This large, single-centre study explores a unique perspective on the safety and efficacy aspects of individual DOACs by analysing a large real-world dataset for patients in different body weight classes. Overall, the results indicate a positive impact of specific DOAC types on specific clinical outcomes, with significant differences between obese and morbidly obese subgroups and other BMI categories.

The association between the BMI category (of the patients on DOACs) and the clinical outcomes was mixed. For example, being underweight significantly raised the odds of any thromboembolic events, all-cause mortality, length of stay, and emergency visits—the odds of ischemic stroke dropped significantly in the underweight patients (OR 0.53, 95% CI: 0.49–0.58, p = 0.001). The negative impact on a safety outcome (all-cause mortality) was expected as it reflects the higher exposure of standard dose DOACs that increase the risk of adverse events. Other factors besides high exposure, like age and comorbidity, may be responsible for the poor efficacy outcomes. A morbidly obese BMI was associated with significantly lower odds of ischemic stroke, length of stay, and all-cause mortality but higher odds of thromboembolic events. The positive outcomes may be attributed to other factors, including obesity paradox and DOAC type or dose—obesity paradox describes a phenomenon in which clinical outcomes like mortality appear to improve with high BMI (e.g., morbid obesity). In contrast, the negative outcome (e.g., thromboembolic events) may reflect underexposure given the pharmacokinetic or pharmacodynamic changes because of obesity (pharmacokinetic changes include increased clearance due to higher metabolic enzyme activity and glomerular filtration rate). Interestingly, the obese BMI category was associated with significantly lower odds of bleeding as expected.

The results from this study agreed with some of the previously published evidence: for example, Whittemore et al. [7], analysed the adjusted odds ratio following multivariate logistic regression and revealed that increased body weight decreased the risk of bleeding events. Wu et al. [24], also found that a higher BMI was negatively associated with bleeding events and mortality compared to normal BMI for DOACs overall (i.e., dabigatran and rivaroxaban). This suggests that DOACs were linked to better survival and lower bleeding risk in the obese (higher BMI) cohort compared to normal BMI patients. This was not the case for thromboembolic events, as no statistically significant association was reported for the morbidly obese cohort. Also, subgroup analysis of rivaroxaban and dabigatran did not yield statistically significant associations.

Furthermore, using a Cox proportional hazards model, Weitz et al. [25], reported a lower risk of all-cause mortality in obesity compared to patients with normal BMI. This suggests better safety and effectiveness (obesity paradox). On the other hand, Briasoulis et al. [26], reported an increased risk of all-cause mortality (HR of 1.12, 95% CI: 1.02–1.23) but no significant difference in stroke and bleeding events for morbidly obese patients when apixaban was compared with rivaroxaban. Also, in the subgroup of individuals with morbid obesity, dabigatran was associated with significantly lower all-cause mortality compared to rivaroxaban; apixaban was associated with greater mortality than dabigatran and rivaroxaban.

Meanwhile, in the studies by Netely et al. [11], and Wang et al. [19], the obese BMI category had no significant impact on bleeding or thrombotic events, agreeing with the findings from Aloi et al. [16], Perino et al. [17], and Deitelzweig et al. [18]. The authors suggested their findings were due to confounding biases associated with most observational studies. Specifically, other observational studies and randomised controlled trials (RCTs) widely suggest that increased BMI in VTE/NVAF has no significant effect on the safety and efficacy of apixaban. In other words, despite exposure being slightly reduced, no dose adjustment was required [20]. However, this does not dismiss the claim that the morbidly obese category had the most incidence of thrombotic events among the BMI classes [8, 9, 21, 22].

Lucijanic et al. [13], established that obesity increased the risk (odds) of stroke and bleeding in DOACs overall: dabigatran conferring lower efficacy (higher thrombosis/stroke risk) and Factor Xa inhibitors (rivaroxaban, edoxaban, and apixaban) increasing the odds of bleeding (somewhat similar to our result). Their study observed a positive association between BMI and time to thrombosis (TTT), and time to bleeding (TTB) for DOACs. Potential factors responsible for the increased risk of bleeding were inappropriate dosage regimens and concomitant interacting medications [23]. It is important to note that reference to specific DOAC doses was not made in the study (absence of subgroup analysis).

In our study, the majority of the patients were on apixaban (84%, n = 82,073), similar to the study by Briasoulis et al. [26]. This reinforces the status of apixaban as the most prescribed DOAC in the NHS (at the time of this study). In addition, the mean age in our study is within the same range as the study carried out by Barakat et al. [1], (74.1 years), which also examined the outcomes of DOACs in patients across different BMI categories—DOACs were linked to a lower risk of bleeding and stroke among obese and morbidly obese patients.

Notably, our study was based on the assumption that bleeding and mortality were measures for safety during ischaemic stroke or thromboembolic events, emergency visits and length of stay were used as measures for effectiveness—for instance, the shorter length of stay (or fewer emergency admissions) implies positive outcomes. A possible explanation for the inverse relationship with BMI for edoxaban could be due to the link with positive outcomes (which may include low risk of bleeding) from treatment and hence early discharge—or the sample size of patients (on edoxaban) was insufficient to draw firm conclusions. It is worth mentioning that bleeding events occurred only in normal BMI and overweight and obese patients, respectively. No deaths were recorded for obese patients who experienced bleeding.

Our results add weight to the potential safety and effectiveness of specific DOAC types (for example, edoxaban) in different patient BMI groups, even though the positive clinical outcomes were inconsistent across DOAC types or BMI groups. Indeed, there is mounting evidence of the safety and effectiveness of DOACs in obese patients, which is comparable to or even better compared to normal-weight patients [3].

A key strength of this study was the inclusion of a large sample size of obese patients who were underrepresented in landmark DOAC trials and remained underappreciated in clinical practice. Machine learning (ML) models are data-driven (i.e., thrive on large datasets) and known for their excellent pattern recognition and predictive abilities. The performance of the ML model relies heavily on the quality of data. The importance of input features towards class attribute prediction in decision trees and random forests increased the confidence in our ML solution. The ML approach is a multifaceted technique where the outcome goals could vary from task to task. Apart from predicting the unknown data, a well-trained and accurate ML model could be used to analyse the critical features in the dataset.

Irrespective of the ML model’s strength, low-quality or noisy data will result in poor predictions. Furthermore, a limitation of the study was its retrospective design hence the susceptibility to confounding biases. Also, there was sampling bias: it was impractical to do regression analysis on dabigatran and edoxaban (by BMI category) due to insufficient sample size (as a result of limited prescribing); the samples for bleeding outcomes were highly imbalanced (with the majority of cases being positive). It is also important to acknowledge that by grouping patients together as per outcomes, our analyses did not quantitatively account for their differential risks of adverse clinical outcomes (e.g., CHA2DS2-VASc score for stroke and HAS-BLED score for bleeding risk); also, there was no justification for recommendation of DOAC based on BMI thresholds. Finally, our use of BMI as a body size descriptor may limit the applicability of our results in terms of dose determination (it fails to account for the distribution of body fat given that DOACs are lipophilic)—more accurate biomarkers that describe the pharmacokinetics of DOACs are required (perhaps waist-to-hip ratio).Undoubtedly, these limitations may undermine the strength of our findings and reemphasise the importance of more accurate (advanced) models and well-designed studies (e.g. randomised controlled trials).