Triage algorithms for early discovery of adverse drug interactions
(English)Manuscript (preprint) (Other academic)
Background: Most methodological research for broad surveillance of drug interactions in large collections of suspected ADR reports has focused on measures of disproportionality. However, recent results indicate that reported clinical information and pharmacological characteristics may be at least as valuable to detect adverse drug interactions early.
Objective: To develop triage algorithms for adverse drug interaction surveillance, and to evaluate the algorithms prospectively relative to expert clinical assessment.
Methods: A previously developed reference set based on Stockley’s Drug Interactions was used to train the algorithms. Logistic regression was used to set the relative weights of the different indicators (information potentially suggestive adverse drug interactions such as pharmacological properties including cytochrome P450 (CYP) activity; explicit suspicions of drug interactions as noted by the reporter in different forms; clinical details such as dose and treatment overlap; and a measure of disproportionality based on the total number of reports on two drugs and one ADR together) of each algorithm. Three triage algorithms were designed. All are logistic regression models producing an estimated probability that a given case series constitutes an adverse drug interaction signal. Two of them are data driven: one which used a very broad set of indicators (full data-driven) and one which used a more narrow set (lean data-driven). The third was manually derived (lean clinical) as a simplified version of the full data-driven algorithm. An independent evaluation set was constructed that consisted of 100 randomly selected case series in the WHO Global Individual Case Safety Report (ICSR) Database, VigiBase, from January 1990 to February 2011. Each algorithm’s ranking of case series was evaluated against an evaluation set. In a complementary analysis the algorithm were compared to a pure disproportionality analysis.
Results: The two lean algorithms were comparable in performance. However both outperformed the full data-driven algorithm on the independent evaluation set. The areas under the curve (AUC) for the receiver operating characteristics (ROC) curves were as follows: 71% (lean clinical) and 69% (lean data-driven). For a false positive rate (FPR) of up to 0.04 the lean algorithms classifies about 14,000 case series as potential interaction signals. Thresholds corresponding to greater FPRs are unlikely to be feasible in practice. The algorithms clearly outperform disproportionality analysis alone.
Conclusions: The value of incorporating clinical and pharmacological information in first-pass screening for adverse drug interactions is clear. Two triage algorithms have been proposed that each effectively identify adverse drug interaction signals and clearly outperforming pure disproportionality analysis in this respect.
Individual Case Safety Reports, Adverse Drug Interactions, VigiBase, Triage algorithms
Medical and Health Sciences
IdentifiersURN: urn:nbn:se:liu:diva-70848OAI: oai:DiVA.org:liu-70848DiVA: diva2:442067