The oral anticoagulant vitamin K antagonist (VKA) drugs such as warfarin have been widely used for decades as the only oral anticoagulant. Despite its known effectiveness, a frequent dose adjustment is required to maintain its therapeutic range, and risk of haemorrhage is significant. The introduction of direct oral anticoagulants (DOACs)—apixaban, dabigatran, edoxaban, and rivaroxaban—allowed clinicians to overcome these limits. In fact, DOACs have shown a lower risk of haemorrhage than that for warfarin; moreover, the unsolicited routine monitoring with the use of DOACs is much more convenient for patients. In the pivotal randomized clinical trials (RCTs), each DOAC was compared to warfarin, but no head-to-head comparison between the individual DOAC has been performed (Connolly et al., 2009; Granger et al., 2011; Patel et al., 2011; Giugliano et al., 2013). Some unanswered safety aspects remain to address, and observational studies utilizing real-world data could play a complementary role to RCTs. Several observational studies have currently emerged to provide supportive evidence of the safety and/or effectiveness of DOACs in a real-world setting (Keshishian et al., 2016; Kohsaka et al., 2017; Li et al., 2017; Kohsaka et al., 2018; Russo-Alvarez et al., 2018; Coleman et al., 2019; Ikeda et al., 2019; Inoue et al., 2019). Although the increasingly widespread use of DOACs, few studies evaluated their safety profile by leveraging pharmacovigilance data. Specifically, results from a study based on reports of suspected adverse drug reactions held in VigiBase have shown, as well as premarketing authorization RCTs, a reduced risk of intracranial haemorrhage, but an increased risk of gastrointestinal haemorrhage in patients treated with DOACs compared to warfarin (Monaco et al., 2017). Moreover, this study has shown several differences between different DOACs in the rate and type of suspected adverse events (AEs). Another study based on spontaneous reporting data from the Japanese database of Pharmaceuticals and Medical Devices Agency (PMDA) has shown that differences in drug safety aspects may exist between dabigatran and FXa inhibitors (apixaban, edoxaban and rivaroxaban), especially in terms of hemorrhagic and ischemic cerebrovascular events (Terayama, 2017). As already demonstrated, “Correspondence Analysis” (CA) could be a useful tool to uncover the relationships among categorical variables (Sourial et al., 2010). In fact, CA has become a popular method for the analysis of data in a wide range of fields, such as archeology, paleontology, linguistics, marketing, social sciences and more (Freudenthal et al., 2009; Alberti, 2013; Glynn, 2014; Hoare and Bock, 2019). However, CA is rarely use in health sciences. Its first use in health sciences was proposed by Greenacre: he introduced CA in an initial simple example using data on the relationship between headache types and age; then, he illustrated a more complex situation when several categorical variables are involved using test data on a collection of bacterial isolates, with the aim of comparing bacterial types and understanding the inter-relationships of the different tests (Greenacre, 1992).

CA is a multivariate graphical technique which contains three basic concepts: 1) that of a point in a multidimensional space; 2) a weight (or “mass”) assigned to each point; 3) a distance function between the points, called the chi-square distance. Once these three concepts are defined, the aim of CA is to reduce the dimensionality of the points by projecting them onto a subspace, a two-dimensional plane (CA “biplot”) that can be seen as a spatial map of the data. In this subspace, each point is weighted by its respective mass, and the measurement of distance between points and subspace is in terms of chi-square distance. In its simplest form, CA applies to a two-way cross-tabulation (Greenacre, 1984; Greenacre, 2007). In pharmacovigilance studies, researchers often interested in exploring the relationships among categorical variables with the goal of examining associations among these variables (Ferrajolo et al., 2014; Donati et al., 2016). One might consider conducting separate chi-square tests (one for each pair of variables), but this pairwise strategy would quickly become cumbersome and render the results difficult to summarize. More importantly, it would not provide us with a global picture of the salient relationships among these variables when taken together. CA approach explores these relationships “simultaneously” (Sourial et al., 2010). Given that, we believe that the use of CA should be considered in pharmacovigilance studies in addition to the established measures (e.g., the Reporting Odds Ratio, ROR), especially if high dimensional categorical data are investigated. As demonstrative example of using CA in this area, this study analyses the association of suspected AEs with the use of DOACs and warfarin, focusing on bleeding events.

The present study evaluated the safety profile of warfarin and DOACs using real-world data sent to the RNF by the Italian Region Campania. During the study period, we analyzed 1,161 ICSRs differently distributed between drug-populations. In line with their date of approval, the most common reported drug was warfarin, followed by dabigatran, rivaroxaban, apixaban and edoxaban. In fact, while warfarin was first introduced in the second half of the 20th century, the DOAC dabigatran and rivaroxaban were both introduced in 2008, apixaban in 2011, and edoxaban in 2015 (EMA, 2021a; EMA, 2021b; EMA, 2021c; EMA, 2021d). The overall average age of patients was 74.8 years (±11.0) with no significant differences in age distribution. Even if age remains one of the strongest risk factors for stroke in patients with atrial fibrillation (AF), real-world data suggest that a significant proportion of older patients are still not receiving stroke prophylaxis treatment in line with guideline recommendations, even in the absence of contraindications to oral anticoagulants (Steinberg et al., 2015). In addition to that, other important factors could influence the safety issue of the DOACs: the indication of use for which the DOAC is given, its dose, the concomitant drugs, linked to the potential risk of pharmacokinetic and pharmacodynamic drug-drug interaction, the comorbidities, and the duration of therapy (Roberti et al., 2021).

ICSRs related to apixaban, rivaroxaban and edoxaban are prevalently associated to serious AEs. Actually, as described in their Summaries of Product Characteristics, apixaban and edoxaban are used in patients who have one or more risk factors, such as having had a previous stroke, having high blood pressure, diabetes, heart failure or being 75 years old or over (EMA, 2021c; EMA, 2021d).

To our knowledge, this is the first study that analyses real-world safety data using the statistical tool of Correspondence Analysis. It is based on the analysis of the contingency table through the row and column profiles to present a unique graphical display showing the relationship among variables (Zhu et al., 2017). CA could represent an additional tool in exploring the relationship among sets of categorical variables. It is a geometric approach for visualizing the rows and columns of a two-way contingency table as points in a low-dimensional space (a representation which retains some meaningful properties of the original data). The aim of CA is to have a global view of the data that is useful for interpretation, also in terms of chi-square distance between different modalities of the same variables. Therefore, based on constituent ratios, we used CA to reveal the relationship between oral anticoagulants and AEs. Results showed that warfarin is associated to AEs grouped in HH, which mainly includes skin, urinary and respiratory tract hemorrhages. Apixaban was associated with CNSH, and dabigatran was associated with nHH. Moreover, dabigatran and apixaban showed a non-association with HH, and warfarin showed a non-association with nHH. Our results cannot confirm the increased tendency of DOACs to develop gastrointestinal bleeding compared to warfarin. Compared to warfarin, DOACS’ lower bioavailability (dabigatran 7%, rivaroxaban 66%, apixaban 50%, edoxaban 68%) suggests a potentially higher persistence in the gastrointestinal lumen with the resulting increased risk of bleeding particularly in case of high dosage (Desai et al., 2013). Our findings are conflicting with those of pre-marketing authorization RCTs, which showed an increased risk of gastrointestinal haemorrhage of DOACs against warfarin (Loffredo et al., 2015). Specifically, as described in the meta-analysis by Loffredo et al., rivaroxaban and high dosages of edoxaban and dabigatran significantly increased gastrointestinal bleeding against warfarin while a null effect was detected with apixaban. Conversely, some register-based observational studies identified no differences in the rates of gastrointestinal bleeding (Abraham et al., 2015; Chang et al., 2015; Sjögren et al., 2017). From our data, even if CA biplot showed a proximity of GIPUOH with edoxaban and rivaroxaban, we cannot conclude that edoxaban and rivaroxaban are associated with GIPUOH, because of their low contribution to both dimensions.

Relating to intracranial haemorrhages, our findings are contrary to the current evidence. In fact, according to the spatial distribution in CA biplot, apixaban resulted associated with CNSH while results from pivotal study showed a significant reduction of risk of bleeding with apixaban against warfarin (Lopes et al., 2010; Granger et al., 2011). The nationwide cohort study by Staerk et al. also showed that treatment with apixaban, as well as dabigatran, was associated with a significant lower risk of intracranial bleeding, compared with warfarin (Staerk et al., 2017). Our finding could be noteworthy, even though it could reflect limitations in pharmacovigilance databases. As expected, our results described a weak association between dabigatran and CSNH, but warfarin showed a similar result.

In line with RCTs and its summary of product characteristics, warfarin showed a high association to HH, which includes respiratory tract haemorrhage such as epistaxis and hemoptysis, genitourinary tract haemorrhage such as hematuria and menorrhagia, and skin bleeding such as ecchymosis and petechiae. Moreover, the association between dabigatran and nHH could suggest its overall safety in term of haemorrhagic events.

Disproportionality analysis with ROR confirmed the results of CA, that could suggest that CA is a reliable method. Moreover, in this analysis, we believe that CA showed many advantages over ROR in terms of summarizing and visualizing data and results. Unlike the pairwise comparison of ROR (i.e., apixaban/CNSH vs. dabigatran/CNSH), CA explores relationships between each category of each variable simultaneously; it could be useful in a context where there are many categories (and many variables, in the case of “Multiple CA”). Moreover, in contrast to the conventional statistical methods such as ROR, CA is not a confirmatory technique, trying to prove a hypothesis, but rather an exploratory technique, trying to reveal the data content. As described by Greenacre, CA could serve as a window onto the data, allowing medical researchers easier access to their results and facilitating discussion of the data and possibly generating hypotheses which can be formally tested at a later stage (Greenacre, 1992).

Both strengths and limitations are related to the study data source, a large database of spontaneously reporting adverse drug events. This kind of studies reflects both real-life events and, partially, real-life drug use, including drug use patterns that cannot be studied in clinical trials for ethical reasons (Rafaniello et al., 2020). However, several limitations need to be considered while interpreting this type of results. On the top of that, under-reporting phenomena (i.e., only a minority of AEs might be identified and reported) is highly probably in a spontaneous report system, leading to an underestimation of the real frequency of adverse events. Moreover, this type of data does not include information on drug-exposed population, thus we cannot estimate the incidence of these events among treated patients. Finally, as already discussed, our analysis lacks an assessment of the main confounders which could explain the differences of DOACs in terms of safety profile. Given that, our analysis has not been intended to detect new safety signal, but to assess the value of CA in pharmacovigilance studies. We showed that CA could represent an alternative or complementary methodological approach to the disproportionality analysis in assessing data from pharmacovigilance databases.