Managing adverse drug reactions in the context of polypharmacy using regression models

This study highlights the feasibility of using Bayesian regression models with Horseshoe or Laplace priors for exploratory analyzes to detect potential drug candidates associated with a specific ADR. We were able to establish a much more comprehensive account of multiple factors in large and sparse data sets than would have been possible with the more commonly used traditional logistic regression, which is known to lead to overfitting or even non-convergence, especially in situations with high dimensional data and rare events³⁹. In situations close to non-convergence one can observe that coefficients take on very high, nonsensical values⁴⁰. It becomes necessary to apply the use of variable selection or shrinkage parameters to avoid overfitting, for example by applying priorities to the regression formula⁴¹. Despite the use of different priors, the horseshoe and lasso models provided consistent and clinically meaningful results, confirming the effectiveness of both regression techniques.

Shrinking horseshoe regression allowed the long list of drugs to be reduced to a small group of candidates previously associated with ADRs in the literature. Given the shrinkage characteristics, the coefficients were naturally mostly zero, although some trends remained evident. Similarly, the lasso regression also shrank most of the coefficients towards zero due to the Laplace distribution, but the shrinkage appeared to be less aggressive compared to the horseshoe model. This resulted in almost twice the number of fall ADR predictors in the Lasso model. Both methods showed greater consistency in identifying predictors of bleeding ADR.

Overall, the horseshoe model applied stricter constraints than lasso regression. While lasso regression could serve as a more comprehensive screening tool, the horseshoe model appears better suited to refining the analysis to identify the strongest drug-ADR associations. This dual approach could improve the accuracy and efficiency of detecting potential ADR-associated drugs and provide a practical and thorough method for prioritizing drugs for further investigation.

Among the positive predictors in the falls analysis, we found several psychotropic medications associated with falls, which is consistent with traditional drug candidate analysis studies¹². However, unlike traditional methods, our models considered more drug variables than potential confounders. This leads to a shift in focus to medicines for which there is currently less evidence than other medicines. Levetiracetam was the strongest predictor of decline among the coefficients tested and was positive in both horseshoe and lasso regressions. It has been described in several sources to be associated with a risk of falls, causing fall symptoms such as drowsiness, headache, asthenia, dizziness and anorexia^42.43. A meta-analysis even found that somnolence and asthenia were dose-independent side effects, with only a small number of side effects leading to harm⁴⁴possibly suggesting that even a small dose prescribed in our data set results in more falls. Citalopram also showed a strong positive trend in our analyzes and was consistently observed to be associated with an increased risk of falls in other studies^12.45. There is also evidence of an increased risk of falls for opioids such as oxycodone¹³which is consistent with known side effects such as dizziness and sedation. Additionally, there is evidence that opioid use may increase the risk of falls, particularly in polypharmacy¹³. Taken together, the data demonstrate the plausibility of our results. Horseshoe regression revealed a homogeneous group of psychotropic drugs as positive predictors compared to lasso, indicating a more conservative approach to the horseshoe technique.

The same applies to the bleeding analysis, where the models found several anticoagulant drugs, including direct oral anticoagulants (e.g. edoxaban, rivaroxaban, apixaban), vitamin K antagonists (e.g. phenprocoumon), and heparins (e.g .Enoxaparin). As expected, clopidogrel and acetylsalicylic acid, as antiplatelet agents, were also strongly associated with bleeding. Other results included ibuprofen (strongly positive) and diclofenac (mildly positive), both NSAIDs (non-steroidal anti-inflammatory drugs) and known to cause gastrointestinal bleeding and a general risk of bleeding by indirectly affecting platelet aggregation⁴⁶. The models therefore also led to very plausible results with this result ADR. However, indication confusion may also occur between metoclopramide and pantoprazole, showing a slight association with bleeding ADRs, as these drugs may be used for gastrointestinal bleeding⁴⁷.

By including drug groups as potential predictors in our models, we can account for potential DDIs in our models as the associations of predictors with the ADR outcome are adjusted for each other. In addition, we considered potential pharmacokinetic DDIs by including metabolic pathways in our models. In all models tested, we found an association between ADRs and the CYP2C19 signaling pathway. In the case of bleeding ADRs, this could possibly be explained by the frequent prescription of proton pump inhibitors for gastrointestinal bleeding⁴⁷There is evidence in fall ADRs that the CYP2C19 pathway may be important in drug-related falls⁴⁸. However, this requires further analysis in hypothesis-based approach studies.

Our analysis has several important limitations that require careful consideration. First, our study cohort includes only individuals who experienced all ADRs. As is often the case in pharmacovigilance studies, we need to compare a specific ADR (e.g. falls or bleeding) with a control group consisting of other ADRs^16,17. This condition can, as an artifact, result in a leftward shift in the posterior distributions and potentially obscure coefficients that would otherwise be important in comparisons such as ADRs with non-ADRs, thereby complicating the interpretation of the metrics. In particular, the identification of negative predictors cannot be generalized and interpreted as protective drugs. In addition to falls, bleeding events are often documented in the data set^2.4. A significant number of patients in the non-case cohort are also taking anticoagulant medications, explaining the negative predictors identified. Another example is tozinameran in bleeding analysis, which is unlikely to protect against bleeding but is associated with other side effects, making it more common in the control group. Regarding negative predictors, our models demonstrate their exploratory nature, as we also find drugs that are more commonly administered in stable patients.

The analysis benefits from a multicenter data set that minimizes regional differences and hospital-specific confounding effects. This reduces noise and increases the representativeness of the results. Consequently, our models exhibit good performance and validity. However, the dataset lacked reliable information on the indications for which the drugs were prescribed and the dosages of the drugs. This issue could be observed in our data set with levetiracetam associated with falls or with metoclopramide and pantoprazole associated with bleeding. In these cases, it is not possible to distinguish whether the medication was prescribed because of pre-existing symptoms associated with the final documented ADR or whether the medication was the cause of the ADR. This problem could be solved by a conventional hypothesis-driven approach to controlling confounding by indication in future research.

A further limitation arises from the choice of the credibility interval. While the horseshoe prior generally provides good interpretability, the aggressive shrinkage applied in our model increases the difficulty of setting reasonable thresholds to identify positive outcomes, such as: B. the 50% or 90% interval. Although we used the default interval settings provided by the software, the effects are not significant at traditional rejection thresholds of frequentist statistical models. This difficulty in establishing appropriate thresholds for this Bayesian model may hinder the practical application of the models and results. In our analysis, any drug that falls outside the 50% credibility interval is considered a potential cause of the decrease in ADRs. However, changing the size of the credibility interval can produce different results. This issue highlights the need for careful consideration when interpreting our results. Additionally, a limitation arises from the fact that Bayesian models are not as widely used as frequentist statistical models. Some commonly used metrics in the frequentist framework, such as: B. p-values ​​have no direct equivalents in Bayesian frameworks. This disparity can cause confusion for users unfamiliar with Bayesian interpretation and pose a barrier to accurately understanding these results.

Although we were able to include a much larger number of potential predictors in our models than typically analyzed, we still excluded drugs from our analysis. While we adjusted our models for the number of excluded drugs, potentially relevant drugs such as valproic acid, lamotrigine, ticagrelor, or doxazosin were excluded from our analyses. This is a clear limitation of our study.

Overall, our results suggest that the use of modified Bayesian regression models incorporating either horseshoe or lasso priors could be an effective approach to identifying drugs associated with a specific ADR in a population with polypharmacy in exploratory research are associated.