Search

Background: Drug-related adverse events remain an important cause of morbidity and mortality and impose huge burden on healthcare costs. Routinely collected electronic healthcare data give a good snapshot of how drugs are being used in 'real-world' settings., Objective: To describe a strategy that identifies potentially drug-induced acute myocardial infarction (AMI) from a large international healthcare data network., Methods: Post-marketing safety surveillance was conducted in seven population-based healthcare databases in three countries (Denmark, Italy, and the Netherlands) using anonymised demographic, clinical, and prescription/dispensing data representing 21,171,291 individuals with 154,474,063 person-years of follow-up in the period 1996-2010. Primary care physicians' medical records and administrative claims containing reimbursements for filled prescriptions, laboratory tests, and hospitalisations were evaluated using a three-tier triage system of detection, filtering, and substantiation that generated a list of drugs potentially associated with AMI. Outcome of interest was statistically significant increased risk of AMI during drug exposure that has not been previously described in current literature and is biologically plausible., Results: Overall, 163 drugs were identified to be associated with increased risk of AMI during preliminary screening. Of these, 124 drugs were eliminated after adjustment for possible bias and confounding. With subsequent application of criteria for novelty and biological plausibility, association with AMI remained for nine drugs ('prime suspects'): azithromycin; erythromycin; roxithromycin; metoclopramide; cisapride; domperidone; betamethasone; fluconazole; and megestrol acetate., Limitations: Although global health status, co-morbidities, and time-invariant factors were adjusted for, residual confounding cannot be ruled out., Conclusion: A strategy to identify potentially drug-induced AMI from electronic healthcare data has been proposed that takes into account not only statistical association, but also public health relevance, novelty, and biological plausibility. Although this strategy needs to be further evaluated using other healthcare data sources, the list of 'prime suspects' makes a good starting point for further clinical, laboratory, and epidemiologic investigation.

A thorough comparison between different QSAR modeling strategies is presented. The comparison is conducted for local versus global modeling strategies, risk assessment, and computational cost. The strategies are implemented using random forests, support vector machines, and partial least squares. Results are presented for simulated data, as well as for real data, generally indicating that a global modeling strategy is preferred over a local strategy. Furthermore, the results also show that there is an pronounced risk and a comparatively high computational cost when using the local modeling strategies.

The present work describes a method that optimizes a compound on the basis of the interpretation of quantitative structure-activity relationship models. The method has been applied to query compounds that have a mutagenicity liability. The substructure that contributes the most to the mutagenicity prediction is identified and replaced for each query compound. Replacement substructures have been generated in a deterministic fashion to produce a range of new, nonmutagen, compounds. A portion of the new compounds already exists in literature, but was unknown to the method during optimization. These results suggest that this method can substitute "toxic" substructures and produce libraries of compounds with lower liability for a given endpoint. This method is intended to complement or replace the database searches that chemists need to undertake when trying to avoid safety problems in compounds.

A method for local interpretation of QSAR models is presented and applied to an Ames mutagenicity data set. In the work presented, local interpretation of Support Vector Machine and Random Forest models is achieved by retrieving the variable corresponding to the largest component of the decision-function gradient at any point in the model. This contribution to the model is the variable that is regarded as having the most importance at that particular point in the model. The method described has been verified using two sets of simulated data and Ames mutagenicity data. This work indicates that it is possible to interpret nonlinear machine-learning methods. Comparison to an interpretable linear method is also presented.