.

AE‐free time curves and risk tables of AIs associated hepatotoxicity.

3.4. Clinical Outcomes

According to MedDRA 27.0, we performed disproportionality analyses to exclude unrelated PTs (Figure 4). As a result, 24 types of PTs involving liver‐related investigations, signs, and symptoms were screened out of 2328 PTs from 1426 cases. The top 5 frequently reported PTs were: ‘Alanine aminotransferase increased’ (360), ‘Aspartate aminotransferase increased’ (335), ‘Gamma‐glutamyl transferase increased’ (305), ‘Ascites’ (283), and ‘Blood alkaline phosphatase increased’ (251). The most frequently reported PT of anastrozole was ‘ascites’, while the most frequently reported PT of letrozole and exemestane was ‘alanine aminotransferase increased’, as ascites are nonspecific in oncology populations, this finding should be interpreted cautiously. The PTs of AI‐related hepatotoxicity are shown in Table 3.

FIGURE 4.

ROR and IC₀₂₅ of AIs associated hepatotoxicity. IC: information component derived from BCPNN, a signal which is defined to satisfy both the lower 95% CI of ROR > 1, n ≥ 3 and the lower 95% CI of IC calculated by BCPNN > 0; PTs: preferred terms; ROR: reporting odds ratio.

TABLE 3.

The PTs in reports of AIs associated hepatotoxicity.

PTs	Anastrozole	Letrozole	Exemestane	Total
Alanine aminotransferase abnormal	7	11	0	18
Alanine aminotransferase increased	0	297	63	360
Ascites	60	193	30	283
Aspartate aminotransferase abnormal	0	10	0	10
Aspartate aminotransferase increased	0	276	59	335
Bilirubin conjugated increased	0	13	0	13
Blood bilirubin increased	0	128	27	155
Blood cholinesterase decreased	0	3	0	3
Gamma‐glutamyl transferase abnormal	0	6	0	6
Gamma‐glutamyl transferase increased	37	217	51	305
Hepatic function abnormal	0	0	59	59
Hepatic pain	10	0	5	15
Hepatomegaly	0	29	12	41
Liver function test abnormal	40	0	35	75
Glutamate dehydrogenase increased	3	0	0	3
Transaminases increased	46	136	34	216
Hepatic mass	12	13	0	25
Blood bilirubin abnormal	6	7	0	13
Blood alkaline phosphatase increased	51	150	50	251
Blood alkaline phosphatase abnormal	6	0	0	6
Pneumobilia	0	8	0	8
Hypertransaminasaemia	0	35	0	35
Liver function test increased	0	84	0	84
Congestive hepatopathy	0	9	0	9
Total	278	1625	425	2328

Note: Reported number refers to total drug‐event combinations.

Abbreviation: PTs: preferred terms.

Among the 1426 cases of AEs, 39 cases did not report outcome, 444 cases reported more than two outcomes, and 2004 outcomes were reported. We counted these 2004 outcomes separately and analysed them by PT groups (Table 4).

TABLE 4.

The outcomes in reports of AIs associated hepatotoxicity.

Outcomes	Anastrozole	Letrozole	Exemestane	Total
Life‐threatening	24	97	8	129
Hospitalisation‐initial or prolonged	79	350	60	489
Disability	2	4	5	11
Death	36	136	27	199
Congenital anomaly	0	1	0	1
Required intervention to prevent permanent impairment damage	3	0	1	4
Other serious	184	778	209	1171
Total	328	1366	310	2004

There were 7 clinical outcomes, including: ‘Death’ (DE); ‘Life‐Threatening’ (LT); ‘Hospitalisation‐Initial or Prolonged’ (HO); ‘Disability’ (DS); ‘Congenital Anomaly’; ‘Required Intervention to Prevent Permanent Impairment/Damage’ (RI); ‘Other Serious’ (OT). In addition to OT, the most common clinical outcome associated with these three drugs was HO, followed by DE. The most common multiple outcome combinations were those involving ‘Hospitalisation‐Initial or Prolonged’ (HO) and ‘Other Serious (Important Medical Event)’ (OT), as shown in Figure 5.

FIGURE 5.

The numbers and relevance of outcomes associated with AIs. CA: congenital anomaly; DE: death; DS: disability; HO: hospitalisation‐initial or prolonged; LT: life‐threatening; non: outcome non‐reported; OT: other serious (important medical event); RI: required intervention to prevent permanent impairment/damage.

4. Discussion

Our study reviewed 1426 cases of liver‐related AEs associated with AIs in the FAERS database from Q1 2004 to Q1 2025. The results of disproportional analysis demonstrated a strong link between AIs and hepatotoxicity, suggesting that the liver‐related AEs of the three drugs anastrozole, letrozole and exemestane all require attention, especially letrozole. The analysis of the reported time‐to‐onset profiles of AIs and liver‐related AEs revealed temporal patterns compatible with a potential drug‐event association, although causality cannot be definitively established within a spontaneous reporting system.

We found that the largest number of hepatotoxicity reports were associated with letrozole, and the case number increased significantly since 2017. This may be partly attributed to policy changes implemented in Germany regarding adverse drug reaction reporting and pharmacovigilance. In 2017, Germany adjusted Chapter VI of the Medicinal Products Act (Arzneimittelgesetz‐AMG) to add a requirement for systematic adverse drug reaction (ADR) collection in pharmacovigilance studies, which could lead to an increase in the number of reports [18]. In addition, since 2017 the European Union has required all non‐serious ADRs to be reported electronically to the EudraVigilance database within 90 days [19], which may also lead to a significant increase in the number of ADR reports in Germany. In addition, we found that liver‐related AEs associated with exemestane had the shortest median time‐to‐onset, followed by letrozole and anastrozole. Our findings indicate potential safety signals and reporting patterns, but they are not representative of the true incidence of hepatotoxicity in patients treated with AIs due to the lack of external denominator data in the real world [20].

AIs have become the standard of care for postmenopausal HR+ breast cancer patients [6, 21, 22, 23]. Randomised controlled trials, including the ATAC trial, the BIG‐198 trial, the FATA‐GIM3 trial, and numerous others, provide detailed evidence for the clinical use of AIs [24, 25, 26]. Although studies have suggested that compared with tamoxifen, AIs have a higher incidence of hepatotoxicity [11], Hepatotoxicity data for AIs based on real‐world data are rarely reported, and clinical trials have focused only on changes in key liver function abnormalities such as AST, ALT and bilirubin, with mixed conclusions [26, 27].

Cases of drug‐induced hepatotoxicity attributed to AIs have been reported over the past few years, ranging in severity from mild and self‐limited liver enzyme elevation to acute liver failure. The hepatotoxicity associated with letrozole is characterised by drug‐induced hepatitis, with significant elevation of serum ALT, AST and other liver transferases, and the risk of liver failure [13, 28, 29]. A 70‐year‐old female patient with breast cancer developed drug‐induced hepatitis with markedly elevated hepatic transaminases after 3 months of taking letrozole [28]. In 2022, there was a case reported clinically apparent liver injury associated with letrozole therapy. A 75‐year‐old woman with breast cancer, clinical stage IIA, presented with icteric syndrome and liver failure 1 month after treatment with neoadjuvant letrozole [13]. Hepatotoxicity associated with anastrozole may manifest as hepatitis, cholestatic liver injury, acute jaundice, etc., and can lead to liver failure in severe cases [30, 31, 32, 33]. A 66‐year‐old woman with a history of breast cancer developed progressively abnormal liver function, associated interface hepatitis, and numerous necroinflammatory foci throughout the liver parenchyma 6 months after starting anastrozole treatment [30]. Four months after beginning anastrozole, the breast cancer patient aged 70 years developed a mixed hepatocellular and cholestatic liver injury, which was dramatically improved in both clinical manifestation and laboratory index after anastrozole discontinuation [31]. Consistent with these two studies, our study found that the median occurrence time of liver‐related AEs associated with anastrozole was longer than that of letrozole and exemestane. However, anastrozole has also been reported to be associated with severe acute hepatitis. In a case report of severe hepatotoxicity due to metabolites of anastrozole, a 58‐year‐old female patient experienced severe asthenia and progressively developed jaundice and dyspnoea after 3 weeks of anastrozole use [32]. Cases of exemestane‐induced hepatotoxicity were relatively few. Serum liver enzymes are elevated in 4%–11% of female patients treated with exemestane [34]. A 47‐year‐old postmenopausal patient, diagnosed with stage I breast cancer, suffered from severe prolonged cholestatic hepatitis 3 weeks after initiating exemestane [35]. Although our results based on FAERS indicated that there are differences in the number of reports and the median occurrence of hepatotoxicity associated with the above three drugs, we cannot determine the precise clinical severity of hepatotoxicity based solely on it. Moreover, confounding factors, such as comorbidities and concurrent medications, cannot be ruled out. More patient‐level data are needed to further clarify the frequency and severity of hepatotoxicity associated with different AIs.

This data provides potential plausibility support for the association between AIs and hepatotoxicity. Hassan and colleagues found that letrozole could induce hepatorenal oxidative stress (OS) and mitochondrial‐dependent apoptotic progression to alter hepatic and renal functions [12]. Some reports indicated that hepatotoxicity induced by letrozole might arise from the toxic or immunogenic metabolite [28, 34]. Anastrozole may induce hepatotoxicity due to a poisonous or immunoallergic intermediate of its metabolism [30, 31, 32, 34]. While the specific metabolite is unknown, metabolism via CYP enzymes could generate reactive intermediates. It has also been suggested that the anastrozole‐GSH conjugate may play an important role in hepatotoxicity [36]. The acute liver injury attributed to exemestane treatment is considered to be an idiosyncratic reaction to a metabolite of the medication [34], and the pharmacokinetics of exemestane are closely related to liver function. Oral clearance of exemestane was reduced in the presence of significant hepatic disease, which could increase systemic exposure and the risk in predisposed individuals [37]. AI related hepatotoxicity involves varied and not fully understood mechanisms, likely including oxidative stress (letrozole), metabolite‐mediated toxicity (anastrozole), and clearance (exemestane), and more preclinical and clinical evidence is needed.

In pharmacovigilance, BCPNN is a Bayesian disproportionality method for detecting frequent drug adverse event associations in spontaneous reporting data. It does not require labelled outcomes and can therefore be viewed loosely as a form of unsupervised signal or anomaly detection in sparse discrete co‐occurrence data. Its core statistic (IC) compares the observed joint reporting probability with that expected under independence [13, 28]. This finding, consistently identified by both the traditional ROR and the BCPNN method, underscores the immense value of employing complementary computational analytical frameworks to differentiate safety profiles among drugs of the same class.

This study has several limitations common to spontaneous reporting systems. Firstly, residual confounding cannot be excluded. Breast cancer patients, particularly those with advanced or metastatic disease, may develop liver‐related abnormalities due to liver metastases, biliary obstruction, portal hypertension, malignant ascites, prior chemotherapy, alcohol use, viral hepatitis, non‐alcoholic fatty liver disease, or other pre‐existing liver disorders. In addition, AIs are frequently used sequentially or concomitantly with other potentially hepatotoxic anticancer therapies, including CDK4/6 inhibitors, HER2‐targeted agents, mTOR inhibitors, and cytotoxic chemotherapy. The increase in letrozole‐related reports may therefore reflect not only pharmacovigilance reporting changes but also evolving breast cancer treatment patterns, particularly the increasing use of letrozole in combination with CDK4/6 inhibitors. Restricting AIs to the primary suspect role may reduce but cannot eliminate confounding by indication or concomitant medication effects. These concerns are further compounded by the inherent limitations of the FAERS database, including repetitive reporting, underreporting, and incomplete data or detailed laboratory information; the spontaneous reporting system does not easily allow for the complete exclusion of synergistic or confounding hepatotoxic effects associated with these common combination therapies, restricting our ability to perform stratified analyses. Secondly, while we integrated both traditional frequentist metrics (ROR) and a well‐established Bayesian signal‐detection algorithm (BCPNN) to enhance reliability, disproportionality analyses inherently cannot establish causality alone and remain sensitive to upstream data quality issues. Thirdly, we did not use external denominator data to reflect the frequency of AI‐related hepatotoxicity. Last but not least, valuable unstructured data in FAERS (such as free‐text clinical narratives) were not utilised in this study, as extracting nuanced toxicological patterns from textual descriptions requires advanced Natural Language Processing (NLP) capabilities that fall beyond our current pipeline.

5. Conclusion

This study summarises the liver‐related AEs and clinical outcomes associated with letrozole, anastrozole, and exemestane based on real‐world data from the FAERS database from Q1 2004 to Q1 2025. The median time to onset of hepatotoxicity related to exemestane was found to be the shortest, the reporting disproportionality signal of hepatotoxicity related to letrozole was the highest, and clinical outcomes were similar for all three drugs. The findings of this study are beneficial for clinicians and clinical pharmacists in enhancing their comprehension of the hepatotoxicity associated with third‐generation AIs. During the clinical course of treatment, liver function should be closely monitored, and medication should be adjusted promptly according to the clinical presentation of patients for all three drugs.

In the future, prospective epidemiological studies and long‐term clinical trials are needed to verify the findings of this study. Moving beyond tabular disproportionality queries, we recommend that future AI‐enhanced pharmacovigilance research could integrate advanced computational architectures: (1) Deep Learning for Unstructured Text: applying Transformer‐based NLP models (e.g., BioBERT) to mine free‐text adverse event narratives in FAERS for hidden clinical phenotypes and severity markers; (2) Integration of Multi‐modal Data: utilising Graph Neural Networks (GNN) or Federated Learning to combine spontaneous signals with Electronic Health Records (EHR) to adjust for confounders; (3) Predictive Modelling: employing machine learning algorithms (such as random survival forests or eXtreme Gradient Boosting) based on patient demographics and concomitant medications to predict individualised risks for early hepatotoxicity. These advanced AI tools are expected to overcome current limitations of spontaneous reporting databases, such as the analysis of unstructured narrative text and the lack of denominators, thereby refining signal detection in pharmacovigilance. Additionally, incorporating Explainable AI frameworks could help translate black‐box signal detection into transparent clinical decision support tools.

Author Contributions

Yuke Li: data curation, conceptualization, investigation, methodology, formal analysis, visualization, writing – original draft, writing – review and editing. Hongmei Zheng: data curation, investigation, methodology, formal analysis, writing – review and editing. Yanting Wang: data curation, investigation, methodology, formal analysis, writing – review and editing. Jun Yang: investigation, formal analysis, writing – review and editing. Suying Xu: investigation, formal analysis, writing – review and editing. Peng Zhan: investigation, formal analysis, writing – review and editing. Yanna Zhu: investigation, methodology, formal analysis, resources, validation, supervision, writing – review and editing. Di Du: conceptualization, data curation, investigation, methodology, software, project administration, resources, supervision, visualisation, writing – review and editing. The manuscript has been read and approved by all the authors.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

Raw data was accessed from the FAERS online public platform. This data can be found here: https://fis.fda.gov/extensions/FPD‐QDE‐FAERS/FPD‐QDE‐FAERS.html.