The original data utilized for our data mining were sourced exclusively from the FAERS database available on the official FDA website. The raw ASCII data files were downloaded for subsequent data mining and statistical analysis. SAS 9.4 software, recommended by the FDA for FAERS database analysis, was employed for this purpose. It is important to note that the FAERS database contains self-reported data, which may include duplicate or withdrawn/deleted reports. In accordance with the FDA’s official guidance documents, we conducted data cleaning as follows: We selected the PRIMARYID, CASEID, and FDA_DT fields from the DEMO table, sorted them by CASEID, FDA_DT, and PRIMARYID, and retained the maximum values in sequence. Since the first quarter of 2019, each quarterly data package has included a list of deleted reports. After deduplication, we removed the deleted reports based on the CASEID provided in the deletion list. The latest version of the Medical Dictionary for Regulatory Activities (MedDRA 27.0), which includes system organ class (SOC) and preferred term (PT), was systematically utilized to encode the AE terminology within the FAERS database.

For the analysis of baseline data, we employed Cramer’s V coefficient in conjunction with the p-value to evaluate both the strength and statistical significance of associations between categorical variables, particularly for those with larger sample sizes. A p-value more than 0.05 accompanied by a V value close to zero suggests a weak association between the variables. Such a weak association may arise by chance and is therefore considered to have limited reliability. And we employed a case-control study design, specifically a case versus non-case approach, utilizing a fourfold table for disproportionality analysis, as illustrated in Table 1 , to investigate the AEs associated with the target medications. In the database, each patient (represented by an individual report) is associated with a unique “Primary Suspect (PS)” medication. When defining the target medication user population, only the patient’s PS medication is taken into account. If the PS medication recorded in the background database matches the target medication of the study, the patient is included in the target medication population; otherwise, the patient is classified into the other medication population.

In this study, we employed widely recognized methods to detect adverse medication event signals. Specifically, positive signals were identified using the reporting odds ratio (ROR) and the Bayesian confidence propagation neural network (BCPNN) (23, 24). For the ROR method, a signal is triggered when the analysis data satisfies a ≥ 3 and 95% confidence interval (CI) lower limit of ROR (ROR₀₂₅) > 1. For the BCPNN method, a signal is generated if 95% CI lower limit of the information component (IC₀₂₅) > 0. As illustrated in Figure 1 , a positive signal is produced when the AE included in the analysis meets the criteria for both methods.

ROR is a disproportionality analysis method based on the fourfold contingency table, characterized by high sensitivity in detecting typical AEs associated with commonly used medications. It considers the total number of reports and allows for adjustment of potential confounding factors (25, 26). And BCPNN is an advanced algorithm grounded in Bayesian theory, designed to evaluate the likelihood of a causal association between medications and AEs. This method demonstrates superior performance in handling sparse data compared to conventional approaches and tends to produce fewer false-positive signals. Furthermore, the Bayesian framework of BCPNN inherently accounts for reporting bias by incorporating prior distributions to mitigate spurious associations (27, 28). Therefore, this study employs both the ROR and BCPNN methods, which together can help minimize bias resulting from proportional imbalances and are widely recognized for their effectiveness in mining and analyzing similar pharmacovigilance databases such as FAERS (29, 30).

The data screening and analysis process is illustrated in Figure 1 . A total of 54 quarters of ASCII packets were downloaded from the FAERS database, covering the period from the first quarter of 2011 to the second quarter of 2024, resulting in the acquisition of 18,657,950 records. After eliminating duplicate reports, 15,886,542 unique reports remained. Subsequently, we screened for reports where the target medication was listed as the PS medication, selecting a total of 118,001 AEs involving 55,052 individuals as the background dataset. Specifically, among sICIs, there were 7,386 reports for ipilimumab, 29,544 for nivolumab, 5,278 for durvalumab, and 25 for tremelimumab. For dICIs, 12,256 participants reported receiving ipilimumab combined with nivolumab, while 563 people received durvalumab and tremelimumab.

Among the six subgroups included in the study, patients who received ipilimumab, nivolumab, tremelimumab, or durvalumab as monotherapies were classified as sICIs. Patients who received combination therapies of ipilimumab plus nivolumab or tremelimumab plus durvalumab were classified into the dual immunotherapy group (DIG). We have presented the baseline data for the sICIs and DIG in Table 2 , and provided detailed data pertaining to the six subgroups in Additional File 1 . After excluding unspecified cases, 51.73% of patients were male, with 51.19% in the sICIs and 53.54% in the DIG, both proportions exceeding those of female patients. The highest frequency of AEs was observed in the age group of 45–64 years, accounting for 20.81% in the sICIs and 25.58% in the DIG. In the DIG, the majority of AE reports were submitted by physicians (44.71%), whereas in the sICIs, consumers (32.26%) and physicians (31.94%) were the primary reporters. Regarding the sources of the reports, Figure 2 presents a choropleth map illustrating the countries contributing to the reports. The majority of the reports originate from the United States (25,788, 46.85%) and Japan (8,794, 15.98%), followed by France (4,031, 7.32%) and China (2,697, 4.90%). Notably, each of these four countries has contributed more than 2,000 reports, whereas the others have not. Serious reports in the FAERS database are categorized at the patient level rather than at the level of specific AEs. Reports that include outcome results are classified as serious (47,350, 86.01%), whereas those lacking outcome information are deemed non-serious (7,702, 13.99%).

Onset time refers to the duration from the initiation of medication until the occurrence of AEs. As shown in Figure 3 , after excluding reports with missing or abnormal onset times, DIG exhibits an earlier first AE compared to sICIs. For this processed dataset, we focus primarily on the median values ( Table 2 ). The median onset time for DIG is 36 days (interquartile range 14–82 days), which is notably shorter than that of sICIs at 50 days (interquartile range 16–131 days). It is important to note that a substantial proportion of onset time data is missing (76.13%), which may potentially introduce significant bias into the interpretation and reporting of onset time results.

We analyzed the target AEs using two algorithms, ROR and BCPNN. An AE positive signal was generated only when both algorithms met their respective signal generation conditions simultaneously. As illustrated in Figure 4 , at the SOC level, DIG (primarily derived from ipilimumab plus nivolumab) demonstrated significant AEs in detecting endocrine disorders (SOC: 10014698), with 19 distinct types of positive signals identified. In contrast, sICIs exhibited fewer types of positive signals. Moreover, DIG detected a higher number of AE types across multiple SOCs, including skin and subcutaneous tissue disorders (10 PT types, SOC: 10040785), nervous system disorders (6 PT types, SOC: 10029205), hepatobiliary disorders (9 PT types, SOC: 10019805), investigations (8 PT types, SOC: 10022891), infections and infestations (5 PT types, SOC: 10021881), metabolism and nutrition disorders (4 PT types, SOC: 10027433), musculoskeletal and connective tissue disorders (5 PT types, SOC: 10028395), cardiac disorders (4 PT types, SOC: 10007541), renal and urinary disorders (3 PT types, SOC: 10038359), blood and lymphatic system disorders (1 PT type, SOC: 10005329), eye disorders (3 PT types, SOC: 10015919), immune system disorders (3 PT types, SOC: 10021428), and vascular disorders (1 PT type, SOC: 10047065) compared to sICIs.

Tremelimumab exhibited a very low number of positive signals, likely attributable to the limited number of reports. Besides tremelimumab, fewer types of positive signals were observed in several other categories: general disorders and administration site conditions (2 PT types, SOC: 10018065), injury, poisoning and procedural complications (1 PT type, SOC: 10022117), respiratory, thoracic and mediastinal disorders (1 PT type, SOC: 10038738), as well as neoplasms benign, malignant and unspecified (incl cysts and polyps) (2 PT types, SOC: 10029104).

We present the AEs with a > 100, summarized in Figure 5 according to the source of the positive signal. In DIG, the most frequently reported AE was intentional product use issue (a = 679), followed by diarrhea (a = 633), colitis (a = 581), pyrexia (a = 530), and others. In sICI, the top three AEs associated with ipilimumab were diarrhea (a = 665), colitis (a = 607), and adverse event (a = 566, PT: 10060933). For nivolumab, the leading AEs were death (a = 6,216), malignant neoplasm progression (a = 3,129), and off-label use (a = 1,605). Durvalumab was associated with death (a = 1,405), malignant neoplasm progression (a = 619), and radiation pneumonitis (a = 456).

The data were analyzed using two methods, ROR and BCPNN, with detailed results provided in Additional File 2 . In Figure 6 , we summarize the IC₀₂₅ of positive signals with a > 100 obtained by the BCPNN method for DIG and sICIs (including non-positive signals). DIG exhibited higher positive signal intensity compared to sICIs across the following PTs: myocarditis (ROR 2.221, IC₀₂₅ 0.486, PT: 10028606), immune-mediated myocarditis (ROR 2.922, IC₀₂₅ 0.610, PT: 10082606), adrenal insufficiency (ROR 2.503, IC₀₂₅ 0.602, PT: 10001367), hyperthyroidism (ROR 1.872, IC₀₂₅ 0.305, PT: 10020850), thyroiditis (ROR 2.669, IC₀₂₅ 0.546, PT: 10043778), immune-mediated enterocolitis (ROR 3.948, IC₀₂₅ 0.937, PT: 10078961), pyrexia (ROR 1.570, IC₀₂₅ 0.290, PT: 10037660), hepatic function abnormality (ROR 2.582, IC₀₂₅ 0.591, PT: 10019670), hepatitis (ROR 2.705, IC₀₂₅ 0.637, PT: 10019717), liver disorder (ROR 2.718, IC₀₂₅ 0.646, PT: 10024670), immune-mediated hepatitis (ROR 5.504, IC₀₂₅ 0.994, PT: 10078962), immune-mediated liver disorder (ROR 5.322, IC₀₂₅ 0.966, PT: 10083521), cytokine release syndrome (ROR 7.650, IC₀₂₅ 1.103, PT: 10052015), autoimmune diseases (ROR 1.754, IC₀₂₅ 0.275, PT: 10061664), sepsis (ROR 1.414, IC₀₂₅ 0.062, PT: 10040047), diabetic ketoacidosis (ROR 2.294, IC₀₂₅ 0.472, PT: 10012671), type 1 diabetes mellitus (ROR 2.421, IC₀₂₅ 0.508, PT: 10067584), arthritis (ROR 1.562, IC₀₂₅ 0.113, PT: 10003246), myositis (ROR 2.204, IC₀₂₅ 0.412, PT: 10028653), and acute kidney injury (ROR 1.708, IC₀₂₅ 0.264, PT: 10069339). In contrast, the occurrence of death (PT: 10011906), product storage error (PT: 10079843), malignant neoplasm progression (PT: 10051398), and lung disorder (PT: 10025082) is lower in DIG compared to sICIs. The relevant ROR analysis results (only DIG ROR₀₂₅ > 1) are presented in Table 3 .

Due to the promotion of the STRIDE regimen, the combination of Durvalumab and Tremelimumab has been more widely adopted in clinical practice compared to Tremelimumab monotherapy, largely due to its improved safety profile. Comparing the safety signals between the Durvalumab + Tremelimumab group and the Durvalumab monotherapy group can more clearly illustrate the differences in their safety profiles. We present the AEs with positive and strongly positive IC signal intensities in Additional File 3 . The results indicate that the use of Durvalumab alone is strongly associated with AEs such as radiation pneumonitis (PT: 10037765), recurrence of non-small cell lung cancer (PT: 10029515), and disease progression (PT: 10061818), whereas these associations are not observed in the Durvalumab + Tremelimumab group. However, it should be noted that the use of this dICI regimen is strongly linked to AEs including cytokine release syndrome (PT: 10052015) and ruptured liver carcinoma (PT: 10050842).

We conducted an extensive analysis of the FAERS database to investigate the real-world toxicity profile of dICI, representing the most comprehensive evaluation of AE reports for dICI since its clinical introduction. This study aims to provide detailed insights into AE occurrences for clinical practice and establish a foundation for risk assessment in clinical decision-making. However, several limitations must be acknowledged. The FAERS database relies on voluntary reporting, where reporters may not possess specialized medical knowledge, and there is no requirement to provide evidence of a causal relationship between a drug and a specific AE. Furthermore, the incidence rate cannot be accurately determined based on these reports. AEs may be associated with the underlying medical condition being treated, concomitant medications, or other external factors. It should be noted that these reports only represent the observations and opinions of the individuals who submitted them. Moreover, the US Food and Drug Administration cannot capture all adverse event or medication error reports, and the quantity and quality of submitted reports are influenced by various factors.